dspace cover page - eth z40989/et… · ph d thesis no 12065, swissfederalinstitute...
TRANSCRIPT
Research Collection
Doctoral Thesis
Projecting scenarios of climatic change and future weather forecosystem modelsderivation of methods and their application to forest in the alps
Author(s): Gyalistras, Dimitrios
Publication Date: 1997
Permanent Link: https://doi.org/10.3929/ethz-a-001889726
Rights / License: In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For moreinformation please consult the Terms of use.
ETH Library
Diss ETHNo 12065
Projecting Scenarios of Climatic Change and
Future Weather for Ecosystem Models:
Derivation of Methods and Their Application to
Forests In the Alps
A dissertation submitted to the
Swiss Federal Institute of technology Zurich
for the degree of
Doctor of Natural Sciences
presented by
Dimitrios Gyalistras
Dipl El Ing ETH
born Sept 17th, 1964
Athens, Greece
accepted on the recommendation of
Prof Dr H Fluhler, examiner
Dr A Fischlin, co-examiner
Prof Dr A Ohmura, co-examiner
Zurich, 1997
Acknowledgements
My warmest thanks are due to
Dr Andreas Fischlin
for initiating this research, supporting it with many fruitful ideas and doing
the careful supervision
Prof Hannes Fluhler and Prof Atsumu Ohmura
for standing by this work, their encouraging support, and the review of the
manuscript
Prof Martin Beniston, Prof Hans von Storch and Prof Heinz Wanner
who helped to establish and pursue this research and took time for many
important discussions
Dr Harald Bugmann, Prof Christoph Schar and Prof Huw Davies
for their stimulating comments and the valuable exchange during different
phases of the project
Jurg Thony, Dr Heike Lischke, Frank Thommen, Dr Olivier Roth,
Dr Daniel Perruchoud, Hans-Peter Laser, Dr Thomas Nemecek,
and all other colleagues from the Institute of Terrestrial Ecology ETHZ for
their help and the good working atmosphere
My partner Barbara Davies
for her support and patience
The financial support of the Swiss Priority Programme Environment, the
Swiss National Science Foundation, and the Swiss Federal Institute for Technology
is gratefully acknowledged
II
Leer - Vide - Empty
Table of Contents
1 introduction 1
2 The Spatial aspect Statistical Downscaling of GCM-simulated
Climatic Changes to the regional scale 5
2 1 Introduction 5
2 2 Data and Methods 10
2 2 1 Observations 10
2 2 2 GCM-Expenments 12
2 2 3 Statistical Procedure 13
2 3 Results and Discussion 17
2 3 1 Model Estimation 17
2 3 2 Model verification 22
2 3 3 Downscaling of GCM-Simulated Climatic Changes 30
2 4 Conclusions 36
3 The Temporal aspect Stochastic Simulation of Monthly
Weather 38
3 1 Introduction 38
3 2 Materia! and Methods 41
3 2 1 Data 41
3 2 2 Stochastic Models 44
3 2 3 Parameter Estimation 46
3 2 4 Biochmatic Variables 50
3 2 5 Statistical Tests 52
3 2 6 Modeling and Simulation Tools 53
3 3 Results 53
3 4 Discussion 57
3 5 Conclusions 60
4 synthesis and application assessing Impacts of Climatic Change
on Forests in the Alps 61
4 1 Introduction 61
4 2 Material and Methods 65
4 2 1 Baseline Climates 66
4 2 2 Climatic Scenarios 67
4 2 3 Forest Model ForClim 68
4 3 Results 69
4 3 1 Climatic Scenarios 69
4 3 2 Forest Responses 72
4 4 Discussion 74
4 4 1 Forest Responses 77
4 4 2 Sensitivities and Uncertainties 80
4 5 Conclusions 82
5 Discussion 84
6 Conclusions 88
7 References 90
List of Figures
2.1 Comparison between the climate simulated by the ECHAM1/LSG-GCM in
the vicinity of the Alps (averages from three gndpoints) and observations 7
2.2 Gndpoints of atmospheric fields and case study locations considered for
downscaling 10
2.3 Block diagram of the method used to derive scenarios of local climatic
changes from large-scale climatic changes as simulated by GCMs 14
2.4 CCA model for Bern in winter 18
2.5 CCA model for Davos in winter 20
2.6 CCA model for Bern in summer 21
2.7 Skill of the CCA models as a function of the predictors, season, location
and variable considered 24
2.8 Comparison of skills between the procedure proposed by VON STORCH et
al (1993) and the improved procedure 25
2.9 Comparison of statistically reconstructed time series of local weather
statistics with observations 26
2.10 1901-1980 linear trends in the large-scale data sets used 29
2.11 Climatic change scenario for Bern in winter (5 yr running means) 31
2.12 Statistically downscaled changes in seasonal mean temperatures for the last
20 yr of the "double CO2" experiment and the last 10 yr of the "IPCC Sce¬
nario A" experiment 31
2.13 Statistically downscaled changes in winter precipitation totals and summer
probabilities of wet days for the last 10 yr of the "IPCC Scenario A"
experiment 32
3.1 Observed (bio-)chmatic parameters at the 89 chmatological stations used to
test the stochastic simulation of monthly weather 43
3.2 Comparison of the performance of the Type IV models relative to the TypeI-III models 55
3.3 Comparison of the annual cycles of the diagonal elements a,, and a22 of
the system matrices A of cyclostationary, first-order autoregressive models
at selected European locations 56
4.1 Method used to derive climatic scenarios and simulate forest responses 65
4.2 Comparison of annual mean temperatures and annual precipitation totals
under the present and the assumed scenario climates at the case study sites 70
4.3 Comparison of in period 1931 -1980 observed annual cycles for monthlymean temperature and total precipitation with the climatic scenarios derived
by statistical downscaling from a "2xCOj"-expenment with the CCC-
GCMII at the case study sites 71
4.4 Simulated species compositions at the case study site Bever in the Swiss
Alps for current climate (800-2040) and a possible, future "2xC02" climate
(2060-3200) as downscaled from the CCC-GCMII 72
4.5 Simulated equilibrium species compositions at the case study sites in the
Alps for current baseline climate (800-2040) 73
4.6 Simulated equilibrium species compositions at the case study sites in the
Alps for a possible, future "2xCOj" climate (2060-3200) as downscaled
from the CCC-GCMII 73
4.7 Using the CLIMSHELL to explore forest dynamics where there are no
measurements available 75
List of Tables
2.1 Meteorological variables and their seasonal statistics considered 11
2.2 Mean linear trends of weather statistics in the period 1901 -1980 as
observed and reconstructed from 84 selected CCA models .28
2.3 Projected deviations of weather statistics from their respective 1901-1980
long-term means for the last 20 yr of the "double C02" experiment and for
the last 10 yr of the "IPCC Scenario A" experiment 33
3.1 Chmatological stations used to study the stochastic simulation of monthlyweather 42
3.2 Types of stochastic models considered 44
3.3 Types of output functions considered 45
3.4 Output functions used in different variants of stochastic models 45
3.5 Overview of the data used to compare distributions of bioclimatic variables
as derived from measured and stochastically simulated weather inputs 52
3.6 Medians of the p-values obtained from the comparison of the distributions
of selected bioclimatic variables derived from measured and stochasticallysimulated monthly weather data at 89 European chmatological stations 54
4.1 Characteristics and major current climatic parameters of the case studysites 66
Abbreviations
2mT 2m above ground temperature
AET Actual evapotranspiration
AR(1) Lag-1 autoregressive model
CCA Canonical Correlation Analysis
CCC Canadian Climate Centre
DC02 "Double CO2" experiment performed with the ECHAM 1/LSG model
ECHAM European Centre weather forecasting model, extensively modified at the
Max-Planck Institute for Meteorology, HAMburg for climate simulations
EOF Empirical Orthogonal Function
GCM General Circulation Model
GHCN Global Historical Climatology Network
IPCC Intergovernmental Panel on Climate Change
LSG Ocean model resolving large-scale geostrophic motions, developped at
the Max-Planck Institute for Meteorology, Hamburg
LTM Long-term mean
PC Principal Component
PCA Principal Component Analysis
PET Potential evapotranspiration
SCNA "IPCC Scenario A" experiment performed with the ECHAM 1/LSG
model
SLP Sea-level pressure
Notation and List of Symbols
The following notation is used x, 4 or H denote a scalar, X denotes a random variable, x
or X a vector, and A a matrix X' and AT denote the transpose of X and A, respectivelyA '
is the inverse, and a,j the (ij)-th element of A E[XJ\ STDEVfX], and SKEW[X]
denote the expected value, standard deviation and skewness of X, respectively, and
COV[X,Y] denotes the covariance of X and Y EIGVEC[A] denotes the matrix of the
eigenvectors of A, and EIGVALIA] is a diagonal matrix containing the eigenvalues of
A Subscripts (t, k etc ) are used to denote dependence on time or other dimensions
X(,) denotes a time series, <x(t)> its arithmetic mean, and MIN(x(1|) its minimum value
Below follow the most important symbols used in each chapter
Symbols Used in Chapter 2
eof, l-th empirical orthogonal function
tic threshold value for canonical correlations used for prediction
T)v threshold value for canonical correlations entering the calculation of 4*
nc total number of canonical correlations available
ncu number of canonical correlations used for prediction
nx, ny numbers of predictors/predictands entering PCA
Nx, Ny number of predictor/predictand PCs retained for CCA
pc^,) l-th principal component
P,, 0_, i-th canonical pattern for the predictors/predictands
Pm m-th canonical correlation coefficient
s,(t), t,(t) i-th canonical time coefficient for the predictors/predictants
as,, at, standard deviations of the s,(tj and t,(t)
t index denoting current time (time step = 1 year)
t,*(|) i-th estimated canonical time coefficient
v,(t) i-th original variable entering PCA
xj(t)> vk(t) j-th predictor/k-th predictant entering PCA
y*k(t) k-th estimated predictand
•P index measuring CCA model performance
Symbols Used in Chapter 3
ae,0 annual mean of an u, or o, (6 stands for u or a)
aeu, beu amplitudes of sines/cosines in Fourier expansion of an a, or a,
A(p) system matrix at phase p
A system matrix in the subspace spanned by the first few Fourier
harmonics of the annual cycle
B(p) input matrix at phase p
DSI index of annual drought stress
£(k) input vector of independent random components from a N-dimensional
normal distribution M0,1) at timepoint k
f|(p)' 8i(p)> ni(p) parameters defining log-normal transformation of the state vector
element X, at phase p
/ system output function
GDD annual growing degree-day total (temperature threshold = 5.5 °C)
rv(p) lag-v covariance matrix of state vector elements at phase p
i index denoting i-th element of the state or output vector
k index denoting current simulation time (time step = 1 month)
m,(p), s,(.) estimated expected value / standard deviation of output Y, at phase p
u1(p), a,(p) expected value / standard deviation of output Y, at phase p
M dimension of subspace used to fit phase-averaged cyclostationarymodels
n sample size of measurements
n6l number of Fourier harmonics used to represent the annual cycle of an
H, or o, (9 stands for u or a)
N dimension of the state and output vectors
p index denoting phase within the year (0=January, 1 l=December)
P monthly total precipitation
P P standardized to zero mean and unit variance
q number of Fourier harmonics used to fit phase-averagedcyclostationary models
t index denoting time in measured time series (time step = 1 year)
T monthly mean temperature
T T standardized to zero mean and unit variance
TWinMin winter minimum temperature
X<k) state vector at time k
yl( ,jmeasured time series for random variable Y, at phase p
y,( t)time series yl(p l(
standardized to zero mean and unit variance
y,( „ log-normal transform of time series y l(p1(
Y(k) output vector at time k
Z<k) auxiliary vector of back-transformed system states at time k
Symbols Used in Chapter 4
DSI index of annual drought stress
FC field capacity
GDD annual growing degree-day total (temperature threshold = 55 °C)
X latitude
P monthly total precipitation
p correlation between T and P
SA( slope-aspect index used to correct PET
S) similarity index used to compare species distributions
T monthly mean temperature
TWinMin winter minimum temperature
X
Leer - Vide - Empty
XI
Abstract
Dimitnos Gyalistras, 1997 Projecting scenarios of climatic change andfuture weather
for ecosystem models derivation ofmethods and their application toforests in the Alps
Ph D thesis No 12065, Swiss Federal Institute of Technology Zurich (ETHZ), 103 pp
Global climate models are the main source of information on possible future climatic
change, whereas local ecosystem models represent important tools to study possible
impacts of a changing climate on ecosystems If one wishes to combine the two kinds of
models, however, major problems occur due to the limited horizontal resolution, statis¬
tical precision and temporal extension of the global climate simulations These problems
have not been satisfactorily resolved until now The present thesis deals with the devel¬
opment, testing and application of a method to construct from the output of global climate
models climatic scenarios suitable to study ecosystems The assessment of possible
impacts of climatic change on forests in the Alps by means of a forest succession model
served as a case study The method was developed in three steps, as follows
The first step served to bridge the gap between the spatial scales at which General Circu¬
lation Climate Models (GCMs) and ecosystem models operate Based on the approach
proposed by von STORCH et al (1993), a statistical downscaling procedure to assess
local climatic changes from large-scale climatic changes as simulated by GCMs was
developed, verified and applied at five Swiss locations for the summer and winter sea¬
sons According to the very diverse and demanding input requirements of different
ecosystem models, at each location 17 seasonal statistics of daily temperatures, precipita¬
tion, sunshine duration, air humidity and wind-speed were considered Year-to-year
variations of the local variables were linked by means of Canonical Correlation Analysis
to simultaneous anomalies of the North Atlantic/European sea-level pressure and near-
surface temperature fields The statistical models were fitted separately for each season
and location for the interval 1901-1940
In all cases physically plausible statistical relationships were found, which quantified the
local effects of changes in major circulation patterns such as the strength of westerly flow
in winter and of large-scale subsidence in summer In the verification interval 1941-
1980, most variables were better reconstructed in winter than in summer, and better at the
three northern Alpine than at the two southern Alpine locations The most reliably recon¬
structed variables were seasonal mean daily temperatures and daily temperature extremes,
winter precipitation totals, and winter numbers of days with precipitation above 1 mm
Seasonal mean daily temperature amplitudes, relative humidities, and wind speeds, as
well as the within-season standard deviations of most daily variables were generally only
poorly reproduced The procedure of VON STORCH et al (1993) was, however, gen¬
erally improved by using in addition to sea-level pressure the near-surface temperature as
Xll
a large-scale predictor Improvement was strongest for temperature related variables, and
for the summer season Application of the statistical models to two simulations of global
climatic change with the ECHAM1/LSG-GCM (CUBASCH et al, 1992) yielded for sev¬
eral important ecosystem inputs time-dependent, spatially as well as between the weather
variables consistent, and regionally strongly differentiated estimates of possible climatic
changes at a spatial resolution far above the resolutions of present global or regional
climate models
The second step focused on the construction of statistically accurate weather inputs over
arbitrary time spans, and under any scenarios of time-dependent climatic change For
this purpose investigated was at 89 long-term European chmatological stations the perfor¬
mance of 12 stochastic models to simulate monthly mean temperature and precipitation
The models differed in the treatment and estimation of the monthly variables' auto-, and
cross-correlation coefficients, the annual cycles of the variables' expected values and
standard deviations, and in the use of a skewness-reducing transformation for precipita¬
tion All models were fitted to local measurements from the period 1951-1980 Model
performance was evaluated by comparing the distributions of monthly winter minimum
temperatures, annual growing degree-days, and an index for the annual drought stress as
derived from simulated and independently measured monthly weather data
It was found that inclusion of a memory term in the stochastic models generally enhances
the statistical accurracy of the simulated monthly mean temperatures, winter minimum
temperatures and growing degree-day totals The use of a log-normal transformation
strongly improved the simulation of the monthly precipitation totals and the drought
stress index It was shown that cyclostationary autoregressive models can be further im¬
proved, and the number of parameters substantially reduced by estimating the monthly
system matrices in a sub-space spanned by the first two Fourier harmonics of the annual
cycle according to the approach proposed by HASSELMANN & BARNETT (1981) The
representation of the annual cycles of the weather variables' expected values and standard
deviations by means of their first two (temperature), respectively three (precipitation),
Fourier harmonics yielded an improvement mainly for the drought stress index, but not
for the two other bioclimatic variables Based on the above findings a new stochastic
model to simulate monthly weather was proposed Compared to the commonly used
models it requires a slightly larger number of climatic parameters, but relies upon more
robust methods to estimate these parameters, and produces significantly improved distri¬
butions of bioclimatic variables
In a third step, the previous two steps were combined with the forest succession model
FORCLIM (BUGMANN, 1994, 1996, FISCHLIN et al, 1995) into an overall method to
project quantitatively possible impacts of climatic change on mountain forests at high tem¬
poral (annual cycle), spatial, and qualitative resolution The method consists of the fol-
XIII
lowing steps (1) Describe weather at the ecosystem location as a stochastic process,
(2) estimate the process/climatic parameters for present climate from measurements,
(3) use a statistical downscaling procedure to estimate from the output of a GCM time-
dependent changes in selected climatic parameters, (4) generate scenarios of monthly
weather by means of stochastic simulation, (5) use the simulated weather sequences to
drive the forest model
The method was applied to four representative sites in the Alps, starting from a 2xC02
scenario of global climatic change as projected by the the CCC-GCMII climate model
(BOER et al, 1992) Similar to the results obtained from the ECHAM 1/LSG model the
downscaled scenarios showed strong spatial and seasonal variation, and depicted a gen¬
eral warming which was larger at the northern than at the southern slope of the Swiss
Alps A tendency towards increased precipitation was found However, decreases were
obtained for some locations and seasons, with partially dramatic effects on simulated
forest compositions The scenarios appeared physically plausible and spatially consis¬
tent Sharply contrasting forest responses were obtained within short distances under the
same 2xC02 scenario of radiative forcing While some forest simulations produced only
small changes in tree species composition, others produced major changes even to the
point of a complete disappearance of the forest In some cases new species assemblages
emerged without any analogue under present conditions The forest responses were con¬
sistent with current understanding of forest dynamics They suggest that some mountain
forests are sensitive to a 2xC02 global change, and that human assistance may be
required to help forests to adapt
In summary, it was found that the proposed method allows to consistently link climatic
changes as simulated by global climate models to local ecosystem models The method is
general, flexible, computationally efficient, has intermediate data requirements, and
conforms to the IPCC guidelines (CARTER et al, 1994) for impacts assessments Its
main limitation is that the used statistical models may not hold under a changed climate
However, the method allows for extensive sensitivity and uncertainty analyses, and
thanks to its modular structure improvements of the individual components can be easily
incorporated at a later stage The overall method used to project forest responses makes
good use of existing models and data, integrates current understanding, and is suitable to
assess impacts of climatic change on any mid to high latitude forests
XIV
Kurzfassung
Dimitnos GYALISTRAS, 1997 Zur Erstellung von Szenarien zukunftiger Klimaverande-
rungen und Wetterverlaufefitr Okosystemmodelle- Herleitung von Methoden undderen
Anwendung aufWalder im Alpenraum Diss ETH No 12065, Eidg Technische Hoch-
schule Zurich (ETHZ), 103 pp
Globale Klimamodelle sind unsere Hauptinformationsquelle uber eine moghche zukunf-
tige Khmaveranderung, wahrend lokale Okosystemmodelle wichtige Werkzeuge darstel-
len, um moghche Auswirkungen ernes sich andernden Klimas auf Okosysteme zu unter-
suchen Mochte man die beiden Modelltypen miteinander kombinieren, so ergeben sich
jedoch aufgrund der begrenzten honzontalen Auflosung, statistischen Prazision und zeit-
hchen Ausdehnung der globalen Klimasimulationen grossere Probleme, die bisher noch
nicht befnedigend gelost worden sind Die vorliegende Dissertation befasst sich nut der
Herleitung, Uberprufung und Anwendung einer Methodik, um aus globalen Klimasimu¬
lationen passende Klimaszenarien fur Okosystemstudien zu erstellen Die Abschatzung
moglicher Khmawirkungen auf Walder im Alpenraum mittels eines Waldsukzessions-
modells diente als eine Fallstudie Die Methodik wurde in drei Schntten entwickelt, die
nachfolgend erlautert werden
Im ersten Schritt wurde die Kluft zwischen den raumlichen Skalen, auf welchen globale
Zirkulationsmodelle (GCMs) einerseits, und lokale Okosystemmodelle andererseits
openeren, uberbruckt Aufbauend auf das von VON STORCH et al (1993) vorgeschla-
gene Verfahren wurde an funf Schweizenschen Standorten fur den Winter und den
Sommer eine statistische Herabskaherungsprozedur entwickelt, uberpruft, und angewen-
det, um die zu den grossraumigen Klimaveranderungen (simuliert durch die GCMs)
zugehongen lokalen Klimaveranderungen abzuschatzen Gemass den sehr diversen und
anspruchsvollen Anforderungen verschiedener Okosystemmodelle wurden an jedem
Standort 17 saisonale Statistiken taghcher Wettervariablen zu Temperatur, Niederschlag,
Sonnenscheindauer, Luftfeuchte und Windgeschwindigkeit betrachtet Die Jahr-zu-Jahr
Schwankungen der lokalen Vanablen wurden mit Hilfe der kanonischen Korrelations-
analyse an simultane Schwankungen des Bodendruck- und Temperaturfeldes uber den
nordatlantisch-europaischen Sektor gekoppelt Die statistischen Modelle wurden separat
furjede Jahreszeit und jeden Standort fur die Penode 1901-1940 erstellt
In alien Fallen wurden physikahsch plausible statistische Modelle gefunden, welche die
lokalklimatischen Auswirkungen von Veranderungen grundlegender Zirkulationsmuster,
wie z B der Starke der Westwinddnft im Winter und der grossraumigen Subsidenz im
Sommer, quantifizierten Im Uberprufungszeitraum 1941-1980 wurden die meisten
Vanablen im Winter besser als im Sommer, und an den drei nordalpinen Stationen besser
als an den zwei sudalpinen Standorten rekonstruiert Die besten Resultate wurden fur
XV
die saisonalen Temperaturnuttel, die saisonalen Mittel der taghchen Temperaturextrema,
die winterlichen Niederschlagssummen, und die Anzahlen Niederschlagstage im Winter
erhalten Die saisonalen Mittel der taghchen Temperaturamphtuden, der relativen Luft¬
feuchte, und der Windgeschwindigkeit, sowie die innersaisonalen Standardabweichun-
gen der meisten taghchen Vanablen konnten im allgemeinen nur schlecht reproduziert
werden Das von VON STORCH et al (1993) vorgeschlagene Verfahren konnte jedoch
allgemein verbessert werden, indem zusatzhch zum Bodendruck die bodennahe Lufttem-
peratur als grossraumiger Pradiktor mitberucksichtigt wurde Die grossten Verbesserun-
gen wurden dadurch fur alle temperaturabhangigen Vanablen und fur den Sommer
erhalten Die Anwendung der statistischen Modelle auf zwei Klimaanderungssimula-
tionen mit dem ECHAM1/LSG-GCM (CUBASCH et at, 1992) ergab fur mehrere
wichtige Eingangsgrossen von Okosystemen zeitabhangige, raumhch, als auch zwischen
den Vanablen konsistente, regional stark differenzierte Abschatzungen moglicher Klima¬
veranderungen mit einer raumlichen Auflosung weit uber derjenigen heutiger globaler
oder regionaler Klimamodelle
Der zweite Schntt befasste sich mit der Herleitung statistisch praziser meteorologischer
Eingangsgrossen uber behebige Zeitraume, und unter behebigen Szenanen einer zeitab-
hangigen Klimaveranderung Hierzu wurde an 89 langjahngen europaischen Khma-
stationen die Gute 12 verschiedener stochastischer Modelle zur Simulation monathcher
Temperaturen und Niederschlage untersucht Die Modelle unterschieden sich in der Be-
handlung und Schatzung der Auto- und Kreuzkorrelations-Koeffizienten der Witterungs-
vanablen, in der Beschreibung der Jahreszyklen lhrer Erwartungswerte und Standard-
abweichungen, sowie in der Verwendung einer schiefereduzierenden Transformation fur
den Niederschlag Alle Modelle wurden anhand lokaler Messungen aus der Penode
1951-1980 erstellt Zur Evaluation der Modellgute wurden die aus unabhangigen Mes¬
sungen und aus stochastisch simuherten Wetterdaten berechneten Verteilungen der mini-
malen Wintertemperatur, der annuellen Tagesgradsumme, sowie eines Indexes fur den
annuellen Trockenheitsstress als Prufgrossen rmteinander verghchen
Es wurde gefunden, dass die Emfuhrung eines Gedachtmsterms in den stochastischen
Modellen die statistische Prazision der simuherten monathchen Temperaturmittel, mini-
malen Wintertemperaturen, und annuellen Tagesgradsummen generell verbessert Die
Verwendung einer lognormalen Transformation verbesserte die Simulation der monat¬
hchen Niederschlagssummen und des Trockenheitsstress-Indexes Es wurde gezeigt,
dass es moghch ist, zyklostationare autoregressive Modelle welter zu verbessern, und die
Anzahl benotigter Parameter erhebhch harabzusetzen, indem die monathchen Systemma-tnzen gemass dem Ansatz von HASSELMANN & BARNETT (1981) in emem Unterraum
geschatzt werden, der durch die ersten zwei Founer-harmonischen Funktionen des
Jahreszyklus aufgespannt wird Die Darstellung der Jahreszyklen der Erwartungswerteund Standardabweichungen der Witterungsvanablen mittels lhrer ersten zwei (Tempera-
XVI
tur) bzw drei (Niederschlag) Founer-harmonischen Funktionen ergab eine Verbesserung
nur fur den Trockenheitsstress-Index, jedoch nicht fur die beiden anderen bioklimati-
schen Vanablen Basierend auf den obigen Ergebnissen wurde ein neues stochastisches
Modell zur Simulation des monathchen Wetters vorgeschlagen Im Vergleich zu den
bisher ublichen Modellen benotigt dieses zwar eine etwas grossere Anzahl von Klima-
parametern, benutzt jedoch robustere Methoden zu deren Schatzung und produziert
significant realistischere Verteilungen bioklimatischer Vanablen
In einem dntten Schntt wurden die beiden vorangehenden Schntte nut dem Waldmodell
FORCLIM (BUGMANN, 1994, 1996, FISCHLIN et al, 1995) zu einer Gesamtmethodik
kombiniert, um quantitative Abschatzungen moghcher Klimawirkungen auf Gebirgs¬
walder nut einer hohen zeithchen (Jahreszyklus), raumlichen und qualitativen Auflosung
zu erhalten Die Methodik besteht aus den folgenden Schntten (1) Beschreibung des
Wetters am Okosystemstandort als einen stochastischen Prozess, (2) Schatzung der
Klima-/Prozessparameter fur das heutige Klima anhand von Messungen, (3) Verwen¬
dung einer statistischen Herabskalierungsprozedur, um aus einem GCM zeitabhangige
Veranderungen ausgewahlter Klimaparameter abzuschatzen, (4) Genenerung von Szena-
nen fur die monathche Witterung mittels stochastischer Simulation, (5) Verwendung der
simuherten Witterungsvanablen, um das Waldmodell anzutreiben
Die Methodik wurde ausgehend von einem nut dem Klimamodell CCC-GCMII (BOER et
al, 1992) berechneten Szenario einer globalen 2xC02 Khmaanderung an vier reprasen-
tativen Standorten in den Alpen angewendet Ahnlich wie beim ECHAM1/LSG-Modell
zeigten die herabskalierten Szenanen grosse raumliche und jahreszeithche Unterschiede,
sowie eine generelle Erwarmung, die grosser am Schweizenschen Alpennordhang als am
Alpensudhang ausfiel Es wurde eine generelle Tendenz zu mehr Niederschlag gefun-
den Fur eunge Standorte und Jahreszeiten wurde jedoch eine Niederschlagsabnahme nut
zum Teil schwerwiegenden Auswirkungen auf die simuherten Waldzusammensetzungen
erhalten Die Szenanen schienen physikalisch plausibel und raumlich konsistent Unter
ein und demselben globalen Szenario wurden stark gegensatzliche Waldreaktionen auf
engstem Raum erhalten Wahrend einige Waldsimulationen nur kleine Veranderungen in
der Artenzusammensetzung ergaben, wurden in anderen Fallen grossere Veranderungen
bis hin zu einem kompletten Zusammenbruch des Waldes simuhert In einigen Fallen
kamen ganzhch neue Artenzusammensetzungen, fur die kein heutiges Analogon existiert,
zum Vorschein Die erhaltenen Resultate stimmten mit dem momentanen Verstandnis der
Dynamik von Waldern uberein Sie legen nahe, dass einige Gebirgswalder sensitiv auf
eine 2xC02 Khmaanderung reagieren, und dass menschhche Unterstutzung notig sem
konnte, um die Anpassung der Walder zu unterstutzen
Zusammenfassend lasst sich sagen, dass die vorgeschlagene Methodik es ermoghcht,
lokale Okosystemmodelle auf konsistente Weise an die von globalen Khmamodellen
XVII
simuherten Klimaanderungen anzukoppeln Die Methodik ist generell, flexibel, rechne-
nsch effizient, hat einen nuttleren Datenbedarf, und erfullt die IPCC-Richthnien (CARTER
et al, 1994) fur Klimawirkungsstudien Ihre grosste Emschrankung ist, dass die ver-
wendeten statistischen Modelle unter einem zukunftigen Klima ihre Gultigkeit verlieren
konnten Die Methodik ermoglicht jedoch ausfuhrliche Analysen von Sensitivitaten und
Unsicherheiten, und dank ihres modularen Aufbaus konnen Verbesserungen der
einzelnen Komponenten gut nachtraghch eingebaut werden Die verwendete Gesamt-
methodik, um Waldreaktionen abzuschatzen nutzt die existierenden Modelle und Daten
gut aus, integrieit den momentanen Wissensstand, und eignet sich, um Auswirkungeneiner Klimaanderung auf behebige Walder der mittleren und hohen Breiten abzuschatzen
I
1. Introduction
Since the beginning of the Industrial Revolution the composition of the Earth's atmo¬
sphere has been increasingly affected by the activities of man. Steadily rising concen¬
trations of greenhouse-gases and aerosols in the atmosphere have probably already
significantly affected the redistribution of solar energy within the climate system
(SANTER et al., 1996; HEGERL et al., 1996). According to most projections global cli¬
matic change1 is likely to continue over the next decades to centuries at a rate and
magnitude unprecedented in the last lO'OOO years (IPCC, 1996a). Such a change can be
expected to have wide-ranging, partially irreversible, and potentially adverse effects on
ecological systems (IPCC, 1996b).
Global climate models (e.g., GATES et al., 1992) are the main source of information on
future climatic change, whereas ecosystem models are of major importance to study pos¬
sible impacts of climatic change on ecosystems. However, several problems occur if one
wishes to combine the two kinds of models for impact assessments: The global models
have a very coarse horizontal resolution (several hundreds of km), they simulate weather
only with a limited statistical precision (e.g., HULME et al., 1993), and most available
simulations cover only relatively short time spans (years to decades). This is in sharp
contrast to the needs of ecosystem studies, which often require local, and statistically pre¬
cise (e.g., NONHEBEL, 1994; FISCHLIN et al., 1995) inputs on weather or climate over
several centuries (e.g., PRENTICE et al., 1993) to millenia (e.g., PERRUCHOUD, 1996).
Due to the many uncertainties and unknowns involved in the projection of future climate
these inputs can generally not be provided as predictions, but only in the form of scenar¬
ios. Climatic scenarios are more or less plausible, internally consistent descriptions of
conceivable future space-time evolutions of the climate system. Several methods to
construct regional climatic scenarios have been proposed, which are reviewed e.g. in
GlORGI & MEARNS (1991), ROBOCK et al. (1993), and CARTER et al. (1994). The
advantages and limitations of the vanous approaches with regard to ecosystem studies are
discussed in sections 2.1, 3.1, and 4.1 of this thesis.
' The present study adopts the classical definitions of "weather" and "climate", according lo which the
former refers to the short-term (seconds to months) state of the atmosphere at the global to local scales,
and the latter to the statistics of weather over decades to centuries The term "climatic change" is used to
refer to a systematic shift in these statistics as caused by changes in the boundary conditions of the cli¬
mate system (e.g., changing land-surface conditions, or increasing concentrations of radiatively active
gases in the atmosphere). "Climatic variability" on the other hand refers lo the natural variability of
climate as caused by variations in forcings of the climate system (e g, solar inpul or volcanic activity)
and the internal dynamics of ihe global and regional climate
2 Chapter 1
The production of regional information from the output of global climate models (a
process often termed'
downscaling ) presents a key element in the construction of re¬
gional climatic scenarios A frequently used approach in ecosystem studies (e g,
BURTON & CUMMING, 1995, SAMPSON et al, 1996, BARROW et al, 1996) is to inter¬
polate climatic changes from a few grid points in the vicinity of the region of interest
However, as is discussed below in section 2 1, this procedure gives inconsistent results
(see also VON STORCH, 1995) Valid alternatives are the simulation of regional climates
with high-resolution global (e g ,CUBASCH et al, 1996) or regional (e g ,
GIORGI et al,
1992) climate models, or the utilization of statistical relationships between the large-scaleand the regional climate (e g ,
VON STORCH et al, 1993)
Regional climate models have the advantage that they are based upon first physical pnnci-
ples However, their enormous computing requirements severely restrict their useful¬
ness for scenario construction (see section 2 1) Furthermore, even if fast enough com¬
puters were available to run a climate model for the time spans and at the horizontal res¬
olution needed by ecosystem studies, some basic limitations would still apply Firstly, all
processes relevant to simulate local climates would have to be sufficiently well known
Secondly, the model would inevitably include empirical approximations of unresolved
processes These must be fitted to finite data sets, such that possible systematic errors
could feed into, and be amplified by the model, thus restricting its ability to correctly
simulate other climates than the present Finally, high-resolution models require accurate
boundary conditions, and therefore are difficult to test High-quality forcing fields for
the atmosphere and/or the ocean are available for some decades, but long-term data on,
e g, land-surface characteristics are already much more difficult to get
The statistical downscaling approach appears more suitable since it can provide directly a
local resolution, and because it can be expected to be computationally much more effi¬
cient A range of statistical downscaling procedures has been proposed, which are re¬
viewed in KATTENBERG et al (1996), GYALISTRAS et al (1998), and in section 2 1 of
this thesis However, the feasibility and suitability of statistical downscaling with regard
to ecosystem studies has not been systematically investigated yet
Moreover, it is also not clear how downscaling should be applied best to provide
scenanos with the needed temporal extension and resolution One possibility would be to
directly use downscaled time senes of weather variables2 However, as is discussed later
in section 3 1, this approach has several major limitations
2 The term weather variable or weather statistic is used to refer to statistics of a meteorologicalvariable over an individual day month or season (e g ,
the average temperature of a particular month or
the number of rainfall events experienced within this month) Climatic parameter refers to a statistical
Introduction 3
An often used alternative technique to simulate local weather in impact studies (e g ,
FISCHLIN et al, 1995, MEARNS et al, 1996) are so-called weather generators (e g ,
RICHARDSON, 1981) These are stochastic models of weather variables, which are fitted
to local measurements WlLKS (1992), OELSCHLAGEL (1995), and BARROW et al
(1996) have proposed procedures to derive scenarios of daily weather by combining
weather generators with the output of global climate models
However, these studies have demonstrated the construction of scenarios either only based
upon general indications from global climate simulations (WILKS, 1992), or only upon
very simple downscaling procedures which relied but upon a few climate model grid
points (OELSCHLAGEL, 1995, BARROW et al, 1996) Furthermore, as is shown in
section 3 1, until now only little work has been done at the monthly resolution, which is
particularly important for several ecosystem studies which operate on time scales of
centuries to millenia (see e g review in PERRUCHOUD & FISCHLIN, 1995)
The aim of the present thesis is to develop, test and apply a general method to construct
scenarios of future climate and weather suitable for ecosystem studies The assessment
of possible impacts of climatic change on forests in the Alps by means of a forest succes¬
sion model (SHUGART, 1984, KlRSCHBAUM & FISCHLIN, 1996) serves as a case studyto test the method The forest model poses only medium requirements to the construcion
of scenarios (see section 4 2 3) However, as is discussed later, several other ecosystem
studies have similar needs Moreover, the method is develloped in a modular manner,
such that it can be easily extended to meet additional requirements, and that all steps are in
principle applicable to any ecosystem model and any region (see Chapter 5)
The main questions addressed are
• Is it possible to derive new, or improve existing techniques to increase the physi¬cal plausibility and consistency of climatic scenarios needed by ecosystem studies9
• Can the scenarios be improved due to
- the use of statistical procedures to downscale the output of global climate
models to the regional scale9
- an improved stochastic simulation of the driving weather inputs9
• By applying such new, improved techniques, what scenarios of possible future
climatic changes and associated forest responses result for selected Alpine loca¬
tions and how plausible are they9
parameter used to describe weather as a stochastic process (e g the expected value of a weather variable or
the cross covariance of two weather variables)
4 Chapter I
The thesis contains three main chapters which correspond to individual papers already
published, or to be published in scientific journals It is structured as follows
Chapter 2 focuses on the consistency of climatic scenarios across spatial scales and pre¬
sents a statistical procedure to downscale climatic changes as simulated by global climate
models to the regional scale This chapter has been published with minor modifications
in GYALISTRAS et al (1994)
Chapter 3 focuses on the temporal aspect of scenario construction It deals with the simu¬
lation of site-specific weather and bioclimatic variables by means of stochastic models
In Chapter 4 the methods developed in two previous chapters are combined with a forest
model into an overall procedure which is applied to project possible impacts of a
changing climate on forests in the Alps This work has been published with minor
changes in FISCHLIN & GYALISTRAS (1997)
The proposed methods for scenario construction are discussed in Chapter 5, and the main
conclusions of the study are summarized in Chapter 6
5
2. The Spatial Aspect: Statistical Downscaling of
GCM-simulated Climatic Changes to the
Regional Scale
Published as
Gyahslras, D ,von Storch, H , Fischlin, A & Beniston, M (1994) Linking GCM-
simulated climatic changes to ecosystem models case studies of statistical downscaling in
the Alps Climate Research, 4 167 189
2.1 Introduction
In order to study possible impacts of climatic change on particular ecosystems models of
these systems are of paramount importance Although ecosystems and the atmosphere
interact on a multitude of temporal and spatial scales, most impact studies consider onlythe "one-way" forcing of the ecosystems by the atmosphere
Yet, in ecosystem studies a large variety of climatic input requirements occurs For ex¬
ample, in order to drive agroecosystem models (PARRY et al, 1988, SPITTERS et al,
1989, ROTH et al, 1995), or insect population models (LlSCHKE & BLAGO, 1990)
specific combinations of several meteorological variables such as temperature, precipi¬
tation, photosynthetically active radiation, wind, and air humidity, with an hourly to daily
time step, and for at least the duration of the vegetation season, are needed On the other
hand, forest gap models (SHUGART, 1984) simulating forest succession require
realisations of monthly mean temperatures and precipitation totals for every month of the
year (FISCHLIN et al, 1995), and SCHRODER'S (1976) population dynamics model for
red deer requires monthly snow-heights
To study and assess the impacts of climatic change, a given set of inputs needs to be de¬
rived from appropriate climatic change scenarios A scenario at a particular location and
moment should represent an internally consistent picture for changes in at least those
climatic parameters which are needed either as direct model inputs (FISCHLIN, 1982,
BRZEZIECKI et al, 1993), or to stochastically generate the weather variables that drive the
ecosystem model (e g ,MEARNS et al, 1984, SUPIT, 1986, SWARTZMAN & KALUZNY,
1987, WILKS, 1992, FISCHLIN et al, 1995)
6 Chapter 2
Climatic scenanos should extend from the present to an appropriate point in the future, or
otherwise cover time spans that are long enough for the purposes considered In some
cases, e g agroecosystems, a few decades can be sufficient When, however, soils or
forests are considered, which respond to climate on time scales of several centuries
(BUGMANN & Fischlin, 1992, 1994), the scenarios should specify the transient be¬
havior of regional climates for at least a few hundred years
Further, the climatic inputs are typically needed at specific representative locations, i e
with a spatial resolution of a few kilometres or less A high spatial resolution is partic¬
ularly important if ecosystems are studied within a complex topography such as the Alps
Three basic approaches could be adopted to derive the needed regional climatic change
scenarios (GlORGl & MEARNS, 1991) (1) Construct analogues of regional climatic
changes from past climatic situations as inferred from proxy data (FLOHN & FANTECHI,
1984, WIGLEY et al, 1986) or instrumental records (e g ,WlGLEY et al, 1980, PlT-
TOCK & SALINGER, 1982, LOUGH et al, 1984, HULME et al, 1990) (2) Use numerical
models, such as General Circulation Models (GCMs) or regional climate models (e g ,
DICKINSON et al, 1989, GlORGl et al, 1990) embedded in GCMs, to simulate possible
future climatic conditions with the best possible spatial detail (3) Use, as an intermediate
solution, semi-empirical approaches establishing a statistical relationship between larger-
scale and regional climatic changes (e g ,KIM et al, 1984, WILKS, 1989) which then can
be applied to derive climatic scenanos from the outputs of spatially coarse numerical
models (KARL et al, 1990, WlGLEY et al, 1990, VON STORCH et al, 1993)
A major advantage of the first, purely empirical, approach is that it is relatively simple to
implement However, paleoclimatic analogues normally provide information only on few
parameters, such as mean temperatures and, less often, precipitation (WlGLEY et al,
1986) In addition, the temporal resolution of proxy data is often coarse, and, even if
yearly data are available (e g from tree rings), the annual cycle may not be well resolved
(STOCKTON et al, 1985, SCHWEINGRUBER, 1988) Though this is not the case for
instrumental records, there still remains the problem that past climatic changes do not
necessarily reflect the effects of increasing greenhouse gas concentrations on climate
(GlORGl & MEARNS, 1991) Also, due to the lack of deterministic models it is difficult to
formulate gradual shifts in future climates in a physically consistent manner
The second approach relies upon GCMs Since GCMs reproduce most large-scale
atmosphenc and oceanic features, present - and even more so future - climate models are
regarded as being capable to reliably simulate at least the broad large-scale aspects of
man-made climatic change (DICKINSON, 1986, GATES et al, 1990) Simulations of
time-dependent changes for the next 100 years or so become increasingly available
(HANSEN era/, 1988, MANABEetal, 1991, 1992, CUBASCHeia/, 1992, GATES et
The Spatial Aspect Downscaling 7
16
12
8
4
0
-4
-8
(a) -12
Temperature (°C)i
i
1
i—4^- ***+\
_
• ^
z*?/£+ **^
oQ»-
ATemperature (°C) (deviations from GCM-"Control")
GCM-simulalions
—— "Control
—— "Scenario A"
(averages 2075-2084)
28
24
20
16
12
8-
4
M 0
Precipitation (cm/month)+—t
Measurements
(averages 1901-1980)
Bever
Bern
Davos
Lugano
Saentis
JFMAMJJASOND
Fig 2 1 Comparison between the climate simulated by the ECHAM 1/LSG-GCM in
the vicinity of the Alps (averages from three gndpoints) and observations (a) Monthly
mean temperatures The following annual mean values were subtracted prior to plotting"Control" and Scenario A" 8 7'C, Bever 1 5'C, Bern 8 4"C, Davos 3 0°C, Lugano11 7'C, Saentis -2 1-C, (b) Differences of the curves shown in (a) from the means of the
Control" experiment (c) Monthly precipitation totals
8 Chapter 2
al, 1992), thereby providing vast amounts of physically consistent data with a time step
of 1 h or less
However, a problem occurs due to the coarse horizontal resolution of most currently used
climate models, which is in the order of 500 km, and which sharply contrasts with the
fine resolution required by ecosystem models Furthermore, the minimum spatial scale at
which GCMs successfully reproduce the observed climate lies above several gndpoint
distances For present GCMs this is at least 2'000 - 4'000 km (VON STORCH et al,
1993) Climatic changes obtained at individual gndpoints seem even less trustworthy
over mountainous areas such as the Alps, which are not well represented in GCMs
This is illustrated in Fig 2 1 which compares the performance of the ECHAM 1/LSG-
GCM (described below in "Data and Methods - GCM Experiments") at three gndpoints
in the vicinity of the Alps with measurements from the five case study locations con¬
sidered in this study Though in the GCM-"control" experiment (simulation of present
climate) the observed annual cycle of temperature is reproduced qualitatively correct, with
a flat maximum in July/August and a minimum in January, the GCM generally overesti¬
mates the amplitude of the annual cycle by several degrees (Fig 2 1a) In particular,
note that the model error is similar in magnitude to the average changes simulated by the
GCM at the three Alpine gndpoints under the "IPCC Scenario A" (IPCC, 1990) for fu¬
ture atmospheric concentrations of greenhouse gases (Fig 2 lb) Precipitation is
severely underestimated at all locations during the summer months Here the projected
changes in gndpoint precipitation under the "Scenario A" are much smaller than the
deviations typically found between the observations and the "control" (Fig 2 lc)
One possibility for improving the simulation of regional climates is to increase the
horizontal resolution of the GCMs or, at least, to use them to drive simulation models
with enhanced resolution over areas of interest However, computational costs of climate
models increase at least quadratically with the spatial resolution (GlORGl & MEARNS,
1991) It may therefore be expected that, even if more powerful computers become
available, the simulation of the full annual cycle over several decades, let alone centuries,
by high-resolution models will not be possible for several years to come Furthermore,
even if gndpoint distances for long-term integrations could be reduced to 100 km or less,
there would still remain a substantial gap between the scales at which climate models
operate reliably and the local detail required by most ecosystem models
An alternative is to use semiempincal approaches which allow site-specific empirical data
to be combined with the physically consistent results of numerical climate models Semi-
empirical approaches are particularly attractive, since, in principle, they can be flexibly
used at any location and for any variable of interest Moreover, due to their computation¬
al efficiency they would even allow transient scenarios far into the future to be derived
The Spatial Aspect Downscaling 9
To date, several semiempncal approaches have been proposed For example, WlGLEY et
al (1990) used monthly means of several regionally averaged atmospheric variables to
predict vanations of monthly mean surface temperature and precipitation at the 32 climate
stations within the region considered, KARL et al (1990) related daily gndpoint data of
22 free-atmosphere variables to daily temperature extrema, precipitation and cloud
ceilings from local surface observations, VON STORCH et al (1993) used anomalies of
large-scale mean sea-level pressure over the Atlantic to predict, using Canonical Correla¬
tion Analysis (BARNETT & PREISENDORFER, 1987) changes in winter mean Iberian pre¬
cipitation, finally, WERNER & VON STORCH (1993) used basically the same approach to
relate January-February mean temperatures from 11 central European stations to the
large-scale circulation
The method of VON STORCH et al (1993) appears to be superior, because, unlike the
other approaches, it considers the large-scale (several 103 km) behavior of a climatic
parameter, according to the assumption that predictions of future climates by GCMs are
most credible within this resolution Also, unlike the approach by Karl et al (1990),
the statistical models obtained do not depend on a particular GCM and can thus be applied
to other GCMs However, the method of VON STORCH et al (1993) has so far been ap¬
plied only to predict one climate variable at a time, t e rainfall or temperature, and only
for winter
Thus the following questions arise Is it possible by means of this method to establish
plausible statistical relationships which allow multivariate descriptions of local climates,
even in a complex orography like the Alps, to be linked to large-scale climatic changes7
If yes, for which ecosystem inputs, seasons and locations can the best and for which the
poorest results be obtained9 What improvements could be gained by modifying the
method, for instance by including large-scale near-surface temperature anomalies as an
additional predictor9 Finally, is it possible to derive regionally varying scenarios from
GCM-projected global climatic changes9
Here we use for the first time the method by VON STORCH et al (1993) to downscale
large-scale climatic changes - at a seasonal resolution, not only for winter, but also for
summer - to 17 local ecosystem inputs at several point locations within a complex topog¬
raphy, i e at five representative case study locations in the Alps We show that the
method proposed by VON STORCH et al (1993) not only yields physically plausible
results for all case study locations, but that it can also be substantially improved, in
particular for temperature-related variables Finally, by applying the obtained statistical
relationships to two climatic change experiments with a GCM, we demonstrate that our
approach allows to derive regionally differentiated, transient climatic scenarios of use in
ecosystem studies
10 Chapter 2
2.2 Data and Methods
2.2.1 Observations
As independent variables (predictors) we tested mean sea-level pressure (SLP) and 2 m
above-ground air temperature for which long-term (several decades) observed data sets
are available The independent variables were given on a 5° x 5° latitude by longitude
grid containing 153 gndpoints and extending from 40° W to 40° E and 30° N to 70° N
(Fig 2 2a)
Winter (December, January, February) and summer (June, July, August) mean SLP for
the years 1901-1980 were calculated at each gndpoint from daily data provided in the
NCAR data set (JESSEL, 1991) Seasonal mean near-surface (2 m above-ground) tem¬
peratures were derived for the same period from monthly data given on a 10° x 5° latitude
by longitude grid (VINNIKOV et al, 1987, JESSEL, 1991), and were then interpolated to
the 5° x 5° standard grid Missing data for some years and gndpoints were replaced by
interpolation between adjacent timepoints
40°W 20°W 0° 20°E 40°E 5°E 10°E 15°E
(a) (b)
Fig 2 2 (a) Gndpoints of sea-level pressure and near-surface temperature fields used in
this study (b) The five Swiss locations considered
The SLP data were initially given as absolute values, whereas the near-surface tempera¬
ture data were given as anomalies from the three different basis periods 1881-1935,
1881-1940 and 1881-1960 However, since fitting of the statistical models required data
with zero means, anomalies were calculated for both predictors relative to the mean states
of the respective years used (see also "Statistical Procedure", below)
The Spatial Aspect Downscaling II
On the regional scale, five different locations within Switzerland, representative of north,
inner and south alpine climates were considered (Fig 2 2b) (1) Bern [valley, 565 m
above sea level (m a s 1), 7 4° E, 46 9° N] at the northern slope of the Alps, (2) Davos
(pronounced valley location, 1590 m a s 1,9 8° E, 46 8° N) in the central alpine region,
(3) Bever (1712 m a s 1,9 9° E, 46 6° N), located in the Oberengadine valley which is
characterized by a central or south alpine climate, (4) Lugano (valley, 273 m a s 1, 9 0°
E, 46 0° N) at the southern slope of the Alps, and (5) Saentis (mountain peak, 2500
m a s 1, 9 2° E, 47 2° N) which represents "free-atmosphere" conditions at the northern
side of the Alps
Variable Seasonal Statistics Explanations
Temperature (a) Tmean Tmm, Tmax, Tampl
(b) s(Tmean), s(Tmin)
s(Tmax), s(Tampl)
(a) means, (b) standard deviations of daily mean,
minimum and maximum temperatures and of daily
temperature amplitude, respectively (in *C)
Precipitation (a) Precip
(b) s(Precip)
(c) Ndays>1mm
(a) mean (b) standard deviation of daily precipita¬
tion totals (in cm),
(c) number of days with total 2 1 mm
Relative sunshine
duration
(a) RelSunsh
(b) s(RelSunsh)
(a) mean (b) standard deviation of daily average
from three measurements per day (in %)
Relative humidity (a) RelHum
(b) s(RelHum)
(a) mean, (b) standard deviation of daily average
from three measurements per day (in %)
Wind speed (a) Wind
(b) s(Wmd)
(a) mean (b) standard deviation of daily average
from three measurements per day (in m s"')
Table 2 1 Meteorological variables and the related seasonal statistics considered in this study
The dependent variables (predictands) were given at each location by a vector of 17
seasonal weather statistics (Table 2 1) Special emphasis was given to the main weather
elements temperature and precipitation, for which 11 variables were defined Relative
sunshine duration was included as a variable allowing us to approximate incoming photo¬
synthetically active radiation, which determines rates of plant growth in many ecosystem
models As additional parameters we considered relative air humidity and wind speed,which are used particularly in agroecosystem models Within-season standard deviations
of all meteorological variables were included, since these parameters are typically needed
to simulate daily weather conditions by means of stochastic weather generators Since
precipitation distributions often show considerable skewness, we used, as an additional
parameter for the wtthtn-season variability of rainfall, the number of days with precip¬
itation > 1 mm, which is proportional to the probability of "wet days" within the season
The threshold value of 1 mm was chosen because it represents a lower limit for the
detection of rainfall valid for most measuring devices (cf UTTINGER, 1970)
12 Chapter 2
All seasonal statistics were derived from daily 1901 to 1980 measurements, extracted
from the database of the Swiss Meteorological Institute (SMA), Zurich (SMA, 1901-
1980, BANTLE, 1989) Daily mean temperatures, relative sunshine durations, relative
humidities and wind speeds, were calculated from three measurements per day, whereas
daily temperature minima and maxima were determined as the extrema from the three
daily measurements and the evening temperature of the previous day After 1971 the time
at which the evening temperature measurements were taken was changed from 21 30 h to
ca 19 00 h, so daily means for the years 1901-1970 and 1971-1980 were calculated
according to different procedures (BANTLE, 1989) The effects of this change in proce¬
dure are still under investigation by the SMA, for locations with large daily temperature
amplitudes it is likely that, due to the new procedure, daily mean temperatures are
overestimated by ca 0 2° C (A DE MONTMOLLIN, SMA, personal communication)
2.2.2 GCM-Experiments
Simulated seasonal mean SLP and near-surface temperature were obtained from ex-
penments performed with a fully coupled global atmospheric/oceanic GCM run at the
Max-Planck Institute for Meteorology (MPI), Hamburg The atmosphere component
(ECHAM 1) of the model is a version of the spectral numerical forecasting model of the
European Centre for Medium Range Weather Forecasts, extensively modified at the MPI
Hamburg for climate simulations Its horizontal resolution is limited by a cut-off at wave
number 21, corresponding to a gaussian grid with an average spacing of 5 6°, and its
vertical resolution is given by 19 levels in a hybrid o-p coordinate system The ECHAM 1
model simulates the diurnal cycle with a time step of 40 min The ocean component
(LSG), which allows for the calculation of large scale geostrophic motions, is based on
11 vanably spaced vertical levels, an effective horizontal grid size of 4° and a basic time
step of 30 d, except for the two uppermost ocean layers, where a time step of 1 d is used
A more detailed descnption of the coupled model is given in CUBASCH et al (1992)
All GCM-simulated data used in this study were derived from daily GCM-outputs and
were interpolated on the 5° x 5° grid of the observed data sets We considered results
from (1) the'
control" integration with constant 1985 atmosphenc C02-concentratton, (2)
the "double CO2" (DC02) experiment, representing the model system's response to an
immediate doubling of greenhouse gas concentrations from 360 to 720 ppm equivalent
CO2, and (3) the "IPCC Scenario A" (SCNA) experiment, where the GCM is forced by
continuously increasing greenhouse gas concentrations according to the "Business-As-
Usual" scenario given by the IPCC (1990) The model's performance in the 'control"
run and its responses to the greenhouse gas forcings are addressed later
The Spatial Aspect Downscaling 13
2.2.3 Statistical Procedure
Our method is based upon a perfect prognosis approach (Glahn, 1985, von STORCH
et al, 1993) a statistical relationship is first established between the large-scale and the
local observations, and then applied unmodified to predict changes in the local variables
from any large-scale observations or GCM-simulated data sets For this purpose, we
considered each season and location separately
The overall procedure consists of (1) the selection of two subpenods, one to fit and one
to verify the statistical models, (2) the calculation of anomalies, plus the weighting of the
predictor and predictand variables, (3) the estimation of a series of differently specified
statistical models linking the predictands to the predictors, (4) the verification of all
specified models with independent observations, (5) the selection of a subset of the best-
performing models, and finally (6) the application of the selected models to GCM anoma¬
ly-fields (Fig 2 3)
In the standard procedure we used the years 1901-1940 for statistical modelling and the
years 1941-1980 for model verification In order to test the sensitivity of the resulting
climatic scenarios to the choice of the estimation period, the two subpenods were re¬
versed at a later stage
All variables were transformed to deviations from the mean states of the years used for
model estimation This is because only anomalies of the large-scale and local variables
were related to each other, whereas the long-term means were not affected by the pro¬
cedure On the local side, this ensured that climatic change scenarios can be specified in a
manner consistent with present long-term mean climate at each location On the side of
the large-scale predictors only the differences occurring between a climatic change
experiment and the control experiment of a GCM were considered This avoided the
problem that the long-term mean fields simulated by GCMs are normally biased with
respect to observations
The predictor variables were weighted with the square root of their latitude cosine, in
order to account for the latitudinal variation of the grid box sizes of the rectangular 5° x 5°
grid When both, SLP and near-surface temperature, were used as predictors, each field
was rescaled with a constant factor in order to account for a predefined proportion (e g
half) of the total predictor variance For graphical representations of predictor patterns,
however, all weightings were removed, such that the same unit applied to all variables of
the same kind The local weather statistics were weighted with their observed standard
deviations from the years used for model estimation, thus eliminating the effects of
different value ranges
14 Chapter 2
1941 80 observed
mean SLP and 2mT
(anomalies from
1901 40 mean state)
GCM simulated
mean SLP and 2mT
(anomalies from
mean state of
control experiment)
Predictors
seasonal mean SLP and 2mT
(years 1901-40)
Predictants
17 seasonal weather statistics
(years 1901 40)
I
weighting,calculation of anomalies,
calculation of EOFs
standardization,
calculation of EOFs
'*zjf
*»+selection of first few EOFs
(one from 169 different combinations
using 2 to 14 predictor/predictand EOFs)
\W_
4*M
Canonical Correlation Analysis
(CCA)
application of CCA-model
(using all canonical correlations > 0 33)
"fEverification of CCA-model
(performance index mean of all
prediction skills > 9%)
1941 80 observed
weather statistics
collection of results from 84
best performing models
T TPredicted Predicted
weather statistics weather staustics
(for years 1901 80) (from GCM experiments)
Fig 2 3 Block diagram of the method used to derive scenarios of local climatic changes
from large scale climatic changes as simulated by GCMs SLP sea-level pressure 2mT
2 m near-surface temperature, EOF = empirical orthogonal function (see
text) Underlined texts = input/output data sets, boxes = procedures, solid lines with
arrows = flow of data, dashed line with arrow = iteration over procedures within grey box
For further explanations see text
Estimation of the statistical models consisted of two steps First, Principal Component
Analysis (PCA, e g , KUTZBACH, 1967, PREISENDORFER, 1988) was used to reduce
the dimensionality of the observed data sets and to remove linear dependencies between
vanables within the same data set PCA determines from a set ofn vanables v,^ (i=l n)
a set of new, mutually uncorrelated vanables pc,^ subject to the constraint that the van-
ance of the new variables is subsequently maximized The new vanables are named prin¬
cipal components (PCs) and are given by the the coordinate transformation pc,(t)=
(21'>XE'uxJ(t)+ZE''JxJ(0
sKt)=J=154J=l
I1+ZE''JxJ(0
306153
signalstwoofsumtheasinterpretedbethencans,
coefficienttimeassociatedTheP,.elementscorrespondingthebydefinedmapsseparate
twoofcomposedasvisualisedwasP,patternindividualanAccordingly,306)(j=154
anomaliestemperaturenear-surfaceand153)(j=lanomaliesSLPintovided
subdi¬bemayx.thepredictors,astemperature,near-surfaceandSLPboth,usingWhen
setsdatarespectivetheofvariance
remainingtheofmaximumaexplaintousedbecannc),(m=2pmfactorawithrelating
cor¬coefficients,timeof(sm,tm)pairfurtherEveryptcorrelationcanonicalmaximuma
shows|),ti(scoefficientstimeofpairfirstthethatbutuncorrelated,mutuallyaret,allas
wellass,allthatconstrainttheunderCCAbydeterminedareQ,andP,Therespectively
k=I
2)(2yk(o,a'lk2
1)(2xj(«).P-'uX
>(t)
and
5<(0
equationsthetoaccordingnc)1(i=t,(tjands,(,)coefficientstimecanonicaltheables,
vari¬ofsetsnewtwoyKandx.thefrominfertousedwerepatternscanonicalThemode
canonicalatocorrespondspatternsofpairEachrespectivelypredictands,anddictors
pre¬thefornc)(i=lQ,andP,patternscanonicalncyieldingspace,physicalthetoback
EOF-thefromtransformedwerecombinationslinearthesedeterminingscalarsThePCs
predictandtheofcombinationslinearrespectivewithbestcorrelatewhichPCspredictor
theofcombinationslinearNy)Min(Nx,=ncidentifiesCCAPCs,usedallofmatrixance
covari¬squaredtheofeigenvectorstheonBased1987)PREISENDORFER,BARNETT&
1984,ANDERSON,,ge(CCA,othereachtoAnalysisCorrelationCanonicalofmeans
byrelatedwerePCspredictandNyfirsttheandpredictorNxfirstthestep,secondaIn
ny)(k=lyk(tjpredictandsnytheandnx)(j=lx./t)predictorsnxtheforrately
sepa¬performedwasPCAThev,^theofmatrixcovariancenxntheofeigenvectorsthe
bygivenareEOFsThe(EOFs)functionsorthogonalempiricalcalledareand3ininbasis
orthogonalanformeofnl,{eofj,vectorsThevn,t))',(V|((\,=v/(jwhereV/(),eof,
15DownscalingAspectSpatialThe
16 Chapter 2
Due to the respective orthogonality of the s, and t,, the element P,, (or Q_1K) of the i-th
canonical pattern represents the covanance between the j-th predictor (the k-th predictand)
and the dimensionless coefficient s, (t,) Local maxima (minima) in the predictor maps of
the i-th canonical mode thus represent regions where positive (negative) anomalies in the
respective x, contribute with large positive (negative) anomalies to the coefficient s, (cf
Eq 2 1 or 2 1') Similarly, a large positive (negative) value Q1K denotes that for the i-th
canonical mode a large positive (negative) anomaly of the k-th weather statistic occurs, if
the time coefficient t, takes a positive value Finally, the canonical correlation p, between
s, and t, measures the strength with which anomalies in the predictands and predictors are
related to each other within the respective canonical mode
The results of CCA can be used to predict linearly, from any anomaly fields describing
changes in the predictors x., the simultaneous responses of all local weather statistics yk
(k=l ny), according to
"ai "cu0s
y*k(t) = X'*i(i)Qik = X p> o7 s'(t)Qik (2 3)i=l i=l
'
Here, ncu < tic is the number of canonical modes used, t,* denotes the estimated i-th time
coefficient of the dependent variables, and oSl, at, are the standard deviations of the time
coefficients s,(t) and t,(t), respectively For each CCA model fitted, we chose ncu
individually as the number of all canonical correlations p, above the threshold value % =
0 33 For the choice of nc we assumed that serial correlation in the predictors and predic¬
tands can be neglected due to the tnterannual variations considered Given that CCA
systematically overestimates the true p,, nc was chosen larger than the critical value of
0 26 (0 31) above which an unbiased estimate of a correlation coefficient would be
significantly different from zero at the 90% (95%) confidence level (two-sided t-test,
n=40, e g , KREYSZIG, 1977)
Performance of CCA may sensitively depend on the numbers Nx and Ny of EOFs
considered Using too many EOFs will fit the statistical models too strongly to the
particular data sets considered, most likely missing an adequate description of the
underlying stochastic process A too small number of EOFs, on the other hand, will omit
part of the significant signal, thus resulting in a poorer predictive ability of the overall
model In the present study, instead of retaining a significant, fixed number of EOFs
based on EOF-selection rules (e g ,PREISENDORFER et al, 1981) we focused on the
effects resulting from possible alternative choices for Nx and Ny To this purpose we
performed CCA for the 132=169 combinations resulting when varying Nx and Nyindividually within the range of 2 14 The upper limit of 14 EOFs was chosen because
The Spatial Aspect Downscaling 17
from this number on typically more than 90% of the total variances of the predictor and
predictand data sets could be accounted for, for all seasons and locations considered
Each of the 169 CCA models was then applied to predict the local variables from the
large-scale data not used for model estimation, and model performance was assessed
according to
^(e*. ey) = ri Max ( °' Cor[ yk<0' y*«t) 1 " "v ) (2 4)y
k=l
4* measures the mean of all correlations above the threshold value n.v = 0 3 between the
observed (y^) and reconstructed (y*\J weather statistics in the 40-year venfication period
The value chosen for r)v can be justified similarly as was done for nc above
The 84 (half of 169) best-performing models were then used to evaluate the performance
of the procedure for the individual weather statistics, and to derive scenanos of climatic
change from the GCM-expenments
2.3 Results and Discussion
2.3.1 Model Estimation
We investigated five different variants of predictor data sets, with the SLP and near-sur¬
face temperature fields accounting for the total variance of all predictors as given by the
ratios 1 0 (i e, only SLP), 2 1,11,12 and 0 1 (only temperature) For the predictors
being SLP alone, SLP and near-surface temperature with equal weight, and temperature
alone, respectively, the canonical patterns and time coefficients of individual CCA models
were graphically represented and visually compared with each other and between
locations
In this subsection we mainly present results using Bern as an example, since results ob¬
tained for Bever, Davos, Lugano and Saentis could be interpreted similarly to those for
Bern We also present only CCA models fitted to a combined predictor data set (SLP and
near-surface temperature with equal weight), because the canonical patterns resultingwhen using only one field were found quite similar to the respective subpatterns of the
combined approach This was generally the case for both seasons considered
18 Chapter 2
A) WINTER
Fig. 2.4 shows the CCA model obtained for Nx = Ny = 4 for Bern in winter. The first
predictor pattern consists of a large positive SLP anomaly centred over the United
Kingdom, a negative temperature anomaly over land areas, and a positive temperature
anomaly over the Atlantic with an increasing amplitude towards the northwest. The two
subpatterns give a consistent picture: in winters with on average higher than normal pres¬
sure over the UK, mean westerly flow and thus advection of relatively mild maritime air
towards the continent are less pronounced, such that temperatures over lands drop below
their long-term means. At the same time, advection of polar air into the north-west
Canonical Correlation Patterns B 1 Q 2 Variances expl. by Canonical Modes H 1 Q 2
Fig 2.4 CCA model for Bern in winter (a) First, (b) second canonical patterns for the
predictors Left sides of (a) and (b) SLP subpatterns, contour interval 1 mb Right sides
near-surface temperature subpatterns, contour interval 0 2"C (c) Canonical patterns, (d) ex¬
plained variances for the local weather statistics The first four predictor EOFs and the first
four weather statistic EOFs were used, explaining ca 75% and 77% of the respective tolal
variances Correlations between time coefficients are 0 89 for the first and 0 82 for the
second mode.
The Spatial Aspect Downscaling 19
Atlantic is also reduced, resulting in above-normal temperatures in this region This was
the dominant predictor pattern in winter for all locations considered
The first canonical pattern of the weather statistics in Bern (Fig 2 4c, solid bars)
confirms the above interpretation temperatures are anomalously low, as are precipitation,
its standard deviation, and the number of days where precipitation exceeds 1 mm The
negative responses in the precipitation-related variables are plausible, indicating that
advection of maritime air is the main source of winter precipitation in Bern Since the
sign of the canonical patterns is arbitrarily determined by CCA, note that the following
interpretation also holds a negative SLP-anomaly centred over the UK, indicating inten¬
sified mean westerly flow, is associated with anomalously high temperatures over the
continent, and also implies positive temperature and precipitation anomalies in Bern
The relevance of the large-scale patterns for the individual weather statistics can be seen
from Fig 2 4d Shown are the proportions of the variances in the local variables which
were explained by the time coefficients t|(tj (solid bars) and tji^ (shaded bars), l e the
squared correlation coefficients (in percent) between the yk(t) and t,(t) for the model esti¬
mation period 1901 to 1940
The second canonical mode mainly explained the variability of winter mean daily tempe¬
rature extremes and amplitudes, mean relative sunshine duration, and its within season
standard deviation (Fig 2 4d, shaded bars) The large-scale pattern depicts intensified
mean advection of warm air from the west/south-west (Fig 2 4b, left), and thus elevated
temperatures over the continent with a maximum over eastern Europe (Fig 2 4b, right)For Bern this setting implies slightly milder winters with on average more sunshine and
higher-than-normal daily temperature amplitudes (Fig 2 4c, shaded bars)
The third and fourth canonical modes (not shown) also contnbuted to explaining the year-
to-year variability of the weather statistics The third predictand pattern denoted predom¬inance of meridional flow from north/northeast, which at Bern was associated with
below-normal temperatures and mean wind speed, with higher-than-normal within-
season temperature variability, and increased relative humidity The fourth predictand
pattern showed a large positive SLP anomaly over mid-latitude Atlantic and correspond¬
ing large-scale positive near-surface temperature anomalies over Finland, it correlated
mainly with negative deviations in the withm-season standard deviations of daily temper¬ature means and extremes The third and the fourth mode showed canonical correlations
of 0 55 and 0 37, respectively, such that they also contributed to the reconstruction of the
weather statistics in the model venfication phase
The canonical conelation patterns obtained for Davos (Fig 2 5) illustrate the regionallydifferentiated relationships established by means of CCA For example, in contrast to
20 Chapter 2
Bern, year-to-year variations of winter mean precipitation, its within-season standard
deviation, and of the probability of precipitation events > 1 mm (Fig. 2.5d, shaded bars)
were explained in Davos mainly by the second canonical mode (cf. Fig. 2.5b), not by
the mean strength of westerly flow. In addition, though the respective SLP and temper¬
ature subpatterns are qualitatively similar for Bern and Davos, regional differentiation
may also be given by shifts in the locations and/or amplitudes of the pattern's troughs and
ridges. Due to sampling uncertainties however, it could be that not all respective patterns
are significantly different from each other.
Canonical Correlation Patterns I I D 2 Variances expL by Canonical Modes | | Q 2
Fig. 2 5 CCA model for Davos in winter, (a) First, (b) second canonical patterns for
the predictors. Left sides of (a) and (b) SLP subpatterns, contour interval 1 mb Rightsides, near-surface temperature subpatterns, contour interval 0.2°C. (c) Canonical patterns,
(d) explained variances for the local weather statistics The first seven predictor EOFs and
the first four weather statistic EOFs were used, explaining ca 88% and 73% of the
respective total variances Correlations between time coefficients are 0 91 for the first and
0.79 for the second mode
The Spatial Aspect Downscaling 21
B) SUMMER
Fig. 2.6 shows the results of CCA with Nx = 4 and Ny = 3 for Bern in summer. Again,
the first predictor pattern depicted here for Bern was typical for all other locations as well.
Its SLP-subpattern (Fig. 2.6a, left) shows a small positive anomaly over Europe (ca. 0.7
mb) when compared to the 1901-1980 standard deviation of mean summer SLP, which
ranges from approximately 0.8 mb over the Mediterranean area to ca. 2 mb over the UK
and Scandinavia (not shown). Considering that in the long-term mean the fringe of the
Azores high-pressure system reaches Eastern Europe (not shown), the pattern can be in-
Fig 2 6 CCA model for Bern in summer (a) First, (b) second canonical patterns for
the predictors Left sides of (a) and (b) SLP subpatterns, contour interval 0 5 mb Rightsides near-surface temperature subpatterns, contour interval 0 1°C (c) Canonical patterns,
(d) explained variances tor the local weather statistics The first four predictor EOFs and
the first four weather statistic EOFs were used, explaining approx 75% and 77% of the
respective total variances Correlations between time coefficients are 0 89 for the firsl and
0 75 for the second mode
22 Chapter 2
terpreted to reflect summers with less pronounced subsidence over the Azores, but with
the Azores high extending further than normal into the continent, thus leading to predomi¬
nantly stable, warm weather conditions over central Europe An alternative explanation
could also be, however, that the SLP-subpattern represents a weak circulation induced by
anomalously high temperatures over central and southern Europe
Consistent with both interpretations, the first canonical mode predicts for Bern positive
anomalies for temperature and relative sunshine duration, and a smaller probability for
precipitation events > 1 mm (Fig 2 6c, solid bars) Interpreted with opposite signs, the
first canonical mode describes summers with a higher-than-normal SLP over the Azores,
but a corresponding low SLP over Europe, denoting that the continent is more frequently
exposed to intrusions of cold air from west/northwest This results in anomalously low
temperatures at Bern, as well as over large areas southeast of the anomalously low SLP
The second mode described mainly changes in the precipitation-related variables
(Fig 2 6d, shaded bars) similar to the situation in winter, higher-than-normal SLP
centred over the UK (Fig 2 6b, left) denotes summers with generally weaker westerlies
and reduced advection of maritime air into central Europe, whereas the associated
intensified northwesterly flow implies negative near-surface temperature anomalies over
eastern Europe (Fig 2 6b, nght) In Bern, precipitation-related variables show negative
deviations, mean daily temperature amplitudes as well as relative sunshine durations are
above their long-term means, and, possibly due to reduced thunderstorm activity, the
less pronounced zonal flow leads to negative deviations for summer mean wind speeds
(Fig 2 6c, shaded bars)
Clearly, since summer weather in central Europe is mainly influenced by smaller-scale
convective processes, the second canonical mode describes but a relatively small part of
the interannual variability of the above mentioned variables (see also "Model Venfication"
below) Nonetheless, this mode can be interpreted to quantify for Bern the effects result¬
ing from fluctuations in the intensity of the European summer monsoon, l e the strength
of the enhanced westerly winds over central Europe which typically onset in mid-June
(SCHUEPP & SCHIRMER, 1977)
2.3.2 Model verification
For all five variants of SLP/near-surface temperature predictor data sets we determined
per season and location, according to Eq 2 4, the respective best-performing 84 of 169
fitted CCA models For each set of 84 CCA models, the models capabilities to predict
the individual weather statistics were assessed by comparing the statistically reconstructed
The Spatial Aspect Downscaling 23
interannual variations of the local variables with the observations For six main weather
statistics we also analysed the resulting 80 yr linear trends
A) INTERANNUAL VARIABILITY
The skill of the CCA models with regard to the individual variables was measured by the
proportions of variance explained (100 r,2, i = 1 84) in the verification period 1941-
1980 Under the assumption of a normal distribution, correlations above 0 31, l e skills
above ca 10%, suggest a statistically significant link to large-scale climate (a=95%)
Fig 2 7 shows that the skill of the CCA models may vary strongly, depending on the
large-scale predictors, the season, location, and weather statistic considered
Though no combination of predictors could be found which yielded optimal results for all
cases, the combined use of SLP and near-surface temperature generally improved the
skill of the procedure (predictor data set variants 2-4 in Fig 2 7) On average over all
locations and variables, the mean variances explained when SLP (near-surface tempera¬
ture) alone was used, amounted to ca 23% (21%) in winter and to 12% (18%) in
summer An optimum was reached when the two fields were combined with equal
weights, such that an average skill of 25% in winter and 19% in summer was attained
Use of the two predictors with equal weights seemed therefore to be a good compromise,
which was maintained for the further discussion and for obtaining climatic change
estimates from the GCM-expenments
As can be seen from Fig 2 8, the most dramatic improvement occurred for the summer,
and for temperature related variables However, the situation remained, that most
variables were better linked to the large-scale state of the atmosphere in winter than in
summer
In both seasons, better results were obtained at the three north alpine locations (Bern,
Davos, Saentis) where, on average over all three locations and all variables, the mean
skill of the CCA models amounted to 29% in winter and to 21% in summer At the more
southern locations (Bever, Lugano) on average only 20% (winter) and 17% (summer)
were attained
Large differences were found between the individual variables (Fig 2 8) Temperatures
were generally better reproduced than precipitation-related variables, the latter being
particularly poorly predicted in summer For seasonal mean temperatures, mean skills
from the respective 84 selected CCA models were between 44% (Bever) and 67%
(Saentis) in winter, and 56% (Bever) and 78% (Saentis) in summer (Fig 2 7a, predictor
24 Chapter 2
(a)
(b)
at
70
60
50
40
30
20
70.
60-
50
40
30
20'
.S-U
(c) 8 S
70
60
50
40-
30-
20
5 5 §
*****
5
Hit
« * i
Hi
(1) (2) (3) 14) (S)
Bever
* i
}M5
* I* I
* * I
a 5
i *
Hi
\i h
? ? i
-is-
(1) (2) <3> <4> (S) (1) (2) (3) 14) (S)
Bern Davos
a) <i) m t4) (5)
Lugano
Winter o Summer
IS,5«
* * «»
jlJ_
liU(I) (2) (3) 14) (S)
Saentis
Fig. 2.7. Skill of the CCA models as a function of the predictors, season, location and
variable considered Shown are percentages of variances (=100 r2) explained by the
reconstruction of the weather statistics in the verification period 1941-1980 Dala pointsdenote the mean variance explained by 84 selected CCA models, error bars are ±1 standard
deviation. From left to right (repeated for each location). (1) predictors were SLP
anomalies alone, (2) predictors were SLP- and near-surface temperature anomalies
accounting for the total variance of the predictor data with the ratios 2 1, (3) 1 1, (4) 1 2,
and (5) predictors were near-surface temperature anomalies alone.
variant 3). Mean daily temperature minima, which were better reproduced in summer
than in winter at all locations except for Bern (not shown), were predicted with mean
skills ranging from 30% (Lugano) to 64% (Saentis) in winter, and from 36% (Lugano) to
73% (Saentis) in summer. Mean daily maximum temperatures were also generally well
predicted, with skills ranging in winter from 59% (Bever) to 69% (Bern), and in summer
from 33% (Lugano) to 74% (Saentis). Mean daily temperature amplitudes were not well
reproduced at any location in winter, but in summer skills of 19%, 24% and 32% were
attained at Bever, Saentis, and Bern, respectively. For precipitation, between 29%
(Saentis) and 55% (Bern) were reached in winter, and between 10% (Lugano, Saentis)
and 28% (Davos) in summer (Fig. 2.7c). Relative sunshine durations, which directlydepend on degree of cloudiness, were also less well reproduced; exceptions, however,
The Spatial Aspect Downscaling 25
Winter Summer
Tmean
T min
T max
T ampl
Precip
Wind
Rel Hum
Rel Sunsh
s( T mean )
s( T mm )
s( T max )
s( T ampl )
s{ Precip )
s( Wind )
s( Rel Hum )
s( Rel Sunsh )
N days >lmm
T ' ' '1
—n
-
i
.1
i
i
j I
0 10 20 10 40 50 60 70 0 10 20 30 40 50 60 70
SLP SLP & near surface temperature
Fig 2 8 Comparison of skills between the procedure proposed by VON STORCH et al
(1993) (shaded bars) and the improved procedure (solid bars) Shown arc mean percentages
ol explained variances (=100 r2) in the verification period 1941 80 from all five case study
locations and all 84 selected CCA models luted per season and location
were Davos (52%) in winter and Bern (53%) and Lugano (47%) in summer (Fig 2 7b)
Skills for the numbers of days with precipitation above 1 mm ranged from 29% (Davos)
to 56% (Bern) in winter, and were 4% (Bever), 18% (Davos), 30% (Saentis), 36%
(Lugano), and 42% (Bern) in summer Seasonal mean wind speeds were not well pre¬
dicted, the highest skills were attained in winter at Davos (18%) and in summer at
Saentis (15%) Mean relative humidities were only well reproduced at Saentis in winter
(43%), whereas in summer at no location more than 15% were reached Within-season
standard deviations of daily mean, minimum and maximum temperatures were predicted
in winter with skills between 20% and 45%, an exception was Lugano, where skills for
all three parameters were below 10%, as was the case for all locations in summer For
the within-season standard deviations of daily precipitation totals, skills above 20% were
reached at Bever (24%), Bern (29%), and Lugano (39%) in winter, but, again, at no
location in summer From all remaining within-season standard deviations, a skill above
ca 20% was reached only for relative sunshine durations at Bern (21%) and Davos
(20%) in winter, and at Bever (19%) and Lugano (52%) in summer
26 Chapter 2
In Fig. 2.9, 5 yr running-means of observed (thick lines) and reconstructed (grey areas)
time series of weather statistics are compared, showing a generally coherent behavior
between the pairs of curves. The better reproduction of temperatures (Figs. 2.9a,c) in
contrast to precipitation related variables (Figs. 2.9b,d) also becomes visible in the
different widths of the grey areas, i.e. the uncertainties resulting from different choices in
the numbers of EOFs used to fit the statistical models.
2T BernAVinter
1900 1910 1920 1930 1940 1960 1970 19X0
(C)
(d)
im j Lugano/Summer
A^W^*^E 34
1900 1910 1920 1930 1940 960 1970
Fig 2.9: Comparison of statistically reconstructed time series of local weather statistics
with observations (5 yr running means) Solid lines, observations 190 J -1980 Grey areas
contain 90% of the reconstructed values from 84 selected CCA models fitted separately for
each season and location for the years 1901-1940 using a combined SLP/near-surface
temperature predictor data set (weights 1.1). Explained variances by the average time series
(not shown) of the 84 CCA models in the verification period 1941-1980 were (a) 64%, (b)
55%, (c) 65%, (d) 36%
The Spatial Aspect Downscaling 27
One main reason for the differences in the predictability of the local vanables between the
different locations, seasons and variables certainly is that our approach describes only the
proportion of the interannual variability of a weather statistic which is controlled by
fluctuations of large-scale SLP and near-surface temperature Differences may thus result
from seasonally (winter vs summer) or regionally (northern vs southern Alps) varying
intensities of smaller-scale convective activity, as well as from different combinations of
local factors such as orography, vegetation, and soil characteristics In particular, these
factors may regionally influence local wind systems, snow cover, or energy exchange
with the atmosphere, thus uncoupling the variability of individual weather elements from
the large-scale circulation For example, seasonal mean minimum temperatures, which
depend on smaller-scale thermal inversions and cold air drainage, were generally less
well reconstructed than the seasonal mean and mean maximum temperatures As could
be expected, this was found to be particularly the case for the four valley locations, and
for the winter season
It should be noted that the inclusion of 17 not equally well cross-correlated variables
within the same CCA model tended to worsen the prediction of individual variables to the
benefit of the ensemble This is because - despite the standardisation of all predictands to
the unit standard deviation (Fig 2 3) - the relative importance of individual variables
was implicitly influenced by typical correlation structures found in the predictand data
sets, e g a clustering of temperature- or precipitation-related variables Possibly, due to
an alternative choice or weighting of the predictands, the modest performance obtained
for e g mean relative humidities and wind speeds could be improved
B) LONG-TERM LINEAR TRENDS
The observed and by means of CCA reconstructed 1901-1980 linear trends of six main
weather statistics are summarised in Table 2 2
Seasonal daily mean, mean minimum and mean maximum temperatures displayed an
upward trend at all five stations in winter The CCA models uniformly failed to repro¬
duce any signs of these trends and instead indicated a decrease of temperature since the
beginning of this century With respect to the interannual variations however, the simi¬
larity between reconstructed time series and the m-situ observations was much better
(Fig 2 9a) A similar finding was reported by WERNER & VON STORCH (1993)
CCA indicated a net cooling because of a large-scale trend in SLP (Fig 2 10a), whereas
the spatially less homogeneous trend in the VINNIKOV et al (1987) temperature data
(Fig 2 10b) did not affect the reconstruction of the temperature trends much (see also
Eq 2 1') The reality of the long-term SLP trend was documented by VON STORCH et
28 Chapter 2
Seasonal Statistics Bever Bern Davos Lugano Saentis
obs tec obs rec obs rec. obs rec. obs. rec
Winter
Tmean (°C) 0 78 -0 51 1 14* -0 49 1 75" -0 51 i 17" -0 47 0 61 -0 99
Tmm (°C) 1 16 -014 2 16" -0 41 2 64" -0 44 2* -0 46 0 52 -101
Tmax (°C) 0 60 -0 56 0 16 -0 76 0 77 -0 59 0 44 -0 77 0 82 -0 98
Rel Sunsh (%) -116* -2 1 -11 -14 -6 2 -4 2 -6 0 -4 5 -0 9 -2 2
Precip (%otLTM) -20 9 7 4 16 1 66 122 05 28 4 18 1 -47 4* -2 7
NdaysS1mm(% ot LTM) -9 1 8 0 2 6 16 5 2 62 27 8 20 9 -9 4 3 4
Summer
Tmean (°C) -0 01 0 25 0 50 0 02 0 05 0 04 0 22 -0 09 0 69 0 16
Tmm (°C) 0 27 0 21 1 24" 0 18 1 22" 0 11 2 13* 0 01 0 61 0 19
Tmax CQ -0 58 0 60 -0 25 -0 12 -0 21 -0 06 -2 21* -0 21 0 80 0 17
Rel Sunsh (%) -5 1* -1 6 -6 6- -0 6 -7 1" -12 -101* -07 6 0" 0 2
Precip (% of LTM) 7 1 5 8 11 -0 8 -17 -12 -5 4 4 7 -16 2 -4 9
NdaysS1mm(% of LTM) 09 05 -11 -0 6 -0 1 -1 2 9 7 14 0 1 -11
Table 2.2: Mean linear trends of weather statistics in the period 1901-1980, both observed (obs.) and
reconstructed (rec.) from 84 selected CCA models. The CCA-models were fitted for the years 1901-1940
to a combined SLP/near-surface temperature predictor data set (weights 11). Quantities are given in
respective units per 80 yr; asterisks denote significance of trends on the 95% level (given for observations
only). LTM = long-term (1901-1980) mean
al. (1993), who found a consistent signal in both ship-of-opportunity observations of
SLP and in situ precipitation records on the Iberian Peninsula. When only SLP was used
for CCA, the discrepancy between the CCA models and the local observations was
increased (mean + standard deviation of trend from 84 selected CCA models for winter
mean temperature in Bern: -1.21±0.28 °C per 80 yr). The differences between the 80-
year trends also remained when the 1941-1980 interval was used to fit the CCA models.
There are several candidates that may account for part or all of the discrepancy:
First, the effect of the systematically changing pressure field on the Swiss temperature
must have been counteracted by another large-scale feature, presumably representing the
thermal structure of the troposphere. One possibility is that, e.g. due to the use of chang¬
ing reference intervals, a real trend in near-surface temperature is not correctly represent¬
ed in the VINNIKOV data set. Preliminary calculations with the JoNES-data set (JONES &
BRIFFA, 1992; BRIFFA & JONES, 1993) actually yielded less negative 80 yr trends for
winter mean temperatures, however by only 0.1 -0.3 °C. An alternative explanation could
thus be a trend in a large-scale parameter not included in our analysis, e.g. 500 mb
geopotential height. Due to the lack of data in the first half of this century, however, this
hypothesis could not be tested.
The Spatial Aspect Downscaling 29
(a) (b)
Fig 2 10 1901-1980 linear trends in the dala sets used for (a) winter mean SLP (contour
interval I mb) and (b) winter mean near-surface temperature (contour interval 0.5 "C).
A second reason could be that the statistical link between the local temperatures and the
large-scale circulation has changed with time. For example, systematic changes in the
frequencies of short-lived weather systems, which have a strong effect on the local
temperatures may not be adequately represented in the seasonal mean fields. Possibly,
the linear CCA models may also have failed to capture any non-linear effects.
Third, it should be noted that the positive trends are significant only for the seasonal
mean and mean minimum temperatures at the three urban stations, whereas the two rural
stations Bever and Saentis show much smaller trends (Table 2.2). From Fig. 2.9a it be¬
comes obvious that the CCA models failed to reproduce the trend not only in the last 40
years of the verification interval but also in the first 40 years of the analysis interval. We
speculate that the positive local trends at the urban stations are overestimated due to the
urban heat island effect and increasing air pollution.
In summer, the discrepancy between the indirectly derived and the in situ trends of the
three temperature parameters was considerably smaller than in winter. Only the trends of
the mean minimum temperatures at the three urban stations were not reproduced. We
suggest that these in situ records are contaminated by a significant urbanisation effect.
Another major difference was found at Lugano, where the local observations of seasonal
mean daily temperature maxima indicate a net cooling of -2.2 °C, whereas the large-scalefields specified a small cooling of only -0.2 °C.
For relative sunshine in winter the statistical models were more successful in reproducing
the sign of the trends, even if the absolute values deviated somewhat. The significantreduction in relative sunshine obtained for Bever, as well as its poor reproduction by the
CCA models (Fig. 2.7b), probably occurs because measurements for this location were
taken from the climate station of St. Moritz 6.9 km away, which furthermore has often
been moved.
30 Chapter 2
Trends in summer mean relative sunshine durations were generally underestimated by
CCA, but the signs were correct in all five cases The largest deviations were found for
Lugano, where relative sunshine decreased during 1901-1980 by more than 10%,
whereas the CCA models estimated a reduction by only 0 7% Again, this discrepancy
could be due to the climate station's being moved (H BANTLE, SMA, personal commu¬
nication) or to increased air pollution (see also GENSLER, 1978, p 9)
Trends of daily mean precipitation and of numbers of days with precipitation exceeding 1
mm were mostly reproduced with an error remaining within 10% of the respective 1901-
1980 long-term means At Bever, observed (-21%) and reconstructed (+7 4%) winter
precipitation trends were not consistent, but both numbers are small compared to the
1901-1980 standard deviation which amounts to ca 49% of the long-term mean The
significant trend found for mean winter precipitation in Saentis, which was also not well
reconstructed by means of CCA, is doubtful, since precipitation may not be reliablymeasured at this peak location
2.3.3 Downscaling of GCM-Simulated Climatic Changes
GCM-simulated predictors for the climatic change experiments performed with the
ECHAM1/LSG-GCM were derived as deviations from the mean SLP and near-surface
temperature fields simulated in the first 40 yr of the "control" run The "control" run
shows a drift in the globally and annually averaged near-surface temperature of less than
-0 4 °C in 100 yr, which may be attributed to low-frequency variability of the model In
the DC02 experiment, after one century temperature has increased by 1 7 °C In the
SCNA experiment, global mean temperature rises during the first 40 years by moderate
0 5 °C, which is likely an underestimate due to the "Cold Start"-problem (see HASSEL-
MANN et al, 1992) The rate of temperature growth increases then to 0 35 °C per
decade, such that by the year 2085, and under greenhouse gas concentrations of ca 1150
ppm equivalent CO2, global warming reaches 2 6 °C This result lies between the lowest
and best estimates of approx 2 °C and 3 °C, respectively, given for the same future time-
point by the IPCC (1990)
Fig 2 11 shows time-dependent changes in the temperature and precipitation statistics of
Bern, downscaled from the SLP and near-surface temperature anomaly fields of the 100-
yr SCNA-expenment The range of values predicted by the different CCA models using
different numbers of EOFs typically increases with time, 1 e with the magnitude of large-scale climatic change This is because differences in the regression coefficients of
Eq 2 3 become increasingly effective, the more the predictors deviate from the mean
state used for CCA model estimation
The Spatial Aspect Downscaling 31
(a)
1985 1995 2005 2015 2025 2035 2045 2055 2065 2075 2085
Fig 2 11 Climatic change scenario lor Bern (5 yr running means) Changes in the local
variables were statistically downscaled Irom large-scale SLP and near surlace temperature
anomalies simulated in the IPCC Scenario A experiment perionned wiih the ECHAM 1/
LSG GCM Dashed lines mean responses ol 84 selected CCA models using dilferent
numbers ol EOFs Grey areas contain 90% ol the CCA model estimates
In Figs 2 12 and 2 13 projected changes in seasonal temperature and precipitation
parameters are compared between locations The regional differentiation obtained is in
strong contrast to the homogeneous changes specified at single GCM-gridpoints on a
spatial scale of several I02 km During the last decade of the SCNA experiment, the
projected temperature deviations differ between locations by as much as 1 8 °C in winter
and 2 0 °C in summer Similarly pronounced differences occurred for most other vari¬
ables as well (e g , Fig 2 13)
Blvlt Bern Davos Lugano Saentis
2 x C02
Bevel Bern Davos Lugano Saentis
Fig 2 12 Statistically downscaled changes in seasonal mean temperatures lor the last 20
yr ol the 'double CO2 experiment and lor the last 10 years of the'
IPCC Scenario A'
experiment Shown are the mean changes obtained Irom 84 selected CCA models dttcd in
the period 1901 1940 separately lor each season and location Changes are given relative
loihe 1901 1940 means
32 Chapter 2
Winter Summer
lU(a) Bever Bern Davos Lugano Saentis
40
30
E
E
20
10
At 0
•s 10
720
< 30
40
(b)
1 U MBever Bern Divos Lugano Sacnlis
1901 40 «• 1941 80
Fig 2 13 Statistically downscaled changes in (a) winter precipitation totals and (b)
summer probabilities of days with precipitation > I mm for the last 10 years ol the
IPCC Scenario A experiment The CCA models were fitted in the periods 1901 1940
(solid points) and 1941 1980 (open points) respectively Data points indicate mean
changes obtained from a set of 84 selected CCA models fined separately for each periodseason and location error bars denote the empirical intervals containing 90% ol the
respective CCA model responses Changes are given relative to the 1901 1980 means
As can be seen from Fig 2 12, the DC02 and the SCNA experiments yielded similar
patterns of change between the different locations Since this was not only the case for
temperature (see also Table 2 3), we will focus below on the last decade of the SCNA
experiment, where the most pronounced changes occurred Further, since for some
variables the downscaled changes strongly depended on the data used for CCA model
estimation (Fig 2 13), the mean responses of CCA models fitted in the period 1901-
1940, as well as in the complementary period 1941-1980 will be considered
The results for six main variables are summarised in Table 2 3 For seasonal mean tem¬
peratures, the best estimates - i e the mean changes from all used CCA models per
season and location - amounted to 2 0 °C in winter and 2 4 °C in summer, averaged over
the three northern locations Bern, Davos and Saentis, and to 1 2 °C in winter and 2 7 °C
in summer, averaged over the two southern locations Bever and Lugano Winter mean
daily temperature extremes showed similar spatial patterns of change to the winter tem¬
perature means winter mean daily minimum temperatures were found to rise at the
northern (southern) slope of the Alps by on average 2 2 °C (1 5 °C) whereas for the
maxima moderate changes by 1 8 °C (1 0 °C) occurred Warming was very differentlydistributed in summer, where a substantial increase by 3 1 °C (3 9 °C) was obtained for
the temperature maxima, as opposed to only 2 6 °C (1 7 °C) for the temperature minima
Projected changes in relative sunshine durations were generally small for winter, whereas
in summer changes of +36% were specified at the northern, and +17% at the southern
slope of the Alps Winter precipitation totals changed by only ca +6% at the northern,
The Spatial Aspect Downscaling 33
Winter Summer
Seasonal
Statistic
Loca¬
tion
1901-80
u o
DC02
low b.e. high
SCNA
low b.e. high
1901-80
u o
DC02
low b.e. high
SCNA
low b.e. high
Tmean
Bever 8 26 2 OS 0 15 0.62 1 11 0 46 1.29 2 14 1068 1 19 141 2.00 2 71 1 98 2.80 1 75
Bern 021 221 0 72 1.25 2 10 1 64 2 31 151 16 82 1 15 1 12 1.72 2 70 1 91 2.52 3 71
Davos 5 58 2 16 0 48 0.93 1 18 1 04 1.82 2 51 11 17 1 25 0 82 1 74 2 19 1 19 2.53 3 48
Lugano 2 92 111 0 21 0.45 0 95 0 15 1.03 ] 64 20 10 111 0 47 1.24 2 11 1 01 1.95 3 01
Saentis -8 17 1 9S 0 16 0.88 1 46 1 24 1.86 2 72 4 20 1 14 1 71 2.21 2 84 2 15 3.15 4 1.1
Tmin
Bever -14 56 2 40 0 09 1.05 2 27 0 29 1.83 162 4 07 0 92 III 1 64 2 60 1 48 2.24 1 11
Bern -101 2 15 0 61 1.40 2 06 1 57 2.51 1 46 1 1 99 1 06 1 12 1.85 2 57 1 47 2.53 3 27
Davos -9 96 2 14 0 64 1.22 2 01 118 2 14 140 5 91 1 07 0 18 1.74 2 85 0 58 2.47 4 15
Lugano 0 60 1 64 0 21 0 45 121 0 15 1 12 1 96 14 85 1 00 -0 11 0.77 1 62 0 10 1.22 2 04
Saentis -10 64 1 97 0 45 0.93 1 41 1 11 1.91 2 60 1 76 1 21 1 48 1.93 2 46 1 99 2.70 1 51
Tmax
Bever -1 87 1 84 0 19 0.54 0 91 0 64 1.17 1 75 1761 201 1 90 3.28 4 62 2 81 4.59 6 29
Bern 2 95 2 28 0 46 0.93 1 61 1 17 1.93 2 94 22 10 1 79 0 94 1.82 1 18 1 79 2.75 4 51
Davos 061 191 0 15 0.77 1 10 0 94 1.61 2 01 16 56 1 52 1 15 1.78 2 19 1 87 2.63 3 50
Lugano 7 29 1 58 0 59 0.36 1 47 -0 11 0.92 2 56 26 76 1 82 0 48 2.19 4 19 0 90 3.29 6 30
Saentis 5 16 1 97 0 26 0.80 1 44 1 08 1.73 2 68 7 11 1 58 2 12 2.67 1 18 2 88 3.82 4 94
Rel Sunsh
Bever 42 1 12 7 15 2 2.5 12 5 -19 6 4.7 27 5 50 8 8 0 -9 1 10.6 41 4 -10 7 22.0 67 4
Bern 24 1 7 9 18 0 -17 8 9 -19 2 1.8 14 1 520 9 5 0 8 9.6 19 6 5 6 16.4 10 0
Davos 49 9 118 -10 2 -3.6 2 0 12 0 -3.1 6 2 49 4 8 1 7 4 3.6 14 1 -4 1 8.5 20 4
Lugano 52 6 119 -II 7 -4.1 0 8 16 1 -4.5 4 4 61 9 7 5 10 6.4 114 2 9 11,4 21 9
Saentis 18 2 12 6 -28 2 -4.3 24 2 .20 4 8.1 66 0 15 0 7 0 -0 2 S4.2 162 1 4 8 82.5 221 9
Precip
Bever 4 5 11 -5 1 17.5 15 4 4 4 34.0 56 9 9 9 14 -2 6 25.7 55 9 -6 9 30.2 71 3
Bern 60 17 21 5 -1.3 14 1 -18 1 -1.0 27 0 115 4 1 25 4 6.9 19 4 -16 5 4.1 44 8
Davos 6 8 40 -162 5.0 120 -21 2 9.7 52 7 12 9 14 -5 1 8.9 18 6 -12 0 7.3 20 9
Lugano 7 4 5 7 2 7 28.9 57 1 17 5 53.0 88 5 18 1 7 8 1 1 22.9 52 4 -16 22.9 59 2
Saentis 20 7 114 -18 7 6.3 42 1 -15 9 9.1 419 27 9 9 7 -60 7 -27.5 1 6 -89 5 -42.5 -0 5
Ndays>lmm
Bever 216 9 4 -4 1 7.2 18 7 0 8 14.7 11 0 11 5 7 2 -119 12.2 12 5 -20 2 13.5 40 5
Bern 28 5 12 1 -22 2 -2.3 14 8 -14 8 -3.0 24 1 15 1 10 4 -21 1 -6.0 11 8 118 -13.1 8 5
Davos 26 1 9 4 -114 0.4 12 9 -18 1 12 207 42 1 7 6 8 2 .0.2 8 0 -114 -3.5 7 0
Lugano 18 7 10 2 0 2 16.7 11 2 9 2 27.5 49 1 111 8 2 8 5 11.8 11 0 17 0 10.8 .14 6
Saentis 418 116 10 5 0.7 118 -124 1.8 186 49 8 9 8 -20 6 -4.9 7 1 28 9 .9.6 5 1
Table 2.3 Projected deviations of weather statistics from their respective 1901-1980 long-term means
(u) for the last 20 yr of the "double CO2" experiment (DC02) and for the last 10 yr of the "IPCC Sce¬
nario A" experiment (SCNA) o = 1901-1980 observed standard deviations of the weather statistics;
b e = "best estimate" changes, i.e. mean deviations obtained from 168 selected CCA models (84 fitted
for the years 1901-1940, and 84 for the years 1941-1980) using a combined SLP/near-surface temperature
predictor data set (weights I I), low/high = lower and upper limits of empirical intervals containing
the responses of 151 (=90% of 168) CCA models 1901 -1980 means and standard deviations are given in
°C for the temperature parameters, in % for relative sunshine duration, and in cm mo-1 for precipitation .
Deviations are given in °C for temperatures and in % of the respective 1901-1980 means for all other
variables
but increased dramatically by 34% (Bever) and 53% (Lugano) at the more southern ones
(see also Fig. 2.13a). Summer precipitation was found to increase at all locations by 4%
to 30%, except for Saentis, where a strong decrease by 42% was specified; the strongest
34 Chapter 2
increase occurred again at the southern slope of the Alps (average 26%) Winter
probabilities of wet days did not change much at the three northern locations, whereas at
Bever it increased by 15% and at Lugano by 28%, summer probabilities of wet days
were found to change at the three northern locations by 3% to -13%, and, again, to
increase at the two southern locations by an average of ca 12% (see also Fig 2 13b)
The following "best estimates" were obtained for the weather statistics not considered in
Table 2 3 Seasonal mean daily temperature amplitudes decreased in winter quite uni¬
formly by on average -0 4 °C, in summer they increased at the three northern locations
by 0 5 °C on average, and at the two southern locations by 2 2 °C Winter mean daily
wind speeds were found to decrease at Bever by 48% and at Davos by 40%, and to
increase at Lugano by 9%, of the respective long-term mean, in summer, mean wind
speeds decreased again at Bever (-75%) and Davos (-64%), but increased by 35% at
Bern, and 61% at Lugano Seasonal mean relative humidities did not change much in
both seasons at all locations, except for Saentis in summer (-15% ) Within-season
standard deviations of daily mean, minimum and maximum temperatures decreased
systematically at all locations, on average by 21%, 21% and 18% in winter and by 11%,
10% and 7% in summer, respectively Within-season standard deviations of daily tem¬
perature amplitudes were generally reduced in winter, by ca 10%, whereas in summer
less coherent changes among the locations, in the order of ±10%, were obtained
Within-season standard deviations of daily precipitation totals increased in winter at
Bever and Lugano by 40% and 46%, respectively, and in summer by 5%-24% at all
locations, except for Saentis, where a reduction by 41% occurred Within-season
standard deviations of daily relative sunshine durations changed in winter only at Saentis
(+10%), whereas in summer they increased at Saentis (+74%) as well as at Bever
(+10%), and decreased by 5% to 10% at the remaining locations
The procedure yielded not only a regionally differentiated but also an internally consistent
picture of possible climatic change Note that plausible changes were even obtained for
several variables which were not well predicted in the model verification phase In some
cases however, e g for summer precipitation and mean relative sunshine duration at
Saentis, the CCA models did not yield sensible results Possible reasons could be that the
linearity assumption (Eq 2 3) was violated, or that inconsistencies in the local data sets
were amplified under the given large-scale climatic change
All in all, winter climate was specified to be milder and wetter than under present condi¬
tions This suggests an increased probability for nighttime cloud cover, which is consis¬
tently mirrored in a stronger rise of the minimum as compared to the maximum tempera¬
tures, a general decrease of mean daily temperature amplitudes, and a smaller within-win-
ter temperature variability The summer was specified to become generally hotter and
wetter, in this connection, strong increases in relative sunshine durations, mean daily
The Spatial Aspect Downscaling 35
maximum temperatures and temperature amplitudes suggested increased irradiation and
surface heating The increases m summer precipitation could correspond to stronger con¬
vective activity, or to more intense intrusions of maritime air, which would compensate
for the more pronounced ocean-continent temperature contrast as simulated by the GCM
In both seasons, the precipitation increases consistently tended to occur with increased
within-season variabilities of daily precipitation totals and increased probabilities of wet
days
It should be noted that the large-scale signals underlying the responses of all weather
statistics (Eq 2 1') were found, in both seasons and for all locations, to be mainly deter¬
mined by the rising near-surface temperature In particular, the projected changes depend¬
ed strongly on the inclusion of near-surface temperature as a large-scale predictor For
example, when SLP was used as the only predictor in winter, anomalously high SLP be¬
tween 35° N and 50° N in the final decade of the SCNA experiment implied precipitation
decreases by 4% to 12% in Bever, Bern and Lugano, and positive changes of a few per¬
cent in Saentis and Davos Yet in Fig 2 12a precipitation increases are shown which,
since they include the information of the temperature field, can be interpreted to be a con¬
sequence of the higher moisture content of the air advected towards the Alps, counterbal¬
ancing the effect of slightly weakened westerlies It is possible, that inclusion of addi¬
tional predictors which show similarly drastic changes to those of near-surface tempera¬
ture in the GCM, e g mid-troposphenc pressure fields, could further modify our results
The downscaled climatic changes strongly reflected the characteristics of the underlyingGCM-simulation The DC02- and SCNA experiments yielded qualitatively similar
changes in local climates (Fig 2 12, Table 2 3) because the GCMs large-scale respons¬
es for SLP (not shown) as well as for near-surface temperature (CUBASCH et al, 1992)
to the respective greenhouse gas forcings show quite similar large-scale patterns in both
experiments In the case of the SCNA experiment, the small changes projected for the
first 50 years of the scenario period (Fig 2 11) reflect the delayed global mean tempe¬
rature response of the ECHAM 1/LSG GCM Further, the GCM-downscaled year-to-
year fluctuations of the weather statistics were generally smaller than under present
climate (Fig 2 9 vs Fig 2 11) This is due to a systematic underestimation of the
interannual variability of seasonal mean SLP by the GCM which occurs in the "control"
run, as well as in the first decades of the SCNA-expenment for both seasons considered
(not shown) A possible reason for this could be the underestimation of mid-latitude
cyclonic activity resulting from the model's low spatial resolution (cf CUBASCH etal,
1992) Finally, it should be noted that the smallest warming occurred for both seasons at
the southern slope of the Alps (Lugano) This is possibly because temperatures at the
north alpine locations (Bern, Davos and Saentis) and at the more eastwardly located
Bever are more strongly determined by advection from higher latitudes and northeastern
Europe, which are the regions where the largest warming takes place in the GCM
36 Chapter 2
2.4 Conclusions
From our case studies at five representative locations in the Alps we conclude:
(1) It is possible to establish plausible statistical relationships which allow the large-scale
climatic changes produced by GCMs to be linked to local ecosystem model inputs, even
in a complex orography like the Alps. The results obtained were meaningful for winter
(Dec, Jan., Feb.) as well as for summer (June, July, Aug.). For each location, the
statistical models quantified the effects resulting from changes in major circulation
patterns, e.g. from changes in the strengths of large-scale zonal or meridional flow or of
subsidence, and from fluctuations in the associated large-scale near-surface temperature
distributions.
(2) The strength of the statistical link to the large-scale circulation depends on the season,
location, and variable considered. The procedure allows several important ecosystem
inputs to be reconstructed on a yearly to decadal timescale reasonably well.
The most reliably reconstructed variables were the seasonal means of daily mean,
minimum and maximum temperatures, whereas the poorest results were obtained for
seasonal mean temperature amplitudes, relative humidities, wind speeds, and, for
summer only, daily precipitation totals, numbers of days with precipitation above 1 mm,
plus most within-season standard deviations of daily variables. Some observed 80 yr
trends, especially for mean daily temperatures and temperature extrema in winter, and for
mean minimum temperatures in summer, could not be correctly reproduced. On the
whole, most variables were better reconstructed in winter than in summer.
On average over all variables, the three north alpine locations (Bern, Davos and Saentis)
could be better linked to large-scale climate than the two more southern locations (Bever,
Lugano). This difference was more pronounced in winter than in summer.
(3) The downscaling procedure originally proposed by von Storch et al. (1993) could be
improved by including large-scale near-surface temperature anomalies as an additional
large-scale predictor. Improvement was possible for almost all variables and generally
for all locations considered in this study. The largest improvements were found for
seasonal mean temperatures and mean daily temperature extremes, and for the summer
season.
(4) The method presented here allows to flexibly regionally differentiated, time-depen¬
dent estimates of climatic change to be flexibly for input variables required by ecosystem
models at a much smaller spatial scale than given by the resolution of present GCMs.
The Spatial Aspect Downscaling 37
However, the following restrictions of this approach must not be overlooked
- Considerable amounts of data must be available and have to be managed
Required are sufficiently long records of the local variables close enough to the
location of interest, simultaneous observations of the large-scale state of the
atmosphere in the sector containing this location, as well as corresponding data
sets from experiments with GCMs
- The method assumes that the established statistical relationships remain valid also
under future climatic conditions For example, towards the end of the "IPCC
Scenario A" experiment the GCM-simulated fields undergo changes which
exceed the range of the observations used to fit the statistical models, thus, to¬
wards the end of the next century the projected changes in the local variables
represent (linear) extrapolations which might not necessarily hold in the real
world
- The downscaled local climatic changes will not be more reliable than the un¬
derlying GCM-expenments For example, the relatively modest temperature
increases obtained in our study at all locations are a consequence of the relatively
small sensitivity of the ECHAM 1/LSG-GCM to increased greenhouse-gas con¬
centrations Local climatic changes should therefore be explored for experiments
with other GCMs, a procedure which is easy to accomplish with our method
In spite of these difficulties, the proposed method represents to our knowledge the best
approach currently available to consistently link GCM-simulated climatic changes to
ecosystem models In combination with stochastic weather generators it can be used to
derive multivariate scenarios of climatic change on a wide range of temporal scales, and
yet it allows to attain the high spatial detail particularly required within a mountainous
region to be attained Moreover, our approach can be easily adapted to the specific needs
emerging from ongoing ecosystem research Yet, further improvements can be envis¬
aged, e g by downscaling only a reduced set of inputs, by using additional large-scale
predictors, and, as recent work showed (GYALISTRAS & FISCHLIN, 1996), by spatial
inter- and extrapolation to points with scanty or even no measurements
38 Chapter 3
3. The Temporal Aspect: Stochastic Simulation of
Monthly Weather
3.1 Introduction
Dynamic simulation models present important tools to study possible impacts of future
climatic change on ecosystems Several of these models require site-specific, monthly
weather data as an input For example, biogeochemical models focusing on C/N dynam¬
ics, such as MBL-GEM (RASTETTER et al, 1991) or CENTURY Forest (METHERELL et
al, 1994), or forest succession models such as FORSKA (PRENTICE et al, 1993) and
FORCLLM (BUGMANN, 1994,1996, FISCHLIN et al, 1995) are dnven by different com¬
binations of monthly weather vanables related to temperature, precipitation and relative
sunshine duration Simulation studies with such models typically require many, e g up
to 200 (BUGMANN et al, 1996), realizations of the weather process, and for all months
of the year Furthermore, the inputs must often be provided for time spans covenng at
least a few decades to millenia
A straight-forward approach to provide such inputs is to sample weather sequences trom
the local weather record (e g ,DAVIS & BOTKIN, 1985, BOTKLN & NlSBET, 1992) To
simulate scenanos of possible future weather the record can be adjusted according to the
assumed climatic scenano, for instance by adding a certain fixed amount (e g ,±2 °C for
temperature and ±20% for precipitation) to all elements of the measured time senes (e g ,
ROBOCK et al, 1993) This approach is computationally efficient and relatively easy to
implement However, it has two major drawbacks Firstly, if not a very long record is
available, the measurements may not cover all possible combinations of regional weather
states relevant for the system under investigation In particular, due to sampling variabili¬
ty the frequencies of events exceeding specific thresholds - such as temperature thresh¬
olds that control the survival and growth of organisms - may not be well represented,
and this may decrease the robustness of the results obtained with an impact model
Secondly, future climatic change might include changes in global climatic vanabihty
(e g ,MEEHL et al, 1994, LUPO et al, 1997), which could strongly affect local precipi¬
tation patterns and/or auto- and cross-correlations of regional weather vanables (e g ,
GREGORY & MITCHELL, 1995, MEARNS et al, 1995a, 1995b, TSONIS, 1996) The
direct manipulation of the local weather record allows however mainly for changes in
means, and therefore becomes impractical when one wishes to investigate more complex
patterns of climatic change
The Temporal Aspect. Stochastic Simulation 39
A second possibility is to use statistical or stochastic models which simulate local weather
conditional on the large-scale state of the atmosphere. Several such "downscaling" pro¬
cedures have been proposed until now, which operate at a seasonal (e.g., VON STORCH
et al., 1993; GYALISTRAS et al., 1994) to daily (e.g., BARDOSSY & PLATE, 1992;
ZORITA etal, 1995; BURGER, 1996) temporal resolution. To simulate scenarios of
future weather these procedures can be applied to the weather patterns produced in simu¬
lation experiments with numerical climate models (e.g., GYALISTRAS et al., 1994;
ZORITAefa/., 1995; BURGER, 1996).
This approach has the advantages that it uses the physically consistent output of the cli¬
mate models, and that the resulting scenarios reflect possible changes also in the higher-order statistics of weather. Moreover, the use of large-scale weather patterns circumvents
the limited regional reliability (GROTCH & MACCRACKEN, 1991; VON STORCH, 1995)
of the climate models. However, some major restrictions occur. Firstly, the daily to in¬
terannual variability of weather is not very reliably simulated by present-day climate mod¬
els (e.g., HULME et al., 1993), and this can lead to large enors in the statistics of the
simulated local weather (BURGER, 1996). Secondly, due to the enormous computing re¬
quirements of climate models normally only one realization of large-scale weather is
available from climate simulations. This prevents extensive sensitivity analyses and ex¬
perimentation with many different scenarios, as this is often required in the context of im¬
pact studies. Furthermore, simulations with climate models typically extend only over a
decade to a century. This restricts the construction of transient scenarios over longer time
periods, which again sharply contrasts with the requirements of many ecosystem studies.
An attractive alternative to both above approaches are so-called "weather generators", i.e.
stochastic models which simulate the local weather based on site-specific climatic parame¬
ters and a pseudo-random number generator. Unlike the direct use of the local weather
record, a weather generator allows to describe local climates based on a comparativelysmall number of well-defined climatic parameters, which may be estimated quite reliably
already from relatively short time series by means of appropriate procedures (e.g.
Richardson, 1981). Also, differently from the downscaling approaches, which
critically depend upon a sufficiently good simulation of the large-scale weather by the
climate models, weather generators can be designed to simulate the local weather with a
high statistical accurracy. The climatic parameters involved can be easily adjusted - either
based on ad-hoc assumptions, or according to estimates of local climatic change as ob¬
tained from a downscaling procedure - to experiment with a wide range of climatic sce¬
narios (see e.g. WiLKS, 1992). Moreover, the relatively modest computing requirementsof this approach make it possible to generate a large number of weather sequences, and
under any time-dependent scenarios of climatic change that may extend far into the future.
40 Chapter 3
Though a large body of literature exists on the stochastic simulation of daily (e.g.,
RICHARDSON, 1981; WILKS, 1992; GREGORY et al, 1993; KATZ, 1996) or hourly
(e.g., COWPERTWAIT, 1991; AGUIAR & COLLARES-PEREIRA, 1992; BO et al. 1994;
GYALISTRAS et al, 1997) weather variables by means of local weather generators, we
are aware only of little work that focuses on the simulation of monthly weather.
BOTKIN & NlSBET (1992) investigated the effect of sampling variability in the definition
of the baseline climate on forest compositions simulated by means of the JABOWA forest
succession model. They resampled monthly weather data from different subperiods of
the 1901-1980 record and found generally only small effects on simulated forest compo¬
sitions. However, they considered but one location, at the transition between boreal and
northern hardwoods forests in North America, and did not examine the stochastic simula¬
tion of monthly weather variables.
BUGMANN (1994, 1996) proposed in collaboration with the authors of the present paper
an improvement of the weather generator commonly used until then to drive forest patch
models. In the traditional approach, uncorrelated variates for the monthly weather vari¬
ables (typically monthly mean temperature and monthly precipitation totals) are drawn
from a seasonally varying, bivariate normal distribution (see e.g. FISCHLIN et al, 1995).
BUGMANN (1994, 1996) introduced seasonally varying cross-correlation between the
monthly temperature and precipitation variables and reported a significant effect on the
distribution of a drought stress index (defined as one minus the ratio of actual to potential
evapotranspiration) used in many forest simulators. However, according to our knowl¬
edge, no systematical analysis of the accuracy and suitability of different stochastic
weather generators to simulate monthly weather variables has been carried out until now.
In the present work we present such an analysis using data from 89 European sites re¬
flecting a wide range of climatic conditions. The focus is on the effects of the simulated
monthly weather variables on the distributions of three derived bioclimatic variables typi¬
cally used to drive forest succession models.
Our analysis reveals some shortcomings of the monthly weather generators commonly
used to study impacts of climatic change on forests and other systems. We propose an
improved stochastic weather generator which requires a comparatively small number of
climatic parameters, relies upon more robust methods to estimate these parameters, and
produces more realistic distributions of bioclimatic variables.
The Temporal Aspect Stochastic Simulation 41
3.2 Material and Methods
3.2.1 Data
We used monthly mean temperatures and monthly precipitation totals from 89 Europeanstations listed in Table 3 1 The data were extracted from the "Global Historical Clima¬
tology Network - GHCN' (VOSE et al, 1992, ElSCHEID et al, 1995) data set, and the
data base of the Swiss Meteorological Institute (BANTLE, 1989) From both data bases
we used only stations within the sector 12° W to 40° E and 42° to 75° N for which at least
20 yrs of simultaneous temperature and precipitation measurements were available in the
30-yr interval 1951-1980, and at least 60 yrs in the 100-yr interval 1894-1993
Data from the period 1951 -1980 were used to define the baseline climates, and to estimate
the parameters of all stochastic models The observed long-term mean annual mean tem¬
peratures, annual precipitation totals, annual growing degree-days totals (threshold 5 5
°C), winter minimum mean temperatures, and annual potential and actual evapotranspi¬
ration at the 89 stations are shown in Fig 3 1 (For the procedures used to compute the
various variables see section "Bioclimatic Variables" below) It can be seen that the sta¬
tions cover a large variety of climates, which range from warm-temperate/lower montane
to boreal/subalpine conditions
42 Chapter 3
Nr ID Name Aim. Lot! Lai Years Nr ID Name AIM. Lon Lai Yean
6
7
8
9
10
109100
4140
600
1940
191700
1127400
111700
5520
9850
115200
ABERDEEN/DYCE
ALTD0RF
AROSA
BASEL B1NNINGEN
BELFAST/A1RP
BEOGRAD
BERGEN/FLORIDA
BERN LIEBEFELD
BEVER
BODOVI
65
451
1847
117
81
112
19
570
1712
NA
22
8 6
9 7
7 6
62
20 5
5 1
7 4
99
144
57 2
46 9
46 8
47 S
54 7
44 8
60 4
46 9
46 6
67 1
1829 1990
1901 1991
1911 1991
1901 1991
1911 1990
1888 1990
1921 1990
1901 1991
1901 1982
1869 1990
46
47
48
49
50
51
52
51
54
55
1608000
2685000
7180
8020
2761200
1086600
722200
6140
1184600
2617900
MILANCVL1NATE
MINSK
MONTANA
MONTREUX CLARENS
MOSKVA
MUENCHEN RIEM
NANTES
NEUCHATEL
NIKOLAEV
NOVGOROD
107
NA
1495
408
NA
529
27
487
NA
NA
9 1
27 5
7 5
6 9
17 6
I 1 7
1 6
7 0
12 0
II 1
45 4
51 9
46 1
46 4
55 8
48 1
47 2
47 0
47 0
58 5
1764 1990
1891 1988
1911 1991
1911 1991
1820 1989
1848 1989
1815 1989
1901 1991
1891 1988
1892 1988
II
12
11
14
15
16
17
18
19
20
1542001
5610
640
460
626000
1800
4410
316200
5740
670000
BUCUREST1/FILARET
CHATEAU DOEX
CHUR
DAVOS
DE BILT
EINSIEDELN
ENGELBERG
ESKDALEMUIR
FR1BOURG
GENEVE-COINTRIN
82
980
586
1590
2
910
1018
242
614
416
26 1
7 1
9 5
9 8
5 2
8 8
8 4
12
7 1
6 1
44 4
46 5
46 9
46 8
52 1
47 1
46 8
55 1
46 8
46 1
1865 1989
1911 1991
1911 1991
1901 1991
1849 1990
1911 1991
1911 1991
1911 1990
1911 1991
1826 1990
56
57
58
59
60
61
62
61
64
65
1181700
148801
165701
715000
182700
917900
1151800
2642200
1471100
9980
ODESSA
OSLO/BLINDERN
OXFORD
PAR1S/LE BOURGET
PLYMOUTH
POTSDAM
PRAHA/RUZYNE
RIGA
ROSTOV NA DONU
S MAR1A/MUESTAIR
NA
96
61
66
27
81
180
NA
NA
1190
10 6
107
1 2
2 5
4 1
11 1
14 1
24 1
19 8
10 4
46 5
59 9
51 7
49 0
50 4
52 4
50 1
57 0
47 1
46 6
1841 1989
1866 1989
1767 1988
1770 1989
1911 1990
1921 1990
1805 1989
1850 1986
1891 1988
1911 1991
21
22
21
24
25
26
27
28
29
30
219600
3410000
297400
1112000
100100
2670200
106500
2662900
2252200
1114500
HAPARANDA
HAR KOV
HELSINKI VANTAA
INNSBRUCK FLUGH
JAN MAYEN
KALININGRAD
KARASIOK
KAUNAS
KEM PORT
KIEV
5
NA
51
581
10
NA
129
NA
NA
NA
24 2
16 1
25 0
II 4
87
20 6
25 5
21 9
14 8
10 5
65 8
49 9
60 1
47 1
70 9
54 7
69 5
54 9
65 0
50 4
1860 1990
1891 1988
1845 1990
1856 1989
1922 1990
1848 1989
1901 1970
1892 1988
1861 1986
1856 1989
66
67
68
69
70
71
72
71
74
75
1100
1526000
7500
1709900
281600
9010
1010
102600
1516000
601100
SCHAFFHAUSEN
SIBIU
SION
SOCI
SODANKYLA
ST GOTTHARD
ST GALLEN
STORNOWAY
SULINA
THORSHAVN
457
441
542
NA
179
2090
664
9
1
44
8 6
24 2
7 4
19 7
26 7
8 6
9 4
61
29 7
68
47 7
45 8
46 2
41 6
67 4
46 6
47 4
58 2
45 2
62 0
1911 1991
1851 1990
1901 1977
1881 1986
1907 1990
1901 1970
1911 1991
1911 1990
1868 1990
1868 1990
11
12
11
14
15
16
17
18
19
40
618001
1492900
1101200
1400900
1119100
8560
6480
2606100
100500
2640600
KOBENHAVN
KRASNODAR
KREMSMUENSTER
KURSK
LVOV
LA CHAUX DE FONDS
LANGNAU1E
LENINGRAD fTOWN)
LERWICK
LIEPAJA
22
NA
182
NA
NA
1060
695
NA
82
NA
12 6
19 2
14 1
16 2
24 0
6 8
7 8
10 1
1 2
21 0
55 7
45 0
48 1
51 7
49 8
47 1
46 9
60 0
60 1
56 6
1821 1989
1895 1988
1820 1990
1891 1988
1876 1990
1901 1981
1911 1991
1740 1989
1911 1990
1881 1986
76
77
78
79
80
81
82
81
84
85
110000
644700
245800
195100
110200
1101800
2671000
2701700
1412200
2281700
TIREE
UCCLE
UPPSALA
VALENTIA OBSERV
VALLEY
VASILEVICl
VILNIUS
VOLOGDA
VORONEZ
VYTEGRA
12
100
21
9
1 1
NA
NA
NA
NA
NA
69
44
17 6
10 1
4 5
29 8
25 1
19 9
19 2
16 5
56 5
50 8
59 9
51 9
51 1
52 1
54 6
59 1
51 7
61 0
1911 1990
1811 1989
1774 1976
1861 1990
1911 1990
1891 1988
1881 1986
1891 1988
1862 1989
1881 1986
41
42
43
44
45
106800
9480
4590
748000
765000
LOSSIEMOUTH
LUGANO
LUZERN
LYON/BRON
MARSEILLE/MARIGN
11
276
456
200
6
1 1
9 0
8 1
5 0
5 2
57 7
46 0
47 0
45 7
41 5
I860 1972
1901 1991
1911 1991
1841 1990
1749 1989
86
87
88
89
1217500
1101500
1242400
1700
WARSZAWA OKECIE
WIEN/HOHE WARTE
WROCLAW II
ZUERICH SMA
106
201
120
556
21 0
16 4
(6 9
86
52 2
48 1
51 1
47 4
1801 1990
I85I I990
I859 (989
I90I 1991
Table 3.1 Chmatological stations used in this study Data for stations with ID < 9999 were extract¬
ed from the data base of the Swiss Meteorological Institute (BANTLE, 1989), all other data from ihe
"Global Historical Climatology Network - GHCN" (VOSE et al, 1992, EISCHElDera/, 1995) data set
Alt = altitude in m above sea level, Lon = longitude (°E), Lat = latitude (°N), NA = not available
The Temporal Aspect Stochastic Simulation 43
2500
on
(mm)2000
<g
pit 1500
o0)
a.1000
eg
o
ten 500
cc
<
0
1 i
a
n
a
DrfflT
3
D
nDrfWSK IT
aD a
-5 0 5 10 15 20
Annual mean Temperature ( C)
10
5
0
-5
-10
-15
-20
P. innii
2n D i i
[ 'ififfl a a
°
fiffODDD
3
a
R*a4
Ul<
800
700
600
500
400
300
]
^iRPD
u
g
DO
nJS3
Db
u
3
0 1000 2000 3000 4000
Annual growing degree-day total (°Cd)
300 400 500 600
PET (mm)
700 800
Fig. 3.1 Observed (bio-)chmatic parameters at the 89 chmatological stations consid¬
ered in this study Shown are long-term means derived from monthly mean temperatures
and monlhly precipitation totals for the years 1951 -1980 Annual growing degree-day totals
refer to the temperature threshold 5 5 °C and were calculated according to the procedure
given in FISCHLIN et al (1995) Monlhly winter minimum temperatures were defined as
Min(TDec, TJan, TFeb), where T denotes a month's mean temperature, PET - annual
potential evapotranspiraiion, calculated from monthly mean temperatures according to
THORNTHWAITE & MATHER (1957), AET - annual actual evapotranspiraiion, computedfrom PET and monthly precipitation totals by means of the soil water balance model byBUGMANN & CRAMER (1997). For details see section "Bioclimatic Variables"
44 Chapter 3
3.2.2 Stochastic Models
All types of stochastic models considered were given by the discrete time system
2£<k) = A(P) 2L(k i)+ B(P) £<k) (3 1)
XoO^CXflO.llrp,) (3 2)
where X is the state vector of dimension N, k > 1 denotes the current time (time step = 1
month), A and B are the NxN system and input matrices, respectively, £ is an input
vector of independent random components from a N-dimensional normal distribution
M0.1). Y is the N-dimensional output vector, f_ is a vector of non-linear output func¬
tions, u denotes the parameters of f_, and p is the phase within the year, i e, p = {(k-1)
MOD 12), where p=0 corresponds to January and p=l 1 to December
Model
Type
Description Characteristics # Parameters
needed in Eq. (3.1)
# Param.
(N=q=2)
I Uncorrelated vanates X, A = 0, B = I I2N 24
II Cross-correlated vanates X, A = 0, B * 0, ftj = 3j, 12N(N+l)/2 36
III AR(1), fixed-phase cyclostationary A * 0, B * 0 12N2+12N(N+l)/2 84
IV AR(1), phase-averaged cyclostationary A * 0, B * 0 (l+2q)N2+12N(N+l)/2 56
Table 3.2 Types of stochastic models considered X, element of state vector X., 0,1 zero and unit
matrix, respectively, Py elements of B, AR(1) lag-1 autoregressive model, N system order, q num¬
ber of Fourier-harmonics used to fit parameters of a Type IV model
In this study were N=2 and X = (X|, X2)' = (T, P)', where T and P denote monthlymean temperatures (T) and monthly precipitation totals (P), transformed according to
Eqs (3 9a) or (3 9b) Four basic types of models which used different matrices A and
B (Table 3 2), and four kinds of output functions / (Table 3 3) were investigated The
procedures used to estimate the two matrices and the vanous parameter vectors u are pre¬
sented under section 3 2 3
The models of Types I and II were included because they are typically used to generate
monthly weather data in the context of climate change studies The models of Type III
were considered to investigate inasmuch the inclusion of temporal autocorrelation (A * 0)
allows to improve the statistical accurracy of the simulated weather sequences This
model requires, however, also more parameters (Table 3 2) Therefore, an additional
type of models, Type IV, was also considered The Type IV models are formally identi¬
cal to the Type III models, but thanks to a more sophisticated parameter estimation pro¬
cedure (see later) they require a considerably smaller number of parameters (Table 3 2)
The Temporal Aspect' Stochastic Simulation 45
OutputFunction
Description Equations
fi Destandardization of state vecior element X, with monthly long-term
means u^.) and variances a,/-) of corresponding output Y,
(3 4a), (3.5a)
flL In addition lo f\ uses a seasonally varying log-normal distribution for
output Y,
(3 4a), (3 5b)
Si Same as /|, but only the first two (temperature), respectively three (pre¬
cipitation) harmonics of the annual cycles of u^-) and a,^ are used
(3 4b), (3 5a)
/2L Same as /|l, but only the first two (temperature), respectively three (pre¬
cipitation) harmonics of the annual cycles of V^in) and O",^, are used
(3 4b), (3 5b)
Table 3.3 Types of output functions considered X,, Y,- i-lh element of the system state vecior X.
and the output vecior Y_, respectively; p phase within the year (0=Jan, 1 l=Dec)
The output function f\ corresponded to the standard output function typically used in
monthly weater generators. Function /^ implements a skewness-reducing transforma¬
tion and was included with the aim to improve the simulation of monthly precipitation to¬
tals, which often show strongly skewed distributions (e.g., FLIRI, 1974; GARRIDO et
al, 1996) Functions fj and fit were similar to f\ and /il, respectively, but aimed at
reducing the number of needed parameters by adopting a more parsimonous description
of the annual cycles of the output variable's long-term means and standard deviations.
For all model Types I-IV we investigated three different model variants (A-C) which
were given by different combinations of output functions according to Table 3.4.
Model Variant Temperature Precipitation
A u u
B f\ /lL
C h fa.
Table 3.4 Output functions used in different variants of stochastic models (cf Table 3 3)
All considered output functions were defined by the equation:
Yi(k) = Zi(k)-a,(P) + ui(P) (3.3)
where Y, denotes the i-th element of the output vector Y, Z, the i-th element of an auxil¬
iary vector Z, and n, and a, are the expected value and standard deviation of an Y,.
In the output functions /( and /1L, the u, and ct, were given as
46 Chapter 3
H.(p) = m,(p)' °.(p) = sKp) (3-4a)
where the m, and s, were calculated according to Eqs. (3.6) and (3.7). In output func¬
tions /2 and /2L, the ul(p) and a1(p) were descibed by the first few harmonics of their
annual cycles according to:
ne.
8|(p) = ae.o+ X < ae.j Cos(2njp/12) + beiJ Sin(2nip/12) ) (3.4b)
J=i
Here e stands for u or o, an10 is the annual mean of 6, n6l is the number of harmonics
used to represent its annual cycle, and aa,, and b9lJ are parameters estimated according to
Eqs. (3.6) to (3.8c). In the present study were used throughout n^, = n0| = 2 for T, and
V = no2 = 3 for P.
For the output functions fi and /2 we used
Z|(k) = X,^) (3.5a)
whereas in functions /il and /2l, which assume a log-normal distribution for skewed
outputs Y,, we used the back-transformation
+Exp{fl(p)Xl(k)+g,(p)}+h,(p) ifSKEW[Y,(p)] > 0
Zl(k)=i Xl(k) ifSKEW[Y,(p)]= 0 (3.5b)
-Exp{f1(p)-Xl(k)+g1(p)}+h1(p) if SKEW[Y,(P)] < 0
Here fl(p), g,(p), and h,(p) are parameters estimated according to Eqs. (3.11)-(3.14), p is
the phase within the year, and SKEW[V] the skewness of the random variable V.
3.2.3 Parameter Estimation
All model parameters were determined for the period 1951-1980, and separately for each
location.
Estimates m,(p) and s,(p) for the expected values u.,(p) and standard deviations c,(p) of the
output variables Yl(p) (i=1..2 for T and P, respectively) were obtained acording to:
m.(P)=<y.(p.t)>-E[Y1(p)] (3.6)
SKP)= ( ^ < y,(P.O
-
mKP)>2 )"2 - STDEV[Y,(p)] (3.7)
The Temporal Aspect. Stochastic Simulation 47
where y,(p t)denotes a measured time series, < yl(p t)
> is the arithmetic mean over all
timepoints t, and n is the sample size. The parameters a^ and b9lJ of the output func¬
tions /2 and /2L were estimated according to (see e.g. Box & Jenkins, 1976, p. 36):
ae.o
a8ii
2'2
T2X v'(p>p=l
2'2
T2S V'(P>p=l
12
Cos(2icjp/12)
bfl„ =
'e,j
= J2"X v1(p)Sin(2)tjp/12)p*i
(3.8a)
(3.8b)
(3.8c)
for j < 5, where 8 and vl(p) stand for n and ml(p), or a and sl(pJ, respectively.
The matrices A and B were estimated using transformed monthly time series y,(P>t) or
y1(p,). The transformations applied were as follows. For the output functions /[ or /2used was
„• _y_iip,t)' "'(Pi
y'<p-«>-c1(p)
(3.9a)
where the u.l(p) and cl(p) were obtained according to Eqs. (3.4a) for fi and (3.4b) for /2.Note that a y1(p t)
had exactly zero mean and unit variance only when output function /|was used. For the output functions /il and /2l used was:
vKp.t) :
where
L"{+y,(p,t)-hKp)}-g,(P)fi(p>
yi(p.o
Lnl-yi(p.t)-h.(p)}-gi:M
<p)
if skew1(p) » 0
if skewl( jsmall
if skewl(p) « 0
(3.9b)
skew.(pl = s,(p3) <y,(p.,) -m,(p) >3 = SKEW[Y,(p)] (3.10)
Here tn1(p) and s,(p) are the estimated (Eqs. 3.6 and 3.7) mean and standard deviation of
y.jp.i), respectively. The condition "skewl(p) small" was defined to be true if a x2-test
(e.g., KREYSZIG 1977, p. 229) yielded at least a 10% probability (p-value > 0.1) that the
measurements y ](p t)stemmed from a normal distribution. Otherwise (first or third case
in Eq. 3.9b) a log-normal distribution for the random variable 7,^ was assumed.
48 Chapter 3
We discuss the parameter estimation only for the case skew,(p) « 0, since a formallyidentical procedure was also applied in the case of a negative skewness. The onlydifference was that, in order to obtain a positively skewed variable with the same mean as
the original variable, all negatively skewed y,(pt) were replaced by 2(fh,(p) - yl(Pil)).
If hl(p) is known, the maximum likelihood estimators for g,(p) and fl(p) are (JOHNSON &
KOTZ, 1970, p. 120):
g.(P)= < yi(P,t) > (3-iD
f,(p) = (i<y,(P,o-g.(P)>2),/2 (3.12)
For hl(p) we considered the range of values given by
hl(p) = MIN(y1(pjt)) - k-s1(p) (3.13)
when k was varied within [0.05 .. 5.0]. Here MIN(y,(pt)) denotes the minimum value in
the time series y,(Pit)- For each h,(p) we estimated the associated gl(p) and fl(p) accordingto Eqs. (3.11) and (3.12), and then selected the combination of parameters fl(p), g,(p),
hl(p) which maximized the likelihood function
Xl(p) = <Ln{pK(=(pt))}> (3.14)
where <•> is again the averaging operator, and pK(x) is the log-normal probability density
function p obtained for a given k, evaluated at point x.
The system matrices were estimated as follows. For models of Type II was
B(p) = EIGVEC[r0(p)] Sqrt(EIGVAL[r0(p)]) (3.15)
where EIGVEC[M] = [E],E2,..EN] denotes the NxN matrix of the eigenvectors of M,
EIGVAL[M] is a NxN diagonal matrix containing the eigenvalues of M, and rap) is the
lag-zero covariance matrix of the y, (respectively y,) at phase p. The elements yyj(p) of
the lag-v covariance matrix were estimated as
^j(p) = (V«P,0 " rai(P>Hvi(P-v,0 - mj(p.v))' « COV[v,(p),Vj(p.v)] (3.16)
where vl(pt) stands for y,(p>t) or y,(Pit),"•" is the vector product, and ml(p) is the mean of
vi(p,0- When output function fl was used for all output variables, Tv( , corresponded to
the lag-v correlation matrix of the observed time series y,(p,t).
20)(3EqfromobtainedA(p)theusing18)(3Eq
solvingbydeteminedfinallywereB(p)The0>qandq(j-l),=1q(i-l)+l,=kwhere
20)(3Sin(2it^p/12))+(Cos(2jtc,p/12)Z"ki+«ki=
a.j(p)
q
toaccording
spacestateoriginalthetobackAprojectingbyobtainedthenwereA(p)matricesThe
17)(3Eqtoaccordingdeterminedwas1)](3Eq[cfAR(l)-processcorresponding
theofAmatrixsystemMxMTheJ'i+2q(k)jnl(k).i+2qJN3(k)'J'2(k)'J|'(k)'[J|
=JvectorstateM-dimensionalphase-independent,new,adefinejJ,variablesThe
10equaledM
ei2,=qthroughoutusedwestudypresenttheIn?,or£,anforstandsV,andstep,
timecurrenttheiskvariables,newthedenoteq,to1fromrunning£,with^,J,theHere
19c)(3V,(k)Sin(2^k/12)=J,2^i(k)
19b)(3=Vl(k)Cos(2^k/12)J.24(k)
19a)(3=V,(k)J,,(k)
1/12frequencyfundamentaltheofharmonicsFourierqfirstthebyspanned
sub-spaceM=Nx(l+2q)-dimensionalaontoprojectedfirstare-spacestate11)p=0
N,(i=INxl2-dimensionaIadefinewhich-variablesoriginaltheprocedure,this
Following1994)GARDNER,alsosee(1981,BARNETT&HASSELMANNbyproposed
procedurethetoaccordingdeterminedweremodelsIVTypeoftheofA(p)matricesThe
Sqrt(L)E=B(p)thatsuchEIGVALfQ],
L=andEIGVEC[Q]=EwithETSqrt(L)Sqrt(L)Edecompositionaexiststhere
symmetricandrealisB[p)B(p)=Qsince15),(3Eqinro(p)ofcasethetoSimilarly
is)(3rUD)a(d)
-
rn(D)-
rl(D)rn(D)r,(0)-rn(D)-Bin)B(d) lip)'"(p)
~
0(P)'
"~
ftp)'0<P)'
l(P)'
~
ro(p)
=B(p)l(p)
o(p)up)''Vp)- 1?)(3ri(D>rO(i)A(Dl=
as55)p1976,Jenkins,&Box,g(eequationsYule-Walkergeneralized
thetoaccordinggivenwereB(p)andA(p)matricestheIIITypeofmodelstheFor
49SimulationStochasticAspectTemporalThe
50 Chapter 3
3.2.4 Bioclimatic Variables
The performance of the stochastic models was assessed by considenng three bio-climatic
variables derived from monthly weather data monthly winter minimum temperature
(TWinMin), annual growing degree-day totals over the temperature threshold of 5 5 °C
(GDD) and an index for the annual drought stress (DSI) applicable to forest vegetation
(cf Fig 3 1) These vanables were chosen because each one of them covers a distinct
aspect of trie annual weather, and because they are of interest to many ecosystem and bto-
geographical studies
Monthly winter minimum temperature for year y was computed as
TWinMiny = Min(T(Dec)y i. T(Jan)y, T(Feb)y) (321)
Due to the non-lineanty introduced by the minimum function, TWinMin is particularly
sensitive to the accurate simulation of T in the winter months
Annual growing degree-day totals were computed according to the procedure given by
FISCHLIN et al (1995) as
Dec
GDDy = XGDD(m)y (322a)
m=Jan
GDD(m)y = 30 5 Max(T(m)y - Tthresh) + 6(T([n)y) (3 22b)
where Ttnresh = 5 5° 8(T) is an empirically derived function which corrects for the bias
introduced by the monthly resolution, the largest values (-60 °Cd) are returned for T close
to Tthresh (for details see BUGMANN, 1994) GDDy is sensitive to the simulation of T in
summertime, where the largest contributions to the annual GDD sum occur, and in the
transition seasons, where at mid-latitude locations T is typically close to Tthresh
The drought stress index used was defined following BUGMANN & CRAMER (1997) as
Dec
X EoffSoil(m)y
DSIy = 1 - -^ (3 23)
/. "^Demand(m)ym=Jan
where EoffSol|(m)y is an estimate of the water transpired by the vegetation, and
^Demand(m)y ,s an estimate of the demand for moisture from the soil
The Temporal Aspect Stochastic Simulation 51
The DSI was used to assess the stochastic model's ability to simulate precipitation and its
interaction with temperature It was calculated according to the simple "bucket' -type
model for the soil water balance proposed by BUGMANN & CRAMER (1997), as follows
The available soil moisture at time step k was determined as (indices m and y are omitted
for clarity)
SM(k) = Max( 0 0, Min( SM(k ,,+ P,oSoll - EoffSo„ ,
b ) (3 24a)
where k denotes the current time (time step 1 month), b is the'
bucket" size, and
Pioso.i =P-EoffVeg (3 24b)
Ecrveg = M'n( k,«.pt P PET ) (3 24c)
^Demand = PET - EofrVeg (3 24d)
MsuPply =k-^^ (3 24e)
EoifSoil = Mln( MSupp|y, MDemand ) (3 240
AET = EofrVeg + EoltSoil (3 24g)
Here PloSo,| is the proportion of monthly precipitation P which reaches the soil, EoffVeg is
the evaporative loss due to interception, as controled by the parameter k,cpl, PET is the
monthly potential evapotranspiration (see below), MSupp|y is the monthly suply of water
from the soil, which is controled by the maximum evapotranspiration rate kcw, and AET
is the monthly actual evapotranspiration in mm/month PET was computed according to
THORNTHWAITE & MATHER (1957) as
PET = [<x(m)+p(m) MIN(A., 50)][^MAX(0,T)a] (3 25a)
Dec
h = X MAX(0, 0 2T(m))1514 (3 25b)m=Jan
a = 6 75 10 7 h3 -7 71 10 5 h2 + 1 792 10 2 h1 +49 239 10 3 (3 25c)
where a = 1 and P = O(io 2) are parameters used to correct for sun angle and day length
depending on latitude X, and h is a "heat index" calculated from long-term monthly mean
temperatures T
The following parameter values were used for all simulations b = 150 mm, kcw = 120
mm/month, and k,c ,= 03 The parameter h was calculated at each location for its 1951-
1980 baseline climate The initial condition for the soil water model was SM(o> = b
52 Chapter 3
3.2.5 Statistical Tests
The data base used to assess the performance of the stochastic models was generated as
follows Firstly, all available measured monthly weather data were used to denve time
series of bioclimatic variables at the 89 locations considered Secondly, for each loca¬
tion, type of stochastic model (I-IV), and variant of output functions (A-C) 100 years of
weather data were generated by means of stochastic simulation All simulations started in
January, and the initial conditions X(o) for the models of Types HI and IV were given by
the long-term mean climate of December Finally, the simulated weather data were used
to compute corresponding time senes of bioclimatic vanables
Data Period Definition # Years MinYrs # Stations
Baseline 1951 1980 30 20 89
Test Period 1 1921 - 1950 30 20 87
Test Period 2 1884-1950 u 1981 1993 70 50 59
Table 3.5 Overview of the data used to compare distributions of bioclimatic variables as derived
from measured and stochastically simulated weather inputs # Years number of years within period,MinYrs minimum years of data required to use a station for the comparison of distributions, # Stations
number of stations with at least MinYrs of data
Model performance was measured by comparing the distributions of stochastically sim¬
ulated bioclimatic vanables with distributions derived from measured weather data in the
two test penods defined in Table 3 5 Note that none of the test penods overlapped with
the baseline penod (1951-1980, see also Fig 3 1), which had originally been used to fit
the stochastic models. The simulated distributions used for comparison with Test
Periods 1 and 2 were defined by the first 30, respectively last 70 years of simulated data
To measure the similarity of any two distributions of bioclimatic variables we considered
the p-values obtained from two non-parametric tests, the Mann-Whitney U-Test (U) and
the Kolmogorov-Smyrnow test (KS) For either test the p-value gives the probability
that the two compared data sets stem from the same parent distribution The U-Test is
sensitive to differences of the medians, less sensitive to differences in skewness, and
insensitive to differences in variance (e g , SACHS, 1984, p 293) The KS-test is the
sharpest homogeneity test and and covers differences in the mean, the median, the
dispersion, the skewness, and the excess of two distributions (SACHS, 1984, p 291)
The following summary results over all locations were computed Firstly, average model
performance was assessed by computing separately for each combination of bioclimatic
variables, types of stochastic models (I-IV), variants of output functions (A-C), test pe¬
riods (1 or 2), and statistical tests (U or KS) the median p-value (pvmed) from all
The Temporal Aspect Stochastic Simulation 53
locations In order to assess the representativity of the baseline climate, the pvmed were
also computed for the distributions of bioclimatic variables which were derived from
measured weather data in the baseline period 1951-1980
Secondly, in order to assess the performance of the different types (I-IV) of models rela¬
tive to each other we determined the numbers N of cases in which a model of a given type
yielded a larger p-value than each other type of model The N were determined using the
p-values from both test periods The significance of a given N was assessed based on
the following reasoning if the two types of models were equally good (null hypothesis),N would be binomially distributed with p = q = 0 5 and n = 87+59 = 146 (cf Table 3 5),
and, since npq = 36 5 > 9, approximately normally distributed with p. = np = 73 and a =
Sqrt(npq) = 6 Evaluation of the normal probability density function then yields that if
the value 100 N/n deviates by more than 5 3% (8 2%) from 50%, the two types of mod¬
els perform significantly different at the 90% (95%) confidence level (one-sided test)
3.2.6 Modeling and Simulation Tools
Two main computer programs were used, which were both written by the authors For
the management and analysis of monthly weather data, the calculation of bioclimatic vari¬
ables, and the estimation of the parameters of the stochastic models we used the program
ClimShell VI 4d The stochastic simulation of bioclimatic variables was carried out
with the program BCSIM VI 0 Both programs were developed based on the modular
software development and modeling environments "Dialog Machine" (FISCHLIN, 1986,
FISCHLIN & SCHAUFELBERGER, 1987), "RAMSES" (Research Aids for the Modelingand Simulation of Environmental Systems, FISCHLIN, 1991, FISCHLIN etal 1994) and
'
RASS" (RAMSES Simulation Server, THOENY et al, 1994, 1995) The Mann-Whitneyand Kolmogorov-Smyrnow tests were computed with the software package Splus V3 3
3.3 Results
The median p-values (pvmed) obtained from all tests are shown in Table 3 6 It can be
seen that model performance generally increased with increasing sophistication of the
models, and that the models of Type IV and III yielded generally the best results
The model's capability to reproduce the test distributions depended on the variable con¬
sidered The mean pvmed-values (averaged over all model types and output function vari¬
ants) were 0 31 for TWinMin, 0 12 for GDD, and 0 17 for DSI The pvmed-values for
the distributions of bioclimatic variables that were derived from stochastically simulated
weather inputs were mostly below the ones that were computed from measured weather
54 Chapter 3
TWinMin GDD DSI
Test Period1
U KS
Test Period2
U KS
Test Period1
U KS
Test Period2
U KS
Test Period1
U KS
Test Period2
U KS
A I
n
m
IV
40 39
31 39
33 44
.44 .50
17 14
.3 7 26
29 24
34 .31
11 19
09 09
17 16
.17 24
10 04
06 06
13 12
.14 16
33 03
30 02
36 03
.37 .03
16 00
19 01
.28 00
23 0 1
B I
n
m
rv
29 24
36 39
36 39
.41 .39
30 16
31 21
31 25
33 .40
16 20
13 07
12 16
.18 24
08 05
06 03
.09 .10
07 10
24 .06
.3 7 04
32 03
33 03
27 05
.33 07
23 .09
26 05
C I
ii
m
IV
23 24
28 33
31 39
.34 .39
21 09
24 15
36 24
42 .30
.17 20
13 11
16 20
14 n
05 03
06 02
.09 08
09 .08
25 06
35 03
.41 .07
34 05
22 0 8
34 06
.35 .08
31 05
Baseline 40 46 50 44 23 39 21 21 43 09 48 14
Table 3.6 Medians of the p values (pvmed0 obtained from the comparison of the distributions of
bioclimatic variables as derived from measured and stochastically simulated monthly weather data at 89
European chmatological stations (cf Table 3 1) TWinMin - monthly winter minimum temperature
GDD - annual growing degree day totals, DSI - annual drought stress index Test Period 1/2 periods of
measured weather data used to derive test distributions of bioclimatic variables (cf Table 3 5) U pv,^refers lo p-values obtained from the Mann-Whitney U-Test KS pvmed refers to p values obtained from
the Kolmogorov Smyrnow test, I IV types of stochastic models considered (cf Table 3 2), AC
variants of functions used to calculate outputs of stochastic models (cf Table 3 4), Baseline
distributions from test periods 1/2 were compared with distributions for the baseline period 1951 1980
The highest value(s) obtained within a column for each model variant A-C is printed in bold
data the mean pvmed-values in the 'Baseline" case were 0 45 for TWinMin, 0 26 for
GDD, and 0 29 for DSI
From Table 3 6 can be seen that for TWinMin the Type IV models yielded for all except
one case (Period 2/U-test) the best result The variants A and B gave generally higher
pvmed-values than variant C compared to variant A, performance in variant C was better
only for the Test Period 2 (U-test) for model Types I, 111 and IV, but was other-wise less
good in 12 out of 16 (= 2 test periods x 2 statistical tests x 4 model types) cases
Compared to variant B, variant C was worse in 11, and better only in two cases
For GDD, the test distributions were under variant A in all cases most successfully repro¬
duced by the Type IV models (Table 3 6) Under variant B, for two cases the best
results were obtained with the Type III models Variant C appeared again to perform less
well than variants A and B Compared to these variants it gave lower pvmed values in 10
respectively 7 out of 16 cases, whereas better results were obtained only in 5 cases each
For DSI, under variant A for three out of four cases the best result was obtained with the
Type IV models However, for variants B and C the highest pvmed values were obtained
The Temporal Aspect Stochastic Simulation 55
GDD DSI
II III I II III
Fig 3.2 Comparison of the performance of the Type IV models relative to the Type I
III models (ct Table 3 2) Each panel shows the percentage of cases (relative to the 50%
line) in which a Type IV model reproduced for a given output function variant (A-C, cf
Table 3 4) a test distribution of a bioclimatic variable better than the respective other type
of model For each comparison were used a total of 146 test distributions as derived from
independently measured weather data at 89 European chmatological stations (cf
Table 3 I) and for two different test periods (cf Table 3 5) Bars above the dashed lines
indicate significantly better perfomance of the Type IV models at the 95% confidence level
(one sided test) Black bars comparisons with the test distributions were based upon the
Mann Whitney U test Grey bars Kolmogorov-Smyrnow test TWinMin - monthlywinter minimum temperature, GDD - annual growing degree-day totals, DSI - annual
drought stress index
with model Types I-III Under the variant B performance increased in 10 out of 16 cases
as compared to variant A Improvement was even more pronounced under variant C,
where in all except two (Test Period 2/U-test) cases higher pvmed values were obtained
than for variant A
The pvraed-values presented random vanables, such that some of the found differences
between variants or model types may not be significant The overall picture obtained
from Table 3 6 is however confirmed in Fig 3 2, which compares the performance of
the Type IV models relative to the three other types of models
It can be seen that for TWinMin the Type IV models performed according to both, the U-
and the KS-tests, significantly better than the Type I models The Type IV models ap¬
peared to perform also systematically better than the Type II models, but this result was
56 Chapter 3
a,, (Temperature) — Type III a22 (Precipitation) — Type III
— Type IV — Type IV
Fig 3.3: Comparison of the annual cycles of the diagonal elements CL\ i and CC22 of the
system matrices A (cf. Eq. 3.1) of cyclostationary, first-order autoregressive stochastic
models of monthly weather variables at selected European locations. The a,, refer to the
state vector X. = (Xj X2/ = (T P)', where T and P denote standardized monthly mean tem¬
peratures (T) and monthly precipitation totals (P). Thin lines: matrix elements were
estimated by considering each month separately ("Type III" models); thick lines: matrix
elements were estimated in a subspace spanned by the first two Fourier harmonics of the
annual cycle according to the approach proposed by HASSELMANN & BARNETT (1981)
("Type IV" models). All values are given in standard deviations of the respective variable.
Analysis interval was 1951-1980.
generally not significant at the 95% level (cf. dashed lines in Fig. 3.6). With regard to
GDD, the Type IV models performed significantly better than the Type I models only ac¬
cording to the sharper KS-test. However, according to both tests, they clearly perfomedbetter than the Type II models. For DSI no clear improvement was obtained due to the
use of Type IV models. Note, however, that the general improvement obtained under
variants B and C as compared to variant A can not be discerned from Fig. 3.2, since in
this figure the different types of models are compared only relative to each other. Except
for two cases (U-tests for TWinMin, model III/A, and DSI, model III/B), no significant
differences in the performance of the Type III and Type IV models were found.
The Temporal Aspect Stochastic Simulation 57
Fig 3 3 compares for selected locations the annual cycles of the diagonal elements ax [
and a22 of the system matrices A (see Eqs 3 1, 3 17, 3 20) used in the Type III and
Type IV models It can be seen that the a,, of the Type IV models followed smoothly the
annual cycles that were obtained for the Type III models The memory terms were gen¬
erally larger for temperature (i=l) than for precipitauon (i=2) Further it can be seen that
the amplitude and form of the annual cycles of the a,, varied across locations However,
for temperature in all cases distinct minima were obtained in the transition seasons Not
shown in Fig 3 3 are the elements a12 and a2|, which were generally of same
magnitude as the a22 The a,, of the Type III models were very close to the respective
lag-1 autoconelation coefficients (not shown, cf Eqs 3 9a, 3 16) of the two vanables
3.4 Discussion
Temporal autocorrelation on the monthly time scale appears to be a non-neglectable
feature of mid-latitude weather The autocorrelation found in local weather variables on
the one hand relates to persistent anomalies in the large-scale circulation, such as blocking
events or long-lasting periods of zonal flow (e g ,JACOBEIT, 1988, KLAUS, 1993) The
larger autocorrelation coefficients obtained for temperature as compared to precipitation
(Fig 3 3) probably reflected the generally stronger dependence of local temperatures on
the large-scale circulation On the other hand, autocorrelation depends also on smaller-
scale processes, which may enhance (e g , long-lasting inversions in valley locations) as
well as reduce (e g ,convective precipitation, or unsteady flow in the lee of a mountain
range) the temporal stability of the local weather Such processes may also have con¬
tributed to the found differences between vanables, seasons and locations (Fig 3 3)
The generally better performance of the Type III and IV models as compared to the
simpler Type I and II models (Table 3 6, Fig 3 2) demonstrated that inclusion of a
memory term in stochastic models of monthly weather allows to enhance the statistical
accuracy of simulated weather sequences and bioclimatic variables The proposed auto-
regressive models thus presented an improvement compared to existing solutions, which
discard autocorrelations (e g ,FISCHLIN et al, 1995), or generally any conelations (e g ,
SOLOMON & WEST, 1987, KRAUCHI, 1994) between monthly weather variables Some
authors (e g , KRAUCHI, 1994) have argued that these correlations can be neglected
because they are generally too small to be statistically significant Though this may be
true for individual months, our analysis revealed distinct annual cycles for the auto-
(Fig 3 3), or cross-correlation (not shown) coefficients, and these would not occur if
the correlations were purely random Furthermore, our tests with independent data
(Table 3 5) clearly showed that in the majonty of the cases the conelations carry valid
58 Chapter 3
information (see e g comparison with the Type I models in Fig 3 2), and hence
significantly help to improve stochastic weather generation
The improvements obtained due to the use of Type III or IV models were largest for
TWinMin and GDD This was because these variables depended strongly on the joint
probability distributions of monthly mean temperatures (see Eqs 3 21,3 22), and these
joint distributions were much better represented when conelations between successive
months (Fig 3 3) were included in the stochastic models In contrast, for DSI the inclu¬
sion of monthly autocorrelations yielded a clear improvement only under the output
function variant A However, this improvement was small when compared to the im¬
provement achieved due to the use of a log-normal transformation for precipitation in
variants B and C (Table 3 6) This was because DSI depended mainly on precipitation
(Eqs 3 24), and this variable showed generally only small autoconelation (Fig 3 3),
but was otherwise found to be strongly skewed (see also discussion below)
The use of different vanants of output functions (Table 3 4) showed contrasting effects
for the different bioclimatic variables The use of variant B did not appear to have any
systematic effect on the simulation of TWinMin and GDD This appears plausible since
this variant mainly affected the simulation of precipitation (Table 3 4) The found
differences betwen vanants A and B therefore probably reflected but sampling variabilityof the pvmed Under variant C, however, the pvmed for TWinMin and GDD quite clearlydeteriorated This result reflected the threshold-effects that entered the calculation of
TWinMin and GDD (Eqs 3 21,3 22), and suggests that stochastic simulation of these
two vanables strongly depends on an accurate representation of the annual cycles of the
expected values and vanances of monthly mean temperature
A different result was obtained with regard to precipitation, where variants B and C
yielded a strong improvement as compared to variant A (Table 3 6) The main reason
was that the monthly precipitation data used for the calculation of this variable were
strongly skewed from the 1089 (= 89 stations x 12 months) 30-yr distributions investi¬
gated, 85% (4%) were positively (negatively) skewed and differed significantly from a
normal distribution (a = 90%) This result conforms with the findings by other authors
(e g , FLIRI, 1974) and suggests that a skewness-reducing transformation should be gen¬
erally used in stochastic models of monthly precipitation The equally good (or even
slightly better) results obtained for variant C as compared to variant B (Table 3 6) further
indicated that the annual cycle of precipitation is generally better represented by means of
its first few Fourier harmonics rather than twelve monthly values
Which types/variants of stochastic models should be used best to simulate the three bio¬
climatic variables considered in this study9 The results obtained for TWinMin and GDD
(Table 3 6, Fig 3 2) clearly ruled out the Type I and II models For DSI the best re-
The Temporal Aspect Stochastic Simulation 59
suits were obtained under variant C, and this would favour the Type III models
(Fig 3 2) However, these models required much more parameters than the Type IV
models (cf Table 3 1) Furthermore, variant C appeared suboptimal with regard to
TWinMin and GDD (Table 3 6) This lead to the definition of a new variant, variant D,
in which for precipitation we still used the output function /2L (see Table 3 2), but for
temperature only output function /, (as in variants A and B) Indeed, simulations under
variant D yielded for the Type IV models and the four statistical tests pvmed-values (cf
Table 3 6) of 0 37, 0 39, 0 30, and 0 25 for TWinMin, 0 15, 0 24, 0 07, and 0 08 for
GDD, and 0 36, 0 03, 0 39, and 0 11 for DSI Hence, compared to variant C, variant D
yielded for TWinMin a slightly reduced, but otherwise for GDD and DSI a generally im¬
proved performance Moreover, under this new variant the Type IV models performed
again equally good as the Type III models, and in several cases significantly better than
the Type I and II models (results not shown, cf Fig 3 2)
Hence, use of the Type IV models in combination with the output function variant D
appears to be a good compromise if all three bioclimatic variables are needed simultane¬
ously On the other hand, if only TWinMin and/or GDD are needed, we recommend the
use of the Type IV models in combination with the output function variant A (cf
Table 3 6)
However, it should be noted that even the best stochastic models did not yield as good re¬
sults as the direct use of the "Baseline" weather record (Table 3 6) A closer analysis
showed that the "Baseline' distributions yielded for TWinMin in 55%, for GDD in 65%,
and for DSI in 60% of all cases better results than the the Type IV models This was
mainly because the simulated bioclimatic variables showed a too small variability The
stochastic models per definition (Eq 3 3) reproduced the variance of the monthly weath¬
er variables Hence, additional improvements would probably require a more accurate
simulation of further higher-order moments (such as the skewness and curtosis) of the
weather variables' distributions This could possibly be accomplished by using a
second-order autoregressive process for temperature and/or a more sophisticated skew-
ness-reducing procedure for precipitation Even better results might also be obtained by
optimizing for each location the numbers of harmonics used to fit the system matrices of
the Type IV models and to represent the weather variables' annual cycles
Our results demonstated that the choice of weather generation scheme involves a trade-off
between the statistical accuracy that can be attained, and the parsimony of the simulation
approach The optimal choice further depends on data availability if only a relatively
short record (say, 20 years or less) is available, use of a stochastic simulation approach
may be superior to the direct use of the weather record This is because the stochastic
models require much less parameters, and thus can be expected to be more robust (cf
Fig 3 3) For example, note that if 20 years of measurements are used to describe a
60 Chapter 3
location's temperature and precipitation climate this conesponds to 2x12x20 = 480
parameters In contrast, the Type IV models (vanant D) required a maximum of only 99
parameters (the exact number depended on the number of fitted log-normal distnbutions)
Simulation experiments with impact models can be used to determine the optimal weather
generation approach with regard to a particular application For instance, preliminary
simulations with the forest succession model FORCLIM (BUGMANN, 1994, FISCHLIN et
al, 1995) at four Alpine locations (Basel, Bern, Bever and Sion, cf Table 3 1) revealed
sensitivity to an improved simulation of the dnving weather inputs simulated forest
compositions at Bever and Sion differed significantly when the Type Ul and IV instead of
the Type I and II models were used The sensitivities appeared reahstic since in at least
one case (Bever) the more accurate weather generators also yielded more realistic forest
compositions Additional expenments where the forest model is dnven with measured
weather data could be conducted to determine in as far the Type IV models are actually
sufficient One possibility to construct realistically auto- and cross-correlated input tune
senes for such expenments would be to sample at random annual cycles of monthly
temperature/precipitation pairs from the weather record
3.5 Conclusions
From our stochastic simulations and analyses at 89 European long-term chmatological
stations we concluded
(1) The inclusion of a memory term in stochastic models of monthly weamer generally
improves the statistical accuracy of simulated monthly mean temperatures, whereas use
of a log-normal transformation strongly improves the simulation of monthly precipitation
totals
(2) By estimating the monthly system matrices in a sub-space spanned by the first two
Founer harmonics of the annual cycle, according to the approach proposed by HASSEL¬
MANN & BARNETT (1981), cyclostationary autoregressive models of monthly weather
can be further unproved, and the number of needed parameters substantially reduced
(3) To accurately simulate monthly winter minimum temperatures, annual growing
degree-day totals, and the studied annual index of drought stress, the input data should be
generated with a stochastic model, which, m addition to the above properties (1) and (2),
uses a smoothed representation of the annual cycle of precipitation by means of its first
three Founer harmonics
61
4. Synthesis and Application: Assessing Impactsof Climatic Change on Forests in the Alps
Published as
Fischlin A & Gyalistras D (1997) Assessing impacts of climatic change on
forests in the Alps Global Ecology and Bwgeography Letters 6 19 37
4.1 Introduction
In a world subject to global change, mountain forests may play an increasingly important
role They moderate the water cycle, stabilise soils, protect human settlements, shape the
landscape, provide wood and other products, and begin playing a key role in preserving
biodiversity while cunent trends of low-elevation deforestation continue Assessing pos¬
sible impacts of climatic change on these important ecosystems poses a particular
scientific challenge, owing to the complex effects of mountainous topography both on
local climates and on their forests
Neither ecological theory (e g , SOLOMON, 1988) nor experimental approaches (e g ,
KORNER, 1993) enable us to predict precisely the consequences of particular climatic
changes on ecosystems On the other hand carefully designed simulation models (e g ,
Fischlin, 1991), which embrace what is known from theory and expenments, not onlymake it possible to trace the consequences of a given set of assumptions, but also render
the underlying theory and experimental data accessible for scrutiny Thus it may be
argued that models cunently provide the most comprehensive means to assess impacts of
climatic change on ecosystems and are likely to remain so for some time
Simulation models are also essential tools for impact assessments according to the IPCC
Guidelines (Carter et al, 1994), which distinguish these basic methods expenmen-
tation, impact projections, empincal analogue studies, and expert judgement In the case
of forests, experimentation is often impractical due to the longevity of the dominant tree
species, to follow conventional experimental approaches, forests would need to be inves¬
tigated over several centuries Forest ecosystem models are therefore essential, althoughthe question remains How much can they actually accomplish'''
62 Chapter 4
Several modelling approaches are currently available to assess the impact of climatic
changes (e g ,SHUGART, 1990, KlRSCHBAUM & FISCHLIN, 1996), offering varying
advantages or disadvantages, depending on emphasis plus resolution in time, space, and
structure
In the first of these approaches, climate and vegetation are empirically correlated by
assuming equilibrium conditions (HOLDRIDGE, 1947, 1967, BOX, 1978, 1981) If such
relationships can be quantified, efficient tools to project vegetation responses in space
result (e g ,EMANUEL et al, 1985, KlENAST et al, 1987, BOX & MEENTEMEYER,
1991, BRZEZIECKI et al, 1993, CRAMER & LEEMANS, 1993, 1995) This approach
assumes a static relationship between climate and climax vegetation, however, which may
be perfectly adequate to predict a potential natural vegetation in an undisturbed, distant
future world, but not necessarily the real, near-future vegetation
The second approach is to use ecophysiological models (e g ,SCHIMEL et al, 1990,
RASTETTER et al, 1991, McGUIRE et al, 1993, MELILLO et al, 1993, PARTON et al,
1993) which offer the advantages of being dynamic and able to reproduce ecosystem
responses to climatic forcings at a high temporal and sometimes relatively high spatial
(e g ,MELILLO et al, 1996) or qualitative resolution (e g ,
KlRSCHBAUM et al, 1994)
These models, however, often focus on particular chemical elements such as C or N
(e g ,RAICH etal, 1991, RAICH & SCHLESINGER, 1992) or on ecophysiological pro¬
cesses that lump together properties, such as species composition (e g ,SCHIMEL et al,
1990, PARTON et al, 1993) or canopy structure, hence providing only poor qualitative
resolution The associated up-scaling problems limit the usefulness of these approaches,
since feed-backs to the atmosphere by structural changes in an ecosystem must be
ignored unless the latter are modelled explicitly For instance, if the species composition
in a mixed-deciduous forest changes, this may greatly affect winter albedo, an effect
which can only be modelled if species composition is modelled explicitly Similar argu¬
ments can be put forward for other structural properties of forests, such as age structure,
stem density or dispersion, all affecting surface roughness
Some additional problems often apply to the above approaches In many instances
paleoecological studies of the effects of climatic change on ecosystems have demonstrated
that structural properties such as the species composition are of essential relevance
(DAVIS, 1986, 1989, 1990) This has been confirmed particularly for forests (e g ,
DAVIS, 1981, HUNTLEY & BlRKS, 1983, HUNTLEY, 1990, OVERPECK et al, 1991,
PRENTICE etal, 1991, SOLOMON & BARTLEIN, 1992, WEBB III, 1992, WRIGHT et
al, 1993) Since most phytosociological as well as ecophysiological approaches ignore
species composition (PERRUCHOUD & FISCHLIN, 1995), their usefulness for studies of
climatic change impacts appears to be limited
Synthesis and Application Assessing Impacts on Foresls 63
The third and perhaps most promising approach is the use of patch dynamic models,
which have already proved relatively successful in mimicking forest succession (e g ,
SHUGART, 1984) By operating at an intermediate level (high spatial, medium temporal,and medium qualitative resolution) they allow the relevant time constants of forest dy¬
namics to be included while still simulating important structural characteristics of forests
such as species composition or age structure
Patch models are able to mimic broad charactenstics of real forests under a wide range of
climates (e g ,WEST etal, 1981, SHUGART et al, 1992, SMITH et al, 1992) and are
thus of relatively high predictive power (SHUGART, 1984) Patch models have also been
successfully used to study forest responses to (i) past climatic changes (SOLOMON et al,
1980, SOLOMON & WEBB III, 1985, LOTTER & KJENAST, 1992, SOLOMON & BART¬
LEIN, 1992, FISCHLIN et al, 1998a, 1998b), (n) direct effect of CO2 (SOLOMON, 1986,
PASTOR & POST, 1988, KIENAST, 1991), or (111) possible future climatic changes
(SOLOMON et al, 1981, SHUGART etal, 1986, OVERPECK et al, 1990, BOTKIN &
NlSBET, 1992, SMITH etal, 1994)
With regard to the Alps, several authors have conducted modelling and evaluation studies
(FORECE, KIENAST, 1987, KIENAST & KUHN, 1989, FORCLIM, BUGMANN, 1994,
FORSUM, KRAUCHI, 1994) and have studied possible impacts of a changing climate on
Alpine forests (KIENAST, 1991, KRAUCHI & KIENAST, 1993, BUGMANN, 1994,
Bugmann & Fischlin, 1994,1996, Fischlin, 1995)
These studies have shown that Alpine forests may be sensitive to climatic change
However, they all relied upon relatively coarse climatic scenarios, which may have
hampered the internal consistency and potential realism of the resuhs Indeed, the highlyvariable topography of the Alps, their location in the transitional zone between the
temperate, Mediterranean and continental climatic regimes, the complexity of forests, and
the high input and precision requirements of forest patch models render the construction
of appropriate climatic scenarios a challenging task
A simple method often adopted is to adjust climatic parameters on an ad hoc basis, e g
by assuming a temperature increase of 1-4 °C within the next, say, 50-100 years (e g ,
IPCC, 1990, OZENDA & BOREL, 1990) While this approach may be useful for initial
sensitivity studies, its main disadvantage is that differences between possible patterns of
global climatic change, as well as the regionally differentiated responses to such patterns
must be completely ignored In complex terrain like the Alps, this approach is therefore
particularly likely to yield inconsistent results (GYALISTRAS et al, 1994)
An often recommended approach with fewer shortcomings (Carter et al, 1994) is to
rely upon simulations with global climate models, in particular three-dimensional,
64 Chapter 4
coupled general circulation models of the atmosphere (WASHINGTON & PARKINSON,
1986) and the oceans (SEMTNER, 1995) ("AO-GCMs") At least three different strate¬
gies to construct scenanos of regional climatic changes from GCMs have been proposed
First, GCM-output at a few gndpoints in the vicinity of the region of interest may be used
(e g ,KARL et al, 1990, SANTER et al, 1990) For instance BUGMANN & FISCHLIN
(1996) directly interpolated climatic scenarios from GCM grids to assess the impacts of
these scenarios on Alpine forests However, GCMs typically show very large errors at
individual model gndpoints (e g ,GROTCH & MacCracken, 1991) This is particular¬
ly true over a complex topography such as the Alps (GYALISTRAS et al, 1994)
The second approach is to use a physically-based regional climate model (RegCM) driven
by boundary conditions taken from a GCM (e g ,GlORGl et al, 1992) A major limita¬
tion is that even if a RegCM is run at very high horizontal resolution (say -20 km, MARI-
NUCCI et al, 1995), the simulated climatic changes can only be trusted at a spatial scale
above several RegCM-gndpoint distances (cf VON STORCH, 1995) Hence, many
mountain features (ridges, valleys, slopes), can not be adequately resolved Further, the
huge computational requirements of RegCMs severely restrict the construction of time-
dependent scenarios over the time spans needed to assess forest responses, i e several
centuries
The third approach is based on the interpretation of large-scale climatic changes as
simulated by GCMs (or RegCMs) using expert knowledge (e g ,ROBOCK et al, 1993)
or statistical models that relate regional climates to the large-scale atmospheric state (e g ,
VON STORCH et al, 1993) The latter approach appears to be practical, offering the
potential to satisfy all requirements simultaneously (GYALISTRAS et al, 1994) Several
authors have demonstrated that the application of empirical' downscaling relationships
can yield physically plausible and internally consistent first-order estimates of possible
climatic changes at least at regional scales (for reviews see e g ,KattenberG et al,
1996, GYALISTRAS et al, 1998) BUGMANN (1994), BUGMANN & FISCHLIN (1994),
and FISCHLIN (1995) have used statistically downscaled scenarios to assess forest
responses in the Alps, but they have used scenarios (GYALISTRAS et al, 1994) which
considered only winter and summer seasons instead of the entire annual cycle
This paper presents a general method to estimate possible impacts of climatic change on
forests in the complex Alpine terrain by combining global climate models, downscaling
techniques, and patch models For several carefully selected sites, each representing a
typical forest ecosystem, it will be shown that the proposed method provides consistent
and plausible (and otherwise unavailable) quantitative projections of possible changes in
key structural characteristics such as species composition of mountain forests, many
forests in the Alps are sensitive to these downscaled climatic scenarios given at a monthly
Synthesis and Application Assessing Impacts on Forests 65
resolution, and, forests within a relatively small distance may show a surpnstngly wide
range of responses to the same pattern of global climatic change
4.2 Material and Methods
Climatic change impacts were assessed at four study sites in the Alps (Table 4 1) For
each site CLIMSHELL (GYALISTRAS et al, 1994, GYALISTRAS & FISCHLIN, 1996) was
used to define baseline climates and to derive climatic scenarios from GCM simulation
results, and FORCLIM (BUGMANN, 1994, FISCHLIN et al, 1995) was used to simulate
forests' responses to the obtained climatic scenanos (Fig 4 1)
(GHGSc)
ForClim
EEnviron
ment
WnMin 1
GDD*P
Plants 1AET*DSI
*
r r tSA, FC X ( Ps J
h
Fig 4 1 Method used lo derive climatic scenarios (CLIMSHELL, GYALISTRAS et al,
1994, GYALISTRAS & FISCHLIN, 1996) and simulate forest responses (FORCLIM, BUG
MANN, 1994, FISCHLIN et al, 1995) Arrows represent data flow large - vectors or ma¬
trices, thin - scalar variable or parameter GHGSc, greenhouse gas emission scenario (in
this study 2xC02-equilibnum), GCM, General Circulation Model (simulation results used
in this study from CCC-GCMII, BOER et al, 1992), MODS, meteorological large-scale
(JESSEL, 1991, JONES & BRIFFA, 1992, BRIFFA & JONES, 1993) and local (BANTLE,
1989) observations, DTM, digital terrain model (in this study ARC/INFO data describing
Swiss Alps), curC, current climate data, CC, data describing climatic change, T, climat¬
ic parameters E[Tm], VAR[Tm] for monthly temperatures Tm, P, climatic parameters
E[Pml VAR[Pm] for monthly total precipitation Pm, p, COV[Tm,Pm], SA|, slope-
aspect index, FC, field capacity, X,latitude, TWinMin, winter minimum temperature,
GDD, annual growing degree-day totals, AET, actual evapotranspiration, DSI, annual
drought stress index, Ps, matrix of species parameters, FD, forest dynamics (e g , age
distribution, species compositions, leaf area index etc ) Downscaling is described in
Gyalistras etal (1994), Interpolation attempts to make best use of local/regional
meteorological measurements and digital terrain model, submodel FORCLIM-C controls
climatic change within FORCLIM, submodel FORCLIM-E computes the abiotic
environment, submodel FORCLIM-P simulates tree dynamics
Each case study site (Table 4 1) represents a particular ecoregton from North to South
Bern located on the Swiss Plateau is a typical submontane site, Bever represents sub-
alpine forests, St Gotthard represents an alpine site, currently above the timberline, and
Sion represents colhne belt conditions (cf ELLENBERG, 1988)
66 Chapter 4
4.2.1 Baseline Climates
The large-scale observations used to fit the downscaling models were gridded (5° x 5°
latitude by longitude) anomalies of monthly mean sea-level pressure (SLP, provided by
NCAR, see JESSEL, 1991) and near-surface air temperature (2mT, from the data set of
Jones & Briffa, 1992; Briffa & Jones, 1993) over the North-Atlantic/European
sector (40°W-40°E and 30°N-70°N). Local chmatological observations were extracted
from the data base of the Swiss Meteorological Institute (BANTLE, 1989).
Site Location Elevation
(m above
sea level)
annual
mean
temperature
CO
annual
total
precipitation(cm)
Current potential natural
vegetation
(Ellenberg & Klotzh, 1972)
A Bern Swiss Plateau
(Northern Alps)
570 86 99 Submontane mixed deciduous
forests dominated by beech
(Fagus silvahca L ) and silver
fir (Abies alba Miller)
B Bever Central/
Southern Alps
1712 1.6 82 Subalpine coniferous forests
dominated by larch (Larixdecuiua Miller) and arolla pine
(Pinus cembra L.)
C St.
Gott-
hard
Transition zone
Northern to
Southern Alps
2300 -1.0 207 Alpine belt,
above current timberhne
D Sion Central Alps 542 9.9 60 Colline mixed-deciduous oak
forest
Table 4.1. Characteristics and major current climatic parameters of the case study sites
The forest model ForClim uses a stochastic weather generator to simulate annual cycles
of monthly mean temperature (T) and precipitation totals (P) based on the following set of
site-specific, climatic input parameters: expected values of monthly mean temperatures
(E[T]) and precipitation totals (E[P]), and monthly 2x2 cross-covariance matrices
(COV[T,P]) for each month of the year (Fischlin et al, 1995).
For Bever, Bern and Sion all climatic parameters were computed from daily measure¬
ments for the period 1931-1980 (BANTLE, 1989). For St. Gotthard, which is -0.6 km
away from the nearest climatological station (at 2090 m a.s.l.), E[T] and E[P] were inter¬
polated using a further 39 stations for T (25 for P) up to a distance of 80 km (60 km for
P), whereas COV[T,P] was taken directly from the nearest station. To interpolate we
applied a linear regression separately for each month to predict E[T] and E[P] from ele¬
vation, and then adjusted the result based upon an inverse distance-weighted (IDW)
Synthesis and Application Assessing Impacts on Forests 67
interpolation of elevation-detrended expected values from the sunounding stations (cf
GYALISTRAS & FISCHLIN, 1995, GYALISTRAS & FISCHLIN, 1996) We used for E[P]
a linear, and for E[T] a piecewise linear, seasonally varying dependency on elevation
with a breakpoint at 1450 m
4.2.2 Climatic Scenarios
All climatic scenarios were derived from a "2xC02" global climate change expenment
simulated by the Canadian Climate Centre CCC-GCMII (BOER et al, 1992) The scenar¬
ios were constructed based on the statistical downscaling method of GYALISTRAS et al
(1994, GYALISTRAS & FISCHLIN, 1995, GYALISTRAS & FISCHLIN, 1996) as follows
(a) For each month Principal Component Analysis (PCA, PREISENDORFER, 1988) was
used to construct a smaller number of new, large-scale vanables (so-called principal com¬
ponents, PCs), from the 306 vanables contained in the SLP- and 2mT-fields of the pen-
od 1931-1980 Between 8 (winter months) and 14 (summer months) PCs were retained,
which explained -90% of the total interannual large-scale vanabthty for all months
(b) Canonical Correlation Analysis (CCA, BARNETT & PREISENDORFER, 1987), again
performed for 1931-1980, was used to establish separately for each location and month a
multivariate regression model to predict from the PCs simultaneous anomalies of local
monthly mean T and monthly Vp from their respective long-term means The CCA
regression parameters define pairs of large-scale patterns and conesponding local
responses (see VON STORCH et al, 1993, GYALISTRAS et al, 1994), which were all
plotted to allow inspection of the statistical relationships
(c) The monthly downscaling models denved from the CCA were applied to five years of
monthly mean large-scale anomalies as simulated by the CCC-GCMII under "2XCO2"
conditions relative to the long-term mean fields from a five year control (" lxC02") simu¬
lation Changes in the seasonal (winter = December, January, February, spring = March,
April, May, summer = June, July, August, and autumn = September, October, Novem¬
ber) mean SLP- and 2mT-fields were plotted to examine the downscaled scenarios
(d) Site-specific changes in E[T] and E[P] were estimated by averaging the five down-
scaled anomalies for each month Since these changes showed jagged annual cycles at all
locations, annual cycles were smoothed using a moving average of the downscaled
changes of the month itself and the preceding and subsequent months The climatic
scenano for St Gotthard was obtained by IDW interpolation of the downscaled monthly
changes from the same surrounding stations as were used to denve the present climate
68 Chapter 4
Since the downscaled changes showed no strong dependency on altitude, no correction
for elevation was applied At all sites the covariance matrices COV[T,P] were held
constant at present (baseline) values
4.2.3 Forest Model ForClim
From the modularized (FISCHLIN, 1991) forest model ForClim (version 2 4 04,
BUGMANN, 1994, FISCHLIN et al, 1995) we used three submodels (Fig 4 1)
FORCLIM-C (climate) simulates location-specific climatic changes as discrete events
FORCLIM-E (environment) simulates the abiotic environment of a patch from the climatic
parameters as produced by FORCLIM-C ForClim-P (plants) is a discrete time (1 year
time step) patch dynamics model simulating the fate of tree cohorts populating a patch of
1/12 ha in a species specific manner
FORCLIM-E contains the stochastic weather generator which simulates interannual
variability in monthly temperature and precipitation From the location-specific parame¬
ters SAi, FC, and X (Fig 4 1) FORCLIM-E then computes the winter minimum temper¬
ature TWinMin, the annual growing degree-day totals GDD, and the annual drought
stress index DSI for input to FORCLIM-P TWinMin is the minimum of the vanates of T
for December, January, and February GDD is computed by correcting for biases and
discretization errors (for a discussion cf FISCHLIN et al, 1995) according to the method
developed by BUGMANN (1994) Following a simple bucket-model approach, actual
(AET) and potential (PET) evapotranspiration are computed according to THORNTH¬
WAITE & MATHER (1957), where calculations depend not only on the usual site-specific
parameters field capacity FC (in this study at all sites 30 cm) and latitude X (site A
46 9°N, B, C 46 6°N, and D 46 2°N), but also on the slope-aspect index SA, The
latter was used to correct the PET for the slope and aspect dependent radiation plus wind
effects (cf BUGMANN, 1994) The drought stress index DSI is the difference between
annual PET and AET divided by the PET (PRENTICE & HELMISAARI, 1991, BUG¬
MANN, 1994, FISCHLIN et al, 1995)
FORCLIM-P simulates three basic processes, which determine annual changes in the tree
cohorts cunently existing within a patch (l) establishment of saplings, (n) death of indi¬
vidual trees [both (i) and (n) formulated as Poisson processes], and (in) tree growth, for¬
mulated determtntstically based on an improved growth equation (MOORE, 1989, BUG¬
MANN, 1994), which mimics explicitly inter- and intraspecific competition for light and
implicitly for water and nutrients Establishment and growth are directly affected by
weather However, since slow growth lasting over a series of years increases the proba¬
bility of tree deaths, weather may also contribute indirectly to mortality
Synthesis and Application Assessing Impacts on Forests 69
Full definitions of all equations forming FORCLIM can be found in BUGMANN (1994,
FORCLIM-P) and FISCHLIN et al (1995, FORCLIM-E)
Simulation results were obtained according to the Monte-Carlo technique by sampling for
each site 200 vanates (BUGMANN et al, 1996) Always assuming an empty patch as the
initial condition, each variate was obtained by sampling the state vector every 20th year
within the simulation period 800 till 3200 FORCLIM-C simulated climatic changes as a
step change in the year 2050 AH simulations were performed on Macintosh computers
(68040 CPU) and Sun workstations using the interactive RAMSES simulation environ¬
ment (Fischlin, 1991, Fischlin et al, 1994) and the RAMSES simulation server
RASS (THOENY et al, 1994, THOENY el al, 1995)
Steady state estimates under current and future climates were computed from the 200
vanates by averaging species abundances from biomasses over 400-year periods, under
the baseline climate (1840-2040), and under future equilibrium conditions assumed to be
reached after 2800 (cf BUGMANN etal, 1996)
To summarize effects of climatic change on species compositions similarity indices, S,,
were computed between pairs of steady state estimates (see above) according to the
formula S, = 1 - L lxi< - ykl / £ (xk + yk) where xk and yk are abundances of species k in
the current (xk) and the future (yk) climate, respectively
4.3 Results
4.3.1 Climatic Scenarios
At all four study sites (Table 4 1) the GCM-based future scenarios showed an increase in
annual mean temperature and a tendency towards wetter conditions compared to the pre¬
sent climate The direction and magnitude of the changes (Fig 4 2) differed from site to
site Annual mean temperatures increased by 2 4 °C at Sion and St Gotthard, by 2 5 °C
at Bern, and by 2 6 °C at Bever Precipitation showed no significant change at Sion
(+0 9 cm or +1%), whereas increases of 8 6 cm (+11%), 18 cm (+31%), and 42 cm
(+20%) were obtained at Bern, Bever and St Gotthard, respectively
The downscaling models provided plausible linkages between observed interannual vari¬
ability of the local weather variables and large-scale anomalies in (0 the strength of the
westerlies in winter, (n) the intensity of more meridional (i e north-south) flow patterns
in the transition seasons, and (in) the strength of large-scale subsidence in summer (see
70 Chapter 4
also GYALISTRAS et al, 1994, GYALISTRAS et al, 1998) The proportions of total
variances of the local variables explained in the fitting interval 1931-1980 averaged over
all locations and winter half-year (October-March) amounted to 58% for temperature and
36% for precipitation For the summer half-year (April-September), the mean variances
were 65% and 32% for temperature and precipitation, respectively
300
250
200
150
100
50
0
E
c
o
Q.U
/ St (Bottha d
BeverBern
-»
i > "><>
- —1 1 1—i—i—i— —i— —T—
Sion
i—i— i
-2 0 2 4 6 8 10 12 14
Temperature (°C)
Fig 4 2 Comparison of annual mean temperatures and annual precipitation totals under
the present () and the assumed scenario climates (0) at the case study sues (Table 4 1)
means for 1931 1980 computed from daily measurements (BANTLE 1989) 0 scenario
values denved by statistical downscaling (GYALISTRAS el al 1994 GYALISTRAS &
FISCHLIN 1996) from a 2xC02 experiment with the CCC GCMU (BOER et al 1992)
Comparison of the annual cycles of temperature and precipitation under present-day and
the simulated future climate shows (Fig 4 3) that the shifts in the annual means
(Fig 4 2) were not equally distributed throughout the year and that the seasonal patterns
vary from site to site Averaged over all locations warming was smallest in spring and
autumn (+2 2 °C) and largest in summer (+2 8 °C) and precipitation increased in all
seasons (+9% to +30%) except for summer (-5%) Among sites the warming differed
the least in autumn (0 3°C) and the most in summer (0 9°C) and the changes in precipi¬
tation differed again the least m autumn (20%) and the most in spring (44%)
These results conform to the underlying circulation changes simulated by the CCC-GCM
In winter the GCM-projected enhanced mean westerly flow over Europe would produce
stronger temperature changes at the more north-westerly sites (Bern and Sion) and a gen
Synthesis and Application Assessing Impacts on Forests 71
Sion
T (°a'£l
_v ^,*'.
S^s^ .'
St ,0
Gotthard 5
P(°m)^.
Jil\^-
r ^ 5-.-5F v*
/s
/ >y\/ \\
*
\V
/\
/
/r.
As
/y V
'f/ s.
' 4~\-'
v ^ r
\^..n_!SIi~ 5
-v-,^^ 12
32-,y
— *fc— ,0.
.7 SL 8- \
4.
^~/~^>£^^s*^;^ "^
JPMAMJJASOND J FMAMJ JASOND
Fig 4 3 Comparison of in period 1931 1980 observed (thick lines) annual cycles for
monthly mean temperature (left) and total precipitation (right) (BANTLE 1989) with the
climatic scenarios (broken lines) derived by statistical downscaling (GYALISTRAS etal
1994 GYALISTRAS & FISCHLIN 1996) from a 2xC02 experiment wilh the CCC
GCMII (BOER et al 1992) at the four case study sites (see also Table 4 1) Note the dif
terent scale for precipitation at Si Gotthard
eral precipitation increase The smaller warming at all locations during spring is associ¬
ated with a distinct strengthening of the northerly flow component simulated by the
GCM In summer, increased anticyclonic activity over central Europe in combination
with strong mid continental heating produced greater temperature increases at the more
easterly locations (St Gotthard, Bever) A strengthening of the north-easterly flow in
autumn was associated with a smaller warming and a slight increase in precipitation at all
sites, except for the more sheltered location at Sion
72 Chapter 4
4.3.2 Forest Responses
A general response to climatic change, probably typical for many mountain toresLs in the
Alps (BUGMANN, 1994, BUGMANN & FISCHLIN 1994), is that observed at the Bever
site (Fig 4 4) The pnmary succession starting in year 800 lasts about six centuries,
alter which European larch (Lartx dectdua MILL ), a pioneer species is substantially
replaced by the late successional species arolla pine (Pinui cembut L) The climax
community is the larch-arolla pine loicst ottcn lound in the Cditi al subalpine belt ot the
Alps (Table 4 1) At this site the simulated step change in climate (cl Figs 4 2, 4 3)
produced a sharp transient response in the forest It then initiated a secondary succession
lesembhng a pnmary one, which eventually lead to a completely new climax community
containing none of the previously dominant species Only larch and Scots pine (P
sylvestrn L ) irom the earlier climax community appear as early successional species
EZ3 Pinus silvebtrts L
(S3 Pinus cembra L
C3 Lanx decidua Mill
m Picea abies (L ) H Kahmen
E£2 Abies alba Mill
LT3 Quercus robur L
ES Quercus petraea (Mai ) Lieb
BB Acer pseudoplatanus L
OS Acer platanoides L
DH Fraxmus excelsior L
Uimus glabra Huds
I all other species
1 ig 4 4 Simulated species compositions at the case study sue Bever (B Table 4 1) in
the Swiss Alps lor current clunau. (800 2040) and a possible future 2xCO, clunilc (2060
3200) as downscaled (GYAL1S1RAS et al 1994) trom ihe CCC GCMII (BObR a al
1992) All simulations were made with the forest model FOKCl IM (BunMANN 1994
1 ISCIILIN a ul 199S) Accumulatively shown mem species abundances m l/ha were
averaged liom 200 vimtes sunpled iccordmg lo the Monte C irlo leehiiK|ue Irom the
stochastic process described by 1 ORC L1M (ci 1 ig 4 S. ind 4 6) £ ill olhei species is
total ol the lollowing lira abundance species Sciln alba L Suibu\ ana (L ) ( RANfZ
S auLuparta L Tilui cordala MILLLR T plai\ph\llo\ SCOV Qmriu\ puln\nn\WILLD Poptilus irimulaL Conlus a\ellunaL Bitttla puulula ROIH CarpinuibitulmL Alnu\ glutmow (I ) GAI R1NI R A incana (I ) MOI Nt H A \mdis
(CHAIX) DC Taxus bauala L Acer campeslre L and Pinus numlami MILLhR
Biomass (t/ha) Bever
2800 32i
Synlhesis and Application: Assessing Impacts on Forests 73
Current Climate EquilibriumESI P«ius silvestris L
KS) Pinus cembra L
LZD Lan* decickia Mill
@23 Picea aO»s (L) H Karsten
Hffl ^b;esa*a Mm
H Farjus stlvattca L
EZ3 Quercus robur L
R^l Quercus pelraea (Matuscmka) Lies
ES3 Castanea saliva Mill
ra Acer pseudoplatanus L
^ flcerpfafanodesL
M Populus nigra L
nrni Fraxmus cxcolsior L.
rzn Ulmus glabra Huos.
E3 1 all olher species
Bern Bever St.Gotthard Sion
Fig. 4.5: Simulated equilibrium species compositions al the case study sites (sec also
Table 4.1, Fig. 4.4, and Pig. 4.7 bottom) in the Alps lor currcnl baseline climalc (800-
2040). All simulations made with FORCLIM (BUOMANN, 1994; FISCHLIN et al, 1995)
Depicted arc mean species abundances averaged Irom 200 variaies sampled according to the
Monte-Carlo-tcchmquc Irom the stochastic process described by FORCLIM (cl. Fig. 4.6).All olher sellings as described under Fig. 4.4.
EE3 Pinus siivestns L
ESS P/nus CBmbra L
C3 Lanx deodua Mill
K52 P/cea afiies (L ) H Karsten
S3] Abies alba Mh i
wm Faqvs silvalica L
EH Ooercus roOurL
ESS Quercus pelraea (MatuschkA) Lier
E23 Casianea satwa Mill
raa .Acer pseudoplatanus L
IEH -4cef p/alancwdesL
H Populus nigra L
rrrm fTaxinus eJtce/.wirL
CZI [JJm(i$ oVabra Huos
E3 I all other species
Bern Bever St.Gotthard Sion
Fig. 4.6: Simulated equilibrium species compositions at the case study silcs (see also
Table 4.1, Fig. 4.4, and Fig. 4.7 bottom) in the Alps lor a possible, future 2x00,climate (2060-3200) as downscaled (GYAIJSIRAS el al., 1994) Irom Ihe CCC-GCMII
(BOER el al, 1992). All other settings as described under Figs. 4.4 and 4.5.
74 Chapter 4
The steady-state biomass densities for all four sites (Table 4 1) are depicted for current
climate in Fig 4 5 and for the projected future equilibrium climates (Figs 4 2 4 3), in
Fig 4 6 At Bern the climatic change caused no significant changes in the climax com¬
munities (S[ = 0 92) Contrastingly, the steady states at Bever differ sharply (S, = 0 07)
A similar dramatic change (S, = 0 002) was obtained at St Gotthard where a forest
appears above the current timberhne (cf Fig 4 7 bottom) At Sion we obtained a quite
different response, i e the collapse of the forest canopy Only very low tree biomass
remains in the climax community under the changed climate (S,= 0 07)
The flexibility of the new technique (CLIMSHELL combined with FORCLIM, cf
Fig 4 1) is demonstrated in Fig 4 7 The technique makes it possible to simulate the
behaviour of forests at locations different from those at which weather records are
available, for instance along an altitudinal gradient This is possible not only for current,
but also for downscaled, future projected climatic scenarios (Fig 4 7 top, middle,
bottom, steady-states from bottom simulation also shown in Figs 4 5 and 4 6)
4.4 Discussion
For a far-future "2xCC>2" equilibrium world, the CCC-GCMII projects a warming in the
mean annual global near-surface air temperature of 3 5 CC (BOER et al, 1992), approxi¬
mately at the centre of the 2 1 - 4 6 °C range projected by state-of-the-art GCMs (KAT-
TENBERG et al, 1996) Transient experiments with coupled GCMs (with climatic sensi¬
tivities in the range noted above) under steadily increasing concentrations of greenhouse-
gases showed a warming of 1 3 - 3 8 °C at the time of doubled CO2, with an average of
only 2 0 °C Such a delay was not considered in our simple, step-like scenarios, but, as
has been shown by BUGMANN (1994), it is probably of minor relevance for the mid and
long-term response of forests
Several reasons indicate that the regional scenarios of climatic change used here are
plausible and spatially consistent (1) The scenarios plausibly reflect the effects of
changes in the spatial distributions of large-scale monthly mean sea-level pressure and
near-surface temperature as projected by the CCC-GCMII (analysis of CCA model
response) (11) The projected warming is to be expected, since it seems unlikely that any
changes in regional climate forcings (which could lead to e g increased frequency of
thermal inversions or of cold air drainage) could override the strong, general warming
prescribed by the CCC-GCM (111) The projected increases in precipitation are consistent
with the general increases in globally averaged precipitation of ca 2%-15% projected byGCMs under a C02-doublmg (IPCC, 1990, 1992), as well as with indications from an
experiment with a regional climate model (SCHAR et al, 1996) (iv) The average ratio
Synthesis and Application Assessing Impacts on Forests 75
St. Gotthard
2090 m a.s.l.
500
400
300
200
100
BS33 Pinus cembra L
i i Lanx decidua Mill
Picea abies (L ) H Kahs>ten
Abies alba Mill
I....1 X allother species
2300 m a.s.l
o800 3200
Tig 4 7 Using Ihe LI IMMILLL (GYALISTRAS & rise HLIN 1996) lo explore forest
dynamics where there are no measurements aviilable (lop 2100 middle 2200 boitom
2300 in a s 1) The applied technique allows to inlerpolaie baseline clim lies md climatic
scenarios by combining measurements Irom long temr base stations (this study station
of St Gotthard at 2090 m asl) with measuiements from auxiliary chmalologicalstations (this study up to additional 39 stations) a digital terrain model and regionalizedGCM-oulput (for bottom cf Fig 4 5 and 4 6) All olher settings as described under
Fig 4 4 and 4 5
76 Chapter 4
between projected changes in annual temperature and precipitation (5 4%/°C) is consistent
with the theoretical value (e g , ROEDEL, 1992, p 65) obtained under the assumption that
the relative humidity of the atmosphere would remain unaltered (e g ,MITCHELL & IN¬
GRAM, 1992) (v) The variability in the regionalized patterns of change appears realistic,
in view of the seasonally and spatially strongly varying link between Alpine and large-
scale climate (e g ,SCHUEPP, 1968, FLlRI, 1974, GENSLER & SCHUEPP, 1991, GYA¬
LISTRAS et al, 1994, SCHAR et al, 1998). (vi) The projected changes were spatially
more uniform for temperature than for precipitation, in agreement with patterns observed
in the Alpine region (FLIRI, 1974, EHRENDORFER, 1987, AUER & BOHM, 1994,
BENISTON et al, 1994) (vu) The scenarios are consistent with climatic changes obtained
at a coarser spatial scale from physically-based regional climate models (for a detailed
comparison of model-based and empincally-downscaled scenanos for the Alpine region,
see GYALISTRAS et al, 1998)
In present climates FORCLIM normally produces realistic species compositions as is
shown not only by our own results (cf Table 4 1 with Figs 4 4 to 4 7), but also by
other studies, which have analysed and validated the behaviour of this model in Europe,
North America, present, and past (e g ,BUGMANN, 1994, BUGMANN & FISCHLIN,
1994, BUGMANN & SOLOMON, 1995, FISCHLIN, 1995, FISCHLIN et al, 1998a)
From all this evidence we claim that the model ForClim has considerable predictive
power and is generally capable of projecting "realistic" results
Similarly, sensitivity analyses at the four study sites (not shown), agree with results from
former studies (e g ,BUGMANN, 1994) and suggest that the projected new equilibria for
the species compositions are relatively robust, i e that small changes in the climatic
scenarios of the order of ±0 5 °C or ±5% for precipitation would show only little effect
(Fischlin et al, 1995)
Although all study sites were affected by the same global pattern of climatic change, the
projected forest responses differed significantly, ranging from enhanced growing con¬
ditions, to no effect, to complete disappearance of the forest (cf Figs 4 5 and 4 6) All
simulations assumed an identical spectes pool disregarding any phytogeographic history
Thus, observed differences among the sites can be explained, at least partly, by the site
specific climatic parameters for monthly temperature and precipitation Since the sce¬
narios of local climatic change contained also a strong shared component, however, the
dramatic differences found in the forest responses cannot be explained only by the phys¬
ical environment To fully understand the obtained patterns, we have to consider also the
intrinsic mechanics of the ecosystem model, which form a complex network of non¬
linear, interdependent causes behind the projected forest responses
Synthesis and Application Assessing Impacts on Forests 77
4.4.1 Forest Responses
A) Bern
Bern shows a remarkable insensitivity to the projected changes It is the only site where
the similarity index is close to one, whereas all other sites show very small, near zero
indices (see "Results'), indicating sharp changes in species compositions under current
vs new equilibrium conditions Although other studies (e g ,NlLSSON & PITT, 1991)
have projected similarly small changes for this zone (Swiss Plateau) our result contrasts
with that obtained by BRZEZIECKI et al (1995), who have applied a statistical, steady-
state model based on correlations between zonal forest communities, temperature, and
edaphic factors (BRZEZIECKI et al, 1993, BRZEZIECKI et al, 1995) Their model uses
temperature as the only climatic factor determining plant communities Although temper¬
ature may serve in some instances as a useful indicator for the water balance regime, this
can not be expected to be the case in general (HOLDRIDGE, 1947, 1967, BOX, 1981,
Woodward, 1987, Box & Meentemeyer, 1991), which may explain some of the
differences between their and our results
Moreover, the Bern site is currently characterized by few periods of drought stress,
whereas in the southern Alpine areas (which BRZEZIECKI et al have used as an analogue
for the future conditions on the Swiss Plateau) precipitation and, hence, drought stress,
show greater high intra- and interseasonal variability For example, Bern receives under
present climate in summer (winter) a mean total precipitation of 35 cm (18 cm) distnbuted
over 35 (28) days, whereas at the southern Alpine location of Lugano a far larger amount
of 55 cm (23 cm) falls within only 32 (20) days
A further explanation for this difference lies in our use of seasonal climatic scenarios,
which projected the strongest warming at Bern during winter, which has only a minor
effect on the vegetation Moreover, BRZEZIECKI et al have assumed no changes in pre¬
cipitation, whereas the downscaled scenarios showed increased precipitation during
winter, spring, and fall, partly compensating for the projected increases in temperature
Unless the insubrian climate typical of the southern Alps should really prevail on the
Swiss Plateau, the projections by FORCLIM appear to be more plausible for the given
climatic scenario The small sensitivity at Bern to warming is further corroborated by the
temperature requirements of the presently dominant tree species, since the projectedclimatic changes do not push any of these species far from their ecophysiological optima,
allowing them to remain within the centre of their realized niches However, the statis¬
tical model by BRZEZIECKI et al might again become superior, in cases of extreme
warming scenarios, where patch models like FORCLIM -eg, due to the use of simph-
78 Chapter 4
fied schemes to compute the local water balance (THORNTHWAITE & MATHER, 1957,
FISCHLIN et al, 1995) -tend to underestimate the sensitivity of forest responses
(KIENAST, 1991, BUGMANN, 1994, BUGMANN & FISCHLIN, 1994)
B) BEVER
At the subalpme site Bever, the projected climatic changes cause major changes in species
composition (Fig 4 4), also expressed in the extremely low similarity index The new
forest composition simulated under a changed climate represents a completely new com¬
munity with no present analogue (ELLENBERG & KLOTZLI, 1972, ELLENBERG, 1988)
This result conforms with the individual species responses postulated by several authors
to explain the patterns found during past climatic changes (e g , Davis, 1981, HUNTLEY
&Birks, 1983, Davis, 1990, Huntley, 1990)
The climatic changes projected for Bever vary considerably with season For instance,
summer warms by -0 7 °C more than winter, while precipitation in summer remains
unchanged, but increases 12 to 55% in other seasons
Not surprisingly, the annually averaged changes in climate at Bever (Fig 4 3) distribut¬
ed equally over all months cause a forest response (results not shown here) which differs
significantly in its transient as well as equilibrium species compositions (S, = 0 62)
Consistent with increased water availability and less warming during summer, Acer
pseudoplatanus was found to be less abundant in the non-seasonal scenario than under
our original, monthly resolved scenario (Fig 4 4), instead, Picea abies produces ca
30% of the total biomass The high abundance of P abies moves this new forest
composition closer to that found under present conditions Hence, we conclude that
using annually averaged scenarios may lead to an overestimation of such a forest's ability
to adapt to potential climatic change
Bever is located in the centre of the Alps and the omnipresence of species, e g ,of sweet
chestnut (C sativa) or European beech (Fagus stlvatica), as assumed in the model, can¬
not be expected to be the case in reality Therefore, the transition to a new equilibrium, is
likely to be further hampered by the migrational barriers posed by mountains This infer¬
ence is corroborated by phytogeographic evidence from the past, which shows that the
Alps have functioned as an insurmountable barrier for slow-migrating species in the pre¬
sent-day climate (HUNTLEY & BlRKS, 1983, HUNTLEY, 1990, OZENDA & BOREL,
1990)
Synthesis and Application Assessing Impacts on Forests 79
C) St Gotthard
At St Gotthard tree growth is completely suppressed under the current climate, whereas
in the projected new climate trees can form a forest canopy (cf Figs 4 5, 4 6, and 4 7)
This site is located just above the current tree line (cf Fig 4 7 top and middle vs 4 7
bottom) where forests are believed to be mainly limited by temperature, since precipi¬
tation is abundant (WOODWARD, 1987, ELLENBERG, 1988)
The forest responses projected by FORCLIM are consistent with ecological expectations,
which require particular temperature regimes for tree growth, l e, a growing season of at
least 30 days with daily mean temperature above 10°C and fewer than 8 months with
mean daily minima below 0°C (WALTER & BRECKLE, 1986, 1991) At St Gotthard the
present climate is characterised by no months with a daily mean temperature above 10°C,
but in the projected new climate (cf Fig 4 3) there are more than 30 days exceeding this
threshold Furthermore, the period with daily minima below freezing decreases from
almost 8 months under the present climate to less than 7 months under the projected,
future climate
Hence, the forest establishment projected by the model appears plausible, although some
details of the initial succession may not be realistic Firstly, existing soils may not be
fully colonizable (RENNER, 1982) requiring a long phase of development Secondly,
growth at the tree line may be reduced because of adverse local conditions such as dis¬
eases, strong winds, browsing by cattle and game, or physical instability on steep slopes
Thirdly, the models ignore the form of the precipitation (tree growth might also be ham¬
pered by mechanical or drought stress due to snow and ice), and changes in incoming
solar radiation and in the duration of snow cover For instance, increases in temperature
and/or its variance during spring may produce warm periods which can melt the snow
cover earlier than at present Since occasional invasions of cold air into central Europe
during spring can be expected to occur also in the future, an early retreat of the protective
snow cover may increase the exposure of plants to frost damage The strengthening of
the northerly flow component over Europe as projected by the CCC-GCMII during
spring (mainly April) even suggests the possibility for an increase of this risk under a
"2xC02"-climate
Even though high latitude tree lines can not be compared directly with high altitude Alpine
tree lines (e g ,because of the different radiation regime), some recent observations from
boreal tree lines (TaUBES, 1995) appear to suggest similar effects of at least transiently
reduced growth under warming conditions
80 Chapter 4
D) SION
The Sion site presents an opposite case from that of St Gotthard the current climate al¬
lows for tree growth, whereas in the projected new climate, trees can no longer form a
forest canopy (cf Figs 4 5, 4 6)
The disappearance of the forest vegetation appears to be plausible, considering the pre¬
sent-day low mean annual precipitation of 600 mm (cf Table 4 1), which places this
xenc site close to the drought tree line (HOLDRIDGE, 1967, Walter & Breckle,
1983, 1984, 1986, 1991, WOODWARD, 1987) Using the FORECE model (KIENAST,
1987, KIENAST et al, 1987, KIENAST & KUHN, 1989) KIENAST (1991) has obtained
similar results for this site
An analysis of the growth factor that quantifies the effect of soil moisture deficits in
FORCLIM (BUGMANN, 1994, FISCHLIN et al, 1995) revealed the simulated degradation
of growth conditions were due to enhanced drought stress In the projected climate,
growth was reduced to a range between 4% and 10% (upper and lower 95% confidence
limits) of the maximum growth potential This increased drought stress was caused not
only by the general rise in temperature, but also by the simultaneous 28% decrease in
summer precipitation The downscaled changes for summer precipitation are particularly
uncertain (e g ,GYALISTRAS et al, 1994), but it is worth-noticing that other experiments
with FORCLIM (not shown here) - where we assumed only warming and no changes in
annual precipitation - still lead to a collapse of the forest canopy
4.4.2 Sensitivities and Uncertainties
The sensitivity of Alpine forests to human interference with the climate system directly
depends on the underlying sensitivities of the global and regional climate We note that in
this study we used a global climate model with an intermediate sensitivity to increased
concentrations of greenhouse gases, and that the downscaling procedure adopted may
over- or underestimate the response of the local climates to the simulated changes in
global climate (GYALISTRAS et al, 1998) The forest responses obtained then revealed a
potential for greatly differing sensitivities to the same global scenario of climatic change,
a result which must be interpreted carefully Despite many recent efforts to improve
ForClim (Bugmann & Fischlin, 1992, 1994, Bugmann, 1994, Fischlin et al,
1995, BUGMANN et al, 1996), this model may not always adequately reflect the sensi¬
tivities of the real forests
Synthesis and Application Assessing Impacts on Forests 81
For instance, the disappearance of the forest at Sion may overestimate the true sensitivity,
since this result depends, amongst others, on the particular and limited pool of species
used in the simulations Olive trees (Olea europaea L ), oaks such as the cork oak (Q
suber L ), holly oak (Q ilex L ), or Pyrenean oak (Q pyrenaica Willd ) could probably
produce higher tree biomasses than those simulated by ForClim Similarly, if we were
to accept the oak species now included in FORCLIM, i e Q robur L, Q petrea, and Q
pubescens, could possess the characteristics of Mediterranean provenances of these
species, we would obtain a similar effect (FISCHLIN et al, 1998b) Without human
assistance, however, it is unlikely that such provenances could quickly migrate to Sion,
leaving the site vulnerable to a transitory, yet long-lasting forest disappearance
Moreover, ForClim may have overestimated the sensitivity at all sites, since, the greater
the genetic plasticity of today s tree species in terms of tolerance to abiotic forcings such
as drought, the lower is their expected sensitivity to a given climatic change Such a
buffering effect, now completely ignored by FORCLIM, might be further enhanced if
climatic change were to trigger novel selection and thus alter the gene pools of the tree
species by affecting gene frequencies Considering such effects would require additional
simulations which include new species or to adjust the corresponding species parameters
within ForClim
Conversely ForClim may overestimate the robustness of forest compositions at the Bern
site, since various simulations under even stronger warming (not shown here) still
showed little effect Nevertheless, as already discussed, for the ranges of climatic change
used in this study, the low sensitivity obtained at Bern is likely to be generally realistic
FORCLIM may also underestimate the sensitivity close to the ttmberline limited by temper¬
ature This is because it ignores possible additional effects resulting from decreases in
the duration of the snow cover, from increases in photosynthetic efficiency, or from
improved nutrient allocation (cf eg KORNER, 1995)
The projections represent no forecasts they merely describe the forest succession if all
prescribed assumptions hold or become actually true Consequently, many unavoidable
and in some instances eventually irreducible uncertainties accompany the projections
For instance the arbitrary selection of particular reference points in a large space of abiotic
and other environmental parameters like the selection of the 2xCC"2 radiative forcing
scenario, the CCC-GCMII, the 1931-1980 baseline climate, the species parameters, etc
all generate uncertainties
Thanks to the modular structure of our method it is easy to explore uncertainties, because
modules can simply be exchanged for others e g ,the CLIMSHELL can be switched to
any other GCM* or FORCLIM can be easily simulated with any combination of sub-
82 Chapter 4
models (FISCHLIN, 1991, FISCHLIN et al, 1994) In some cases this allows quantifi
cation of uncertainties For instance based on a random resampling procedure (EFRON,
1979) we obtained ca +0 5 CC for temperature and ±20% for precipitation relative to the
"best estimate scenario values obtained from the 2xC02 experiment with the CCC-
GCMII (averaged over 22 Alpine locations and all months of the year by determining
95% percentiles, for details see GYALISTRAS & FISCHLIN, 1995)
4.5 Conclusions
This study demonstrated the feasibility and viability of a new method to assess possible
transient responses and equilibrium states of forests in a topographically complex region
such as the Alps In combining results from a GCM-expenment with a downscaling
technique and a forest patch model (Fig 4 1) we obtained regionally differentiated, inte¬
grated, reproducible, quantitatively consistent assessments of climatic change and their
associated impacts on mountain forests at a relatively high temporal and spatial resolu
tion The method is spatially flexible and general and has the potential to be applied di¬
rectly without modification in the mid and high latitudes of the Northern Hemisphere
(BUGMANN & SOLOMON, 1995) at every location of ecological interest Finally, it con¬
forms to the IPCC guidelines for impact assessments (CARTER et al, 1994) and is
modular and efficient The latter is important for exploring sensitivities and uncertainties
and provides a basic framework for replacing modules with better vanants or enhancing
the method as new procedures or submodels become available
The climatic scenarios obtained and the associated forest responses appeared reasonable
in light of current knowledge of the climate system and the local ecology It mattered for
the forest responses that we used climatic scenarios resolving the annual cycle, indicating
that annually averaged scenarios may yield misleading results We obtained highly
variable forest sensitivities within close vicinity Insensitivity was found for low altitude
deciduous forests, whereas high sensitivities were found at the water- and temperature
tree lines These results indicate that mountain forests will be highly sensitive to climatic
change At present little is known about the proportions of sensitive vs insensitive
forests in a region such as the Alps Consequently, the relative importances of critical
forest responses (e g , temporary die-back during phases of major changes in species
composition or even complete disappearance) versus beneficial effects (e g ,establish¬
ment above the current timber-line) remain to be studied further If climatic change
should occur in the patterns suggested here, it is likely to result in severe local damage,
even if the total area affected is relatively small Once more the results of this study
indicate that tree populations respond individually and not as a community, and that
Synthesis and Application Assessing Impacts on Forests 83
without human assistance new steady states are reached only after several centuries or
millennia, especially in the presence of migrational barriers
The results ought not be confounded with actual forecasts they represent mere
projections which give possible answers to "what if scenarios The validity some of
these projections may hold, only the future can reveal We believe that they represent
valid first-order estimates of possible impacts of climatic change on forests in the Alps if
greenhouse gas forcing doubles
84 Chapter 5
5. Discussion
The demanding input requirements of ecosystem models present a particular challenge for
the construction of climatic scenanos The present work demonstrated that by combimng
global and local measurements with simulation results by GCMs it is possible to denve
appropriate and physically plausible (sections 2 3 3, 4 4) scenanos at comparatively
small expenses The use of a downscaling procedure in particular allowed hnkage of
local ecosystem responses to in principle any given scenano for the future radiative fore
ing of the global climate system (Fig 4 1) Thus, in contrast to simpler approaches,
such as ad-hoc adjustments and analogue techniques, the proposed method provides a
mean to explore precisely the propagation of uncertainties (amplification vs damping)
through climatic scenanos and ecosystem models
A commonly used approach in impact studies is to interpolate the climatic scenanos from
a few climate model gndpoints in the vicinity of the region of interest The pronounced
spatial gradients found in the present study (Figs 2 12, 2 13, 4 3), however, were in
strong contrast to the smooth patterns of change (e g ,BACH et al, 1985, OZENDA &
BOREL, 1990) that result from this chmatologically inconsistent (cf Fig 2 1, VON
STORCH, 1995) procedure In particular, as shown by GYALISTRAS & Fischlin
(1995), application of gndpoint-interolated scenanos would have yielded in some cases
misleading forest responses - at least m a complex terrain such as the Alps
Different from regional climate models the proposed method provided a very high (local)
resolution This resolution was obtained based on a conceptually and technically relative¬
ly simple procedure (Eqs 2 1-2 4 and 3 1-3 4), and at medium data requirements
Hence, it appears feasible that with the aid of two newly develloped computer programs,
ClimShell and MonWeathGen, a graduate student could construct a first iteration of
scenanos already within a few days Quite differently, simulation expenments with a
regional climate model would have required to develop considerable expertise, and would
have depended on the availability of large data sets for model initialization, testing, and
application All this would have required a considerably larger amount of time
The here proposed procedure was also found to be computationally efficient For
example, downscaling of hundred years of global climate model output to a particular
location (Fig 2 11) took only ~1 min on a modern workstation The stochastic simu¬
lation of 200 x 1'400 years of monthly weather data (section 4 2 3, Fig 4 4) required
only an additional -10 min per location In contrast, simulations with e g the NCAR
Discussion 85
second generation regional climate model would have required at a resolution of 60 km
per simulated year 60 hours of CPU time on a Cray Y-MP (GlORGl, 1996)
A further advantage of the proposed procedure was that it matched the limited precision
of GCMs This was achieved by using a local weather generator instead of GCM-
simulated time series (Fig 4 1), by removing GCM biases prior to downscaling, and by
using but the first few EOFs of monthly to seasonaly averaged fields (Fig 2 3) Quite
differently, in regional climate simulations the errors of the driving model are directly
propagated into, and in some instances even amplified by the high-resolution model For
example, based on similar boundary conditions to the ones which we used to construct
the "ECHAM 2xC02' scenario in section 2 3 3, ROTACH et al (1997) obtained in a
simulation experiment with a high resolution regional climate model over the Alps a sum¬
mer warming of 4-6 °C This result was probably unrealistic since it was mainly caused
by shortcomings in the simulation of the jet stream by the driving GCM in combination
with an inappropriate representation of the regional soil moisture (BENISTON et al, 1995,
MARINUCCI et al, 1995) Our downscaling procedure yielded a smaller, much more
plausible warming in the order of 1-3 °C (Fig 2 12, Table 2 3) This, however, does
not defy the use of regional models in principle, it only demonstrates that it is easier to
obtain plausible scenarios with our method Thus, our result may have again to be
revised as improved high-resolution simulations under similar boundary conditions
become available
All scenarios were based on the critical assumption that the statistical models used for
downscaling and stochastic weather generation would also hold under a changed climate
The validity of this assumption can not be strictly proven Generally, the errors of the
downscaling procedure can be expected to increase with increasing strength of climatic
change (cf Fig 2 11) One may argue that for the near future, l e the next few decades,
the use of first-order terms (Eq 2 3) could be sufficient This view is to some extent
supported by analyses of GCM-expenments which suggest that future climate might not
show basically new circulation types, but only a redistnbution of situations already found
in this century s weather record (e g ,ZORITA et al, 1995) Moreover, a recent com¬
parison of statistically downscaled changes in Romanian precipitation with results from
dynamic simulations (BUSUIOC & VON STORCH, 1997) suggested that a simple linear
model like the one used in the present study can correctly capture regional first-order
effects of changes in the large-scale circulation Further studies should investigate in as
far this could apply also to the climatically much more complex Alpine region
For several variables, such as summertime precipitation, performance of downscaling
was relatively poor (Fig 2 8) Possibly, better results could have been obtained by in¬
cluding additional large-scale predictors (such as upper-level and moisture fields),
optimizing the size of the large-scale sector used for downscaling, and using a higher
86 Chapter 5
(e g , daily) temporal resolution However, regional climatic changes also depend on
systematic changes in smaller-scale climate forcings (e g ,land use), or feedbacks to the
regional climate system (e g ,due to changes in vegetation cover) Such effects were not
considered by the here proposed method Nevertheless, indications from simulation
studies addressing these issues (e g ,GlORGl, 1996) could be included at any time in our
method by adjusting the corresponding parameters used for weather generation (cf
sensitivity analyses in section 4 4 1, sites Bever and Sion)
The used global models were far from perfect and thus also imposed their limitations
upon the denved scenarios The monthly to seasonaly averaged fields used for down-
scaling were nevertheless reasonably well reproduced by both used models for "lxC02"
conditions (VON STORCH et al, 1993, ZORITA et al, 1995, McFARLANE et al, 1992,
and own analyses) However, the "2xC02" responses of the two models differed con¬
siderably for example, the change in global mean near-surface temperature in the
ECHAM1/LSG model was only 1 7 °C, whereas in the CCC-GCMII it was 3 5 °C
These uncertainties in the simulation of global climate were found to affect mainly the
magnitudes and seasonal patterns of the downscaled scenarios, whereas the spatial
patterns of change were quite similar (sections 2 3 2 and 4 3 1) The stability of this
latter result should be studied further based on the use of additional global simulations
and more sophisticated downscaling procedures
All derived scenarios were based upon specific assumptions on the future radiative forc¬
ing of the global climate For instance, more recent projections of global climatic change,which included the cooling effect of aerosols, showed a much reduced summer warming
over Europe (KATTENBERG et al, 1996) This can be expected to have a strong effect on
forest responses The proposed procedure is certainly flexible enough (Fig 4 1) to
investigate such alternative scenarios as soon as they become available
It could be argued that the suitability of the proposed method was tested but in the context
of one case study However, this completely independent case study pursued its own
research goals (Chapter 4) and was therefore considered to represent a realistic research
situation Furthermore, the input requirements of the ForClim model (monthly mean
temperature and precipitation at a resolution of-100 m, and for at least several centuries)
were representative for an entire family of models, including JABOWA (BOTK1N &
NISBET, 1992), FORSKA (PRENTICE et al, 1993), or PnET-II (ABER et al, 1995)
Finally, the results of the presented analyses and the general conclusions depended only
to a small extent on the case study
More detailed dynamic models than FORCLIM would have required additional weather
variables, and/or a daily temporal resolution (section 2 1) The downscaling procedure
was however shown to be applicable to in principle arbitrary combinations of weather
Discussion 87
variables (Chapter 2) For some of these variables downscaling involved large uncer¬
tainties (Fig 2 7), but these were at least quantifiable (Fig 2 11, Table 2 3, see also
GYALISTRAS & FISCHLIN, 1995) In order to provide scenarios at a high temporal
resolution, the monthly weather generator could have been easily extended by a second
weather generator which simulates daily or even hourly weather based on monthly input
data (GYALISTRAS et al, 1997)
Static models of climate-vegetation relationships (e g, Stephenson, 1990, BRZEZIECKI
et al, 1995) would have had generally less demanding requirements in terms of temporal
resolution, but they would have required spatially extended scenarios These could have
been derived by applying the spatial interpolation procedure (Figs 4 1,4 7, GYALIS¬
TRAS & FISCHLIN, 1996) on a regular grid The feasibility of this approach was demon¬
strated in GYALISTRAS & FISCHLIN (1995)
Hence the proposed method appears to be generally applicable to a wider range of eco¬
system studies In particular, its application to modelling studies dealing with possible
impacts of climatic change on grasslands (RlEDO et al, 1997), snow cover (ABEGG,
1996), and run-off (STADLER et al, 1997) suggest that it can be useful also in other
sectors of climate impact research
88 Chapter 6
6. Conclusions
The proposed method to derive climatic scenarios enables consistent assessments of pos¬
sible impacts of global climatic change on local ecosystems. It bridges the gap between
the spatial scales at which global climate models and local ecosystem models operate, and
yet allows to derive transient scenarios at a high temporal resolution and over arbitrary
time spans. The method is general, flexible, and efficient, and conforms to the IPCC
guidelines for impact assessments (CARTER etal, 1994). Thanks to its modular structure
improvements of the individual components can be easily incorporated at a later stage.
Downscaling
Use of a statistical downscaling procedure enhances the consistency of local climatic sce¬
narios. For all Alpine locations considered and for all seasons physically plausible statis¬
tical downscaling models were found. The model's reliability varied however with the
time of the year, the variable and the region considered, and was partly low for some
ecologically important variables (such as growing season precipitation). The downscaled
scenarios depended plausibly on the large-scale climatic changes simulated by two global
climate models, they appeared realistic in view of current understanding of the involved
physics, and were consistent between locations. Compared to the use of regional climate
models statistical downscaling is easy to test, computationally efficient, and provides
sufficient spatial detail far above the resolution of these models.
Stochastic Weather Generation
The newly proposed stochastic weather generator improves the simulation of monthly
weather variables and relies upon more robust procedures to estimate the needed climatic
parameters. Inclusion of autocorrelation mattered for a statistically more accurate simula¬
tion of winter minimum temperatures and growing degree-day totals, whereas use of a
skewness reducing procedure for precipitation improved the simulation of a commonly
used index for annual drought stress. Compared to the direct use of output from climate
models the local weather generator circumvents the presently limited temporal precision
of these models, is more flexible, and provides a concise description of local climates.
Preliminary investigations indicated sensitivity of simulated forest compositions to
changes in higher-order statistics of climate. The weather generator provided a good
starting point for extensive sensitivity studies.
Conclusions 89
Climatic Scenarios
The application of downscaling to three model-generated scenanos of global climatic
change yielded significant, complex, and spatially as well as seasonally highly vanable
climatic changes in the Alpine region The scenanos depended on the choice of large-scale predictors, were dominated by changes in the the large-scale temperature field, and
reflected the global climate sensitivities of the dnvmg climate models All scenanos
depicted a general warming, which was larger at the northern than at the southern slopeof the Swiss Alps A tendency towards increased precipitation was found However,
decreases were found for some locations and seasons, with partially dramatic effects on
simulated forest compositions The consistency of the climatic scenanos should be
assessed further based on comparisons with alternative statistical downscaling proceduresand simulations with regional climate models
Forest Responses
The downscaled scenarios and the weather generator were combined with a forest
succession model to assess possible impacts of climatic change on forests in the Alps It
was shown that under one and the same 2xC02 scenano of global climatic change within
short distances sharply contrasting forest responses may occur minor changes in tree
species composition, major changes with temporary die-back, complete disappearance of
the forest, or the establishment of a new forest The forest responses were consistent
with current understanding of forest dynamics In some cases they depended on the use
of seasonally resolved scenanos It was once more found that completely new species
compositions may emerge under a changing climate, and that human assistance may be
needed to help some forests to adapt
The above results appeared reasonable in tight of current knowledge of the climate system
and the ecology of forests However, they are no forecasts They represent only projec¬
tions of possible future developments which are likely to occur if the assumed global
scenanos plus a senes of further assumptions (Fig 4 1, sections 2 4 and 4 4 2) would
hold or actually come true Nevertheless, our findings clearly demonstrate the potentialfor drastic changes in Alpine climate and forests The here proposed methods could
contribute to the development of corresponding mitigation and adaptation strategies
90 References
7. References
ABEGG.B (1997) Khmaanderung und Tounsmus vdf, Hochschulverlag AG an der ETH Zurich, 222 pp
ABER, J D, OLLINGER, S V
,FEDERER, C A , REICH, P B
, GOULDEN, M L, KICKL1GHTER, D W ,
MELILLO, J M &LATHROP, RG (1995) Predicting the effects of climate change on water
yield and forest production in the northeastern United States Clim Res,5 207-222
AGUIAR, R J & COLLARES-PEREIRA, M (1992) TAG a time dependent, autoregressive, Gaussian
model for generating synthetic hourly radiation Sol Energ ,49 167-174
ANDERSON, C W (1984) An introduction to multivariate statistical analysis 2nd ed, Wiley & Sons,
New York a o, 675 pp
AUER, I & BOHM, R (1994) Combined temperature-precipitation variations in Austria during the
instrumental period Theor Appl Climatol, 49 161-174
BACH, W , JUNG, H J & KNOTTENBERG, H (1985) Modeling the influence of carbon dioxide on Ihe
global and regional climate methodology and results Munstersche geographische Arbeiten, 21,
Schoeningh, Paderborn, 114 pp
BANTLE, H (1989) Programmdokumentation Klima Datenbank am RZ ETH Zurich Schweiz
Meteorol Anstalt, Zurich, Switzerland, 8 pp
BARDOSSY.A &PLATE.EJ (1992) Space-time model for dally rainfall using atmospheric circulation
patterns Water Resour Res,28 1247-1259
BARNETT, T & PREISENDORFER, R (1987) Origins and levels of monthly and seasonal forecast skills
for the United States surface air temperatures determined by canonical correlation analysis Mon
Weather Rev,115 1825-1850
BARROW, E, HULME, M & SEMENOV, M (1996) Effect of using different methods in the
construction of climate change scenarios examples from Europe Clim Res ,7 195-211
BEN1STON, M , REBETEZ, M , GlORGl, F & MARINUCCI, M R (1994) An analysis of regionalclimate change in Switzerland Theor Appl Climatol, 49 135 159
BENISTON, M , OHMURA, A ,ROTACH, M
, TSCHUCK, P, WILD, M & MARINUCCI, R M (1995)
Simulation of climate trends over the alpine region - Development ofa physically based modelingsystemfor application to regional studies ofcurrent andfuture climate Internal Report, Depart¬ment of Geography, Swiss Federal Institute of Technology (ETH), Zurich, Switzerland, 198 pp
BO, Z, ISLAM, S &ELTAHIR, EAB (1994) Aggregation-disaggregation properties of a stochastic
rainfall model Water Resour Res,30 3423 3435
BOER, G J, MCFARLANE, N A & LAZARE, M (1992) Greenhouse gas-induced climate change
simulated with the CCC second-generation general circulation model J Clim, 5 1045-1077
BOTKIN, DB &NISBET, RA (1992) Forest response to climatic change Effects of parameterestimation and choice of weather patterns on the reliability of projections Clim Change, 20
87-111
BOX, GEP & JENKINS, GM (1976) Time series analysis forecasting and control Holden-Day, San
Francisco a o,575 pp
BOX, EO (1978) Geographical dimensions of terrestrial net and gross productivity Radiat Environ
Biophys ,15 305-322
BOX, EO (1981) Macroclimate and plant forms an introduction lo predictive modeling in
phytogeography Junk, The Hague a o, 258 pp
References 91
BOX, EO & MEENTEMEYER, V (1991) Geographic modelling and modern ecology In Esser, G &
Overdieck, D (eds ), Modern ecology basic and applied aspects Elsevier, Amsterdam a o , pp
773-804
BRIFFA, K R & JONES, P D (1993) Global surface air temperature variations during the twentieth
century - Part 2 implications for large-scale, high-frequency paleoclimatic studies Holocene, 3
77-88
BRZEZIECKI, B, KIENAST, F & WILDI, O (1993) A simulated map of the potential natural forest
vegetation of Switzerland J Veg Set, 4 499-508
BRZEZIECKI, B, KIENAST, F & W1LDI, O (1995) Modelling potential impacts of climate change on
the spatial distribution of zonal forest communities in Switzerland J Veg Sci, 6 257-268
BURGER, G (1996) Expanded downscaling for generating local weather scenarios Clim Res, 7 111-
128
BUGMANN, H & FISCHLIN, A (1992) Ecological processes in forest gap models - analysis and im¬
provement In Teller, A , Mathy, P & Jeffers, J N R (eds ), Responses offorest ecosystems to
environmental changes Elsevier Applied Science, London and New York, pp 953-954
BUGMANN, H (1994) On the ecology of mountainous forests in a changing climate a simulation
study Diss ETH No 10638, Swiss Federal Institute of Technology, Zurich, Switzerland, 258
PP
BUGMANN, H & FISCHLIN, A (1994) Comparing the behaviour of mountainous forest succession
models in a changing climate In Beniston, M (ed ), Mountain Environments in ChangingClimates Routledge, London, pp 237-255
BUGMANN, H & SOLOMON, A M (1995) The use of a European forest model in North America a
study of ecosystem response to climate gradients J Biogeogr ,22 477-484
BUGMANN, H (1996) A simplified forest model lo study species composition along climatic gradients
Ecology, 77 2055 2074
BUGMANN, H & FISCHLIN, A (1996) Simulating forest dynamics in a complex topography using
gndded climatic data Clim Change, 34 201-211
BUGMANN, H , FISCHLIN, A & KIENAST, F (1996) Model convergence and state variable update in
forest gap models Ecol Model, 89 197 208
BUGMANN, H & CRAMER, W (1997) Improving the behaviour offorest gap models along droughtgradients Internal report, Potsdam Institute for Climate Impact Research, Potsdam, Germany, 23
PP
BURTON, P J & CUMMING, S G (1995) Potential effects of climatic change on some western
Canadian forests, based on phenological enhancements to a patch model of forest succession
Water Air Soil Pollut, 82 401414
BUSUIOC, A & VON STORCH, H (1997) Verification of GCM generated regional precipitation and of
statistical downscaling estimates Clim Res, (accepted)
CARTER, T R , PARRY, M L, HARASAWA, H & NISHIOKA, S (1994) IPCC technical guidelinesfor
assessing climate change impacts and adaptations Dep of Geography, Univ College London &
National Institute for Environmental Studies, Japan, London, Tsukuba, 59 pp
COWPERTWAIT, P S P (1991) Further developments of the Neymann-Scott clustered point process for
modeling rainfall Water Resour Res,27 1431-1438
CRAMER, W P & LEEMANS, R (1993) Assessing impacts of climate change on vegetation usingclimate classification systems In Solomon, A M & Shugart, H H (eds ), Vegetation dynamicsand global change New York & London, Chapman and Hall, pp 190-219
92 References
CUBASCH, U , HASSELMANN, K, HOCK, H , MAIER-REIMER, E ,MIKOLAJEW1CZ U
,SANTER, B D
& SAUSEN, R (1992) Time-dependent greenhouse warming computations with a coupled ocean-
atmosphere model Clim Dyn ,8 55-69
CUBASCH, U, VON STORCH, H, WASZKEWITZ, J & ZORITA, E (1996) Estimates of climate changein Southern Europe derived from dynamical climate model output Clim Res, 7 129-149
DAVIS, M B (1981) Quaternary history and the stability of forest communities In West, DC,
Shugart, H H & Botkin, D B (eds ), Forest succession concepts and application New York a o,
Springer Verlag, pp 132-153
DAVIS, MB & BOTKIN, DB (1985) Sensitivity of cool-temperate forests and their fossil pollen record
to rapid temperature change Quaternary Res,23 327-430
DAVIS, MB (1986) Climatic instability, time lags and community disequilibrium In Diamond, J &
Case, T (eds ), Community Ecology Cambridge a 0, Harper & Row Publishers Inc
, pp 269-
284
DAVIS, MB (1989) Lags in vegetation response to greenhouse warming Clim Change, 15 75-82
DAVIS, M B (1990) Biology and paleobiology of global climate change Introduction Trends Ecol
Evol.S 269-270
DICKINSON, RE (1986) How will climate change'' The climate system and modelling of future
climate In Bolin, B, Doos, B R
, Jager, J & Warrick, R A (eds ), The greenhouse effect,climatic change and ecosystems (SCOPE Vol 29) Wiley, Chichester a o
, pp 206-270
DICKINSON, R E, ERRICO, R M , GIORG!, F & BATES, G T (1989) A regional climate model for
the western United States Clim Change, 15 383-422
EFRON, B (1979) Bootstrap methods another look at the Jackntfe The Annals of Statistics,! 1-26
EHRENDORFER, M (1987) A regionalization of Austria's precipitation climate using principalcomponent analysis J Climatol,! 71 89
EISCHEID, J K, BAKER, C B
,KARL T R & DIAZ, H F (1995) The quality control of long-term
chmatological data using objective data analysis J Appl Meteorol, 34 2787-2795
ELLENBERG, H & KLOTZLI, F (1972) Waldgesellschaften und Waldstandorte der Schweiz Eidg Anst
Forstl Versuchswes, Mitt, 48 589-930
ELLENBERG, H (1988) Vegetation ecology of Central Europe 4th ed, Cambridge University Press,
Cambridge a o, 731 pp
EMANUEL, W R, SHUGART, H H & STEVENSON, M P (1985) Climatic change and the broad-scale
distribution of terrestrial ecosystem complexes Clim Change, 7 29-43
FISCHLIN, A (1982) Analyse eines Wald-Insekten-Systems Der subalpine Larchen-Arvenwald und der
graue Larchenwickler Zeiraphera diniana Gn (Lep, Tortncidae) Diss ETH No 6977, Swiss
Federal Institute of Technology, Zurich, Switzerland, 294 pp
FISCHLIN, A (1986) Simplifying the usage and the programming of modern working stations with
Modula-2 The Dialog Machine Internal report. Department of Automatic Control and Industrial
Electronics, Swiss Federal Institute of Technology Zurich (ETHZ), 15 pp
FISCHLIN, A & SCHAUFELBERGER, W (1987) Arbeitsplatzrechner im techmsch-naturwissenschaftli-
chen Hochschuluntemcht Bulletin SEV/VSE, 78 15-21
FISCHLIN, A (1991) Interactive modeling and simulation of environmental systems on workstations
In Moller, D P F (ed ), Proc of the 4th Ebernburger Working Conference on the analysis ofdynamic systems in medicine, biology, and ecology Informatik-Fachbenchte 275, Springer,Berlin ao, pp 131-145
References 93
FISCHLIN, A,GYALISTRAS, D , ROTH, O , ULR1CH, M ,
THOENY, J, NEMECEK, T, BUGMANN,
HK & THOMMEN, F (1994) ModelWorks 2 2 An interactive simulation environment for
personal computers and workstations Swiss Federal Institute of Technology ETH, Zurich,
Switzerland, Systems Ecology Report No 14, 324 pp
FISCHLIN, A (1995) Assessing sensitivities of forests to climate change experiences from modelling
case studies In Guisan, A , Holten, J I, Spichiger, R & Tessier, L (eds ), Potential ecological
impacts of climate change in the Alps and Fennoscandian mountains Conserv Jard Bot,
Geneve, pp 145 147
FISCHLIN, A, BUGMANN, H & GYALISTRAS, D (1995) Sensitivity of a forest ecosystem model to
climate parametnzation schemes Environ Pollut, 87 267-282
FISCHLIN, A & GYALISTRAS D (1997) Assessing impacts of climatic change on forests in the AlpsGlobal Ecol and Biogeogr Lett ,6 19 37
FISCHLIN, A,GUISAN, A & LISCHKE, H (1998a) A glance to the future Modelling the response of
vegetation in the Alps to climate change In Cebon, P, Dahinden, U , Davies, H C , Imboden,
D & Jager, C (eds ), Views from the Alps regional perspectives on climate change MIT Press,
Cambridge, Massachusetts a o,28 pp (in press)
FISCHLIN, A ,LISCHKE, H & BUGMANN, H (1998b) The dynamic approach Evaluating the fate of
forests In Cebon, P, Dahinden, U , Davies, H C
, Imboden, D & Jager, C (eds ), Views fromthe Alps regional perspectives on climate change Cambridge, Massachusetts a o
,MIT Press,
22 pp (in press)
FL1RI, F (1974) Niederschlag und Luftlemperatur im Alpenraum Wissenschaftl Alpenvereinshefte,24, Innsbruck, 1II pp
FLOHN, H &FANTECHI, R (eds ) (1984) The climate of Europe past present andfuture D Reidel
Publishing Company, Dordrecht a o,355 pp
GARDNER, W A (ed)(l994) Cyclostationanty in communications and signal processing IEEE, New
York, 493 pp
GARRIDO, J, GARCIA, J A & MATEOS, V L (1996) Effect of seasonal adjustment on empirical
rainfall distributions Theor Appl Climatol, 49 41-51
GATES, W L, ROWNTREE, P R & ZENG, Q-C (1990) Validation of climate models In Houghton,
J T,Jenkins, G J & Ephraums, J J (eds ), Climate change The IPCC scientific assessment
Cambridge Umv Press, Cambridge a o, pp 99-137
GATES, W L, MITCHELL, J F B
, BOER, G J,CUBASCH, U & MELESHKO, V P (1992) Climate
modelling, climate prediction and model validation In Houghton, J T,Callander, B A &
Varney, S K (eds ), Climate change 1992 The supplementary report to the IPCC scientificassessment Cambridge Umv Press, Cambridge a o
, pp 97 134
GENSLER, G A (1978) Das Klima von Graubunden ein Beitrag zur Regionalklimatologie der Schweiz
Arbettsberichte der Schweiz Meteorol Zentralanstalt No 77, Zurich
GENSLER, G & SCHUEPP, M (1991) Witterungsklimatologie von Graubunden In Elsasser, H &
Boesch, M , Beitrage zur Geographic Graubundens Bundesamt fur Landestopographie, Wabern,
Switzerland, pp 7-17
GlORGl, F, MARINUCCI, M R & VISCONTI, G (1990) Application of a limned area model nested in
a general circulation model to regional climate simulation over Europe J Geophys Res,95
18413 431
GlORGl, F & MEARNS, LO (1991) Approaches lo the simulation of regional climate change a
review Rev Geophys,2 191-216
GlORGl, F,MARINUCCI, M R & VISCONTI, G (1992) A 2xC02 climate change scenario over
Europe generaled using a limited area model nested in a General Circulation model - Part II
climate change scenario J Geophys Res D, 9(97D) 10011-10028
94 References
GlORGl, F (1996) Perspectives for Earth system modeling on the regional scale In Ghan, S J,
Pennell, W T, Peterson, K L , Rykiel, E , Scott, M J & Vail, L W (eds ), Regional impacts of
global climate change assessing change and response at the scales that matter, Battelle Press,
Coulumbus, Ohio, pp 3-19
GLAHN, HR (1985) Statistical weather forecasting In Murphy, A H & Katz, RW (eds).
Probability, statistics and decision making in the atmospheric sciences Westview Press, Boulder,
pp 289-366
GREGORY, J M, WlGLEY, T M L & JONES, P D (1993) Application of Markov models to area
average daily precipitation series and interannual variability in seasonal totals Clim Dyn ,8
299-310
GREGORY, J M & MITCHELL, J F B (1995) Simulation of daily variability of surface temperature and
precipitation over Europe in the current and 2xC02 climates using the UKMO climate model
Quart J R Met Soc,121 1451 1476
GROTCH, S L & MACCRACKEN, M C (1991) The use of general circulaton models to predict regionalclimate change J Clim
,4 286 303
GYALISTRAS, D ,VON STORCH, H
,FISCHLIN, A &BEN1STON, M (1994) Linking GCM-simulated
climatic changes to ecosystem models case studies of statistical downscaling in the Alps Clan
Res,4 167-189
GYALISTRAS, D & FISCHLIN, A (1995) Downscaling Applications to Ecosystems Modeling In
Muircheartaigh, I O (ed ), Proc of the 6th International Meeting on Statistical Climatology
Galway, Ireland, Jun 19-23,1995 University College, Gal-way,Ireland,pp 189-192
GYALISTRAS, D & FISCHLIN, A (1996) Derivation of climate change scenarios for mountainous
ecosystems a GCM-based method and the case study of Valais Switzerland Swiss Federal
Institute of Technology ETHZ, Zurich, Switzerland, Systems Ecology Report No 22, 21 pp
GYALISTRAS, D , FISCHLIN, A & RIEDO, M (1997) Herleitung stundlicher Wetterszenanen unler
zukunftigen Khmabedingungen In Fuhrer, J (ed ), Khmaanderung und Grunland - eine Modell
studie uber die Auswirkungen zukunftiger Klimaveranderungen auf das Dauergrunland in der
Schweiz vdf, Hochschulverlag AG an der ETH Zurich, pp 207-263
GYALISTRAS, D , SCHAR, C, DAVIES, H C & WANNER, H (1998) Future Alpine climate In
Cebon, P, Dahinden, U ,
Davies, H C, Imboden, D & Jager, C (eds ), Views from the Alps
regional perspectives on climate change MIT Press, Cambridge, Massachusetts a o ,36 pp (in
press)
HANSEN, J, FUNG, I, LACIS, A , RIND, D , LEBEDEFF, S
, RUEDY, R & RUSSELL, G (1988)Global climate changes as forecast by the Goddard Institute for Space Sciences three dimensional
model J Geophys Res,93 9341-9364
HASSELMANN, K & BARNETT, T P (1981) Techniques of linear prediction for systems with periodicstatistics J Atmos Set, 38 2275-2283
HASSELMANN, K,SAUSEN R
, MAIER-RE1MER, E & VOSS, R (1992) On the cold start problem in
transient simulations with coupled ocean-atmosphere models Max-Planck-Institut fur Meteo
rologie, Hamburg, Report No 83
HEGERL, G C ,VON STORCH, H , HASSELMANN, K , SANTER, B D , CUBASCH, U & JONES, P D
(1996) Detecting greenhouse gas induced climate change with an optimal fingerprint method
J Clim,9 2281-2306
HOLDRIDGE, LR (1947) Determination of world plant formations from simple climatic data Science,
105 367-368
HOLDRIDGE, LR (1967) Life zone ecology Tropical Science Center, San Jose, Costa Rica, 206 pp
References 95
HULME, M , WlGLEY, T M L & JONES, P D (1990) Limitations of regional clmate scenarios for
impact analysis In Boer, M M & de Groot, R S (eds ), Landscape ecological impact ofclimatic change IOS Press, Amsterdam, pp 111 129
HULME, M , BRIFFA, K R, JONES, P D & SENIOR, C A (1993) Validation of GCM control
simulations using indices of daily airflow types over Ihe British Isles Clim Dyn ,9 95-105
HUNTLEY, B & BIRKS, H J B (1983) An atlas of past and present pollen mapsfor Europe 0-13'000
years ago Cambridge Umv Press, Cambridge a o,667 pp
HUNTLEY, B (1990) European post glacial forests compositional changes in response to climatic
change J Veg Sci, 1 507-518
IPCC (1990) Climate change The IPCC scientific assessment, Houghton, JT, Jenkins, GJ &
Ephraums, J J (eds ) Cambridge Umv Press, Cambridge, UK, 365 pp
IPCC (1992) Climate Change 1992 The supplementary report to the IPCC scientific assessment,
Houghton, J T, Callander, B A & Varney, S K (eds ) Cambridge Umv Press, Cambridge, UK,
198 pp
IPCC (1994) Climate change 1994 Radiative forcing of climate change and an evaluation of the IPPC
IS92 emission scenarios, Houghton, J T, Filho, L G M, Bruce, J
, Lee, H , Callander, B A,
Haites, E , Harris, N & Maskell, K (eds ) Cambridge University Press, Cambridge, UK, 339
PP
IPCC (1996a) Climate change 1995 The science of climate change Contribution of Working Group I
to the Second Assessment Report of the Intergovernmental Panel on Climate Change, Houghlon,J T
,Meira Filho, L G
, Callander, B A, Harris, N , Kattenberg, A & Maskell, K (eds )
Cambridge University Press, Cambridge and New York, 572 pp
IPCC (1996b) Climate Change 1995 Impacts, adaptations and mitigation ofclimate change scientific-technical analyses Contribution of Working Group II to the Second Assessment Report of the
Intergovernmental Panel on Climate Change, Watson, R T, Zinyowera, M C & Moss, R H
(eds), Cambridge University Press, Cambridge and New York, 880 pp
JACOBE1T, J (1988) Intra seasonal fluctuations of mid-tropospheric circulation above the eastern
Mediterranean In Gregory, S (ed ), Recent climatic change, Belhaven Press, London, pp 91-
101
JESSEL, I (1991) Inventory of available observed data sets Deutsches Klimarechenzentrurn GmbH,
Hamburg, Tech Report No 1
JOHNSON.NL &KOTZ, S (1970) Continuous univariate distributions Wiley, New York a o
JONES, P D & BRIFFA, K R (1992) Global surface air temperature variations during the twentieth
century Parti spatial, temporal and seasonal details Holocene,! 165-179
KARL, T R, WANG, W -C, SCHLESINGER, M E, KNIGHT, R W & PORTMAN, D (1990) A method
relating General Circulation Model simulated climate to the observed local climate Part I
seasonal statistics J Clim, 3 1053-1079
KATTENBERG, A , GlORGl, F, GRASSL, H
, MEEHL, G A, MITCHELL, J F B
, STOUFFER, R J,
TOKIOKA, T , WEAVER, A J & WlGLEY, T M L (1996) Climate models - projections of future
climate In Houghlon, J T, Meira Filho, L G , Callander, B A,Harris, N , Kattenberg, A &
Maskell, K (eds ), Climate Change 1995 - The Science of Climate Change Contribution ofWorking Group I to the Second Assessment Report of the Intergovernmental Panel on Climate
Change, Cambridge University Press, Cambridge and New York, pp 289-357
KATZ, RW (1996) Use of stochastic models to generate climate change scenanos Clim Change, 32
237 255
KIENAST, F (1987) FORECE - A forest succession model for southern central Europe Oak RidgeNational Laboratory, Oak Ridge, Tennessee, USA, Report ORNL/TM-10575, 69 pp
96 References
KIENAST, F, BRZEZIECKI, B & WILDI, O (1987) Computergestutzte Simulation der raumlichen
Verbreitung natumaher Waldgesellschaften in der Schweiz Schweiz. Z. Forstwes, 145 293-309
KIENAST, F & KUHN, N (1989) Simulating forest succession along ecological gradients in southern
Central Europe Vegetatio, 79 7-20
KIENAST, F (1991) Simulated effects of increasing atmospheric C02 and changing climate on the
successional characteristics of Alpine forest ecosystems Landscape Ecol, 5 225-238
KIM, J W,CHANG, J T, BAKER, N L
, WILKS, D S & GATES, W L (1984) The statistical problemof climate inversion determination of the relationship between local and large-scale climate
Mon Wea Rev ,112 2069-2077
KlRSCHBAUM, MUF, KING, DA, COMINS, HN, MCMURTRIE, RE, MEDLYN, BE,
PONGRAC1C, S, MURTY, D , KEITH, H , RAISON, R J
, KHANNA, P K & SHERIFF, D W
(1994) Modelling forest response to increasing C02 concentration under nutrient-limited
conditions Plant Cell Environ,17 [081-1099
KlRSCHBAUM, M & FISCHLIN, A (1996) Climate change impacts on forests In Watson, R,
Zinyowera, M C & Moss, R H (eds), Climate Change 1995 Impacts, adaptations and
mitigation of climate change scientific-technical analyses Contribution of Working Group II to
the Second Assessment Report of the Intergovernmental Panel on Climate Change CambridgeUniversity Press, Cambridge a o, pp 95-129
KLAUS, D (1993) Zirkulations- und Persistenzanderungen des Europaischen Wettergeschehens im
Spiegel der Grosswetterlagenstatistik Erdkunde, 47 85 104
KORNER, C (1993) C02 fertilization the great uncertainty in future vegetation development In
Solomon, A M & Shugart, H H (eds), Vegetation dynamics and global change Chapman &
Hall, New York & London, pp 53-70
KORNER, C (1995) Impact of atmospheric changes on alpine vegetation the ecophysiologicalperspective In Guisan, A, Holten, JI, Spichiger, R & Tessier, L (eds), Potential ecologicalimpacts of climate change in the Alps and Fennoscandian mountains Conserv Jard Bot,
Geneve, pp 113-120
KRAUCHI, N (1993) Potential impacts of a climate change on forest ecosystems Eur J For Pathol,23 28-50
KRAUCHI, N & KIENAST, F (1993) Modelling subalpine forest dynamics as influenced by a changingenvironment Water Air Soil Pollut, 68 185-197
KRAUCHI, N (1994) Modelling forest succession as influenced by a changing environment Diss ETH
No 10479, Swiss Federal Institute of Technology, Zurich, Switzerland, 271 pp
KREYSZIG, E (1977) Statistische Methoden und ihre Anwendungen Vandenhoeck & Ruprecht,Gottingen, 451 pp
KUTZBACH, JE (1967) Empirical eigenvectors of sea-level pressure, surface temperature and
precipitation complexes over North America J Appl Met, 6 791-802
LISCHKE, H & BLAGO, N (1990) A model to simulate the population dynamics of the codling moth
(Cydia Pomonella L (Lepidoptera, Torticidae)) development and male moth flight Acta
Horttcultura,276 43-52
LOTTER, A & KIENAST, F (1992) Validation of a forest succession model by means of annuallylaminated sediments In Saarnisto, M & Kahra, A (eds), Proceedings of the workshopLaminated sediments, Lammi Biological Station, 4-6, June 1990 Geological Survey of Finland,
pp 25-31
LOUGH, J M, WlGLEY, T M L & PALUTIKOF, J P (1984) Climate and climate impact scenarios for
Europe in a wanner world J Clim and Appl Met ,22 1673-1684
References 97
LUPO, A R, OGLESBY, R J & MOKHOV, I I (1997) Chmatological features of blocking
anticyclones a study of Northern Hemisphere CCM1 model blocking events in present-day and
double C02 concentration atmospheres Clim Dyn ,13 181-195
MANABE, S,STOUFFER, RJ, SPELMAN, M J & BRYAN, K (1991) Transient responses of a
coupled ocean atmosphere model to gradual changes of atmospheric C02 Part I Annual mean
response J Clim,4 785-818
MANABE, S, SPELMAN, M J & STOUFFER, R J (1992) Transient responses of a coupled ocean-
almosphere model lo gradual changes of atmospheric C02 Part II Seasonal response J Clim,5 105-126
MARINUCCI, M R,GlORGl, F
,BEN1STON, M
,WILD, M
, TSCHUCK, P & BERNASCONI, A (1995)High resolution simulations of January and July climate over Ihe Western Alpine region with a
nested regional modeling system Theor Appl Climatol, 51 119-138
MCFARLANE, N A, BOER, G J
, BLANCHET, J -P & LAZARE, M (1992) The Canadian Climate
Centre second-generation general circulation model and its equilibrium climate J Clim,5
1013 1044
MCGUIRE, A D, JOYCE, L A KICKLIGHTER D W
,MELILLO, J M
, ESSER, G & VOROSMARTY,
C J (1993) Productivity response of climax temperate forests to elevated temperature and carbon
dioxide a North American comparison between two global models Clim Change, 24 287-310
MEARNS, L O, KATZ, R W & SCHNEIDER, S S (1984) Extreme high temperature events changes in
their probabilities with changes in mean temperature J Clim Appl Meteorol ,23 1601-1613
MEARNS, L O, GlORGl, F
,SHIELDS BRODEUR, C
, & MCDAN1EL, L (1995a) Analysis of the
variability of daily precipitation in a nested modeling experiment comparison with observations
and comparison with 2xC02 results Global and Planetary Change, 10 55-78
MEARNS, L O, GlORGl, F SHIELDS BRODEUR, C ,
& MCDANIEL, L (1995b) Analysis of the
diurnal range and variability of daily temperature in a nested modeling experiment comparisonwith observations and comparison with 2xC02 results Clim Dyn ,
11 193-209
MEARNS, L O, ROSENZWEIG, C, & GOLDBERG, R (1996) The effect of changes in daily and
interannual variability on CERES-wheal a sensitivity study Clim Change, 32 257-292
MEEHL, G A, WHEELER, M & WASHINGTON, W M (1994) Low-frequency variability and C02
transient climate change Part 3 Intermonthly and interannual variability Clim Dyn, 10 277-
303
MELILLO, J M, MCGUIRE, A D
, KICKLIGHTER, D W,MOORE III, B , VOROSMARTY, C J &
SCHLOSS, AL (1993) Global climate change and terrestrial net primary production Nature,363 234 240
MELILLO, J M, PRENTICE, I C
, FARQUHAR, G D, SCHULZE, E -D & SALA, O E (1996)
Terrestrial biolic responses to environmental change and feedbacks to climate In Houghton, J T,Meira Filho, L G
, Callander, B A, Harris, N , Kattenberg, A & Maskell, K (eds ), Climate
change 1995 The science of climate change Contribution of Working Group I to the Second
Assessment Report of the Intergovernmental Panel on Climate Change Cambridge UniversityPress, Cambridge a o
, pp 445 481
METHERELL, A K, HARDING, L A
, COLE, C V & PARTON, W J (1994) Century - Soil OrganicMalter Model - Environment Great Plains System Research Unit, Technical Report No 4, Fort
Collins, Colorado
MITCHELL, J FB & INGRAM WJ (1992) On C02 and climate mechanisms of changes in clouds J
Clim ,5 5 21
MOORE, A D (1989) On the maximum growth equation used in forest gap simulation models Ecol
Model, 45 63 67
98 References
NILSSON, S & PITT, D (1991) Climate change and mountain forests In Nilsson, S & Pitt, D ,
Mountain world in danger climate change in the mountains andforests of Europe Earthscan
Publications, London, pp 57-81
NONHEBEL, S (1994) Inaccuracies in weather data and their effects on crop growth simulation results -
I Potential production Clim Res,4 47-60
OELSCHLAGEL, B (1995) A method for downscaling global climate model calculations by a statistical
weather generator Ecol Model, 82 199-204
OVERPECK, J T, RIND, D & GOLDBERG, R (1990) Climate-induced changes in forest disturbance and
vegetation Nature, 343 51-53
OVERPECK, J T, BARTLEIN, P J & WEBB HI, T (1991) Potential magnitude of future vegation
change in eastern North America comparisons with the past Science, 254 692-695
OZENDA, P & BOREL, J -L (1990) The possible responses of vegetation to a global climatic change -
Scenanos for western Europe with special reference to Ihe Alps In Boer, M & Degroot, R S
(eds ), Landscape-ecological impact of climatic change - Proceedings ofa European conference,IOS Press, Amsterdam a o
, pp 221-249
PARRY, M L, CARTER, T R & KONJIN, N T (eds ) (1988) The impact of climatic variations on
agriculture. Vol I assessments in cool temperate and cold regions Kluwer Academic Publisher,Dordrecht a o, 876 pp
PARTON, W J, SCURLOCK, J M O
, OJIMA, D S, GILMANOV, T G
, SCHOLES, R J, SCHIMEL,
D S, KIRCHNER, T, MENAUT, J -C
,GARCIA MOYA, E & APINAN KAMNALRUT, K
,J I
(1993) Observation and modelling of biomass and soil organic matter dynamics for the grasslandbiome worldwide Global Biogeochem. Cycles, 7 785 809
PASTOR, J & POST, W M (1988) Response of northern forests to C02-induced climate changeNature, 334 55-58
PERRUCHOUD, D & FISCHLIN, A (1995) The response of the carbon cycle in undisturbed forest
ecosystems to climate change A review of plant-soil models J Biogeogr, 11 759-774
PERRUCHOUD, D O (1996) Modeling the dynamics ofnonliving organic carbon in a changing climate
a case studyfor temperate forests Diss ETH No 11900, Swiss Federal Institute of Technology,Zunch, Switzerland, 196 pp
PITTOCK, AB & SALINGER, M J (1982) Towards regional scenarios for a C02-warmed earth Clim
Change, 4 23-40
PREISENDORFER, R W , ZWIERS, F W & BARNETT, T P (1981) Foundations of principal componentselection rules Scripps Inst Ocn
,La Jolla, California, SIO Ref Series 81-4
PREISENDORFER, R W (1988) Principal component analysis in meteorology and oceanographyElsevier, Amsterdam a o
,425 pp
PRENTICE.IC & HELMISAAR1, H (1991) Silvics of North European trees compilation, comparisonsand implications for forest succession modelling For Ecol Manage ,
42 79-93
PRENTICE, I C, BARTLEIN, P J & WEBB III, T (1991) Vegetation and climate changes in eastern
North America since the last glacial maximum Ecology, 72 2038-2056
PRENTICE, IC , SYKES, M T & CRAMER, W (1993) A simulation model for the transient effects of
climate change on forest landscapes Ecol Modelling, 65 51-70
RAICH, J W, RASTETTER, E B
, MELILLO, J M, KICKLIGHTER, D W , STEUDLER, P A
, PETERSON,B J
, GRACE, A L,MOORE III, B & VOROSMARTY, C J (1991) Potential net primary
production in South America application of a global model Ecol Appl, 1 399-429
RAICH, J W & SCHLESINGER, W H (1992) The global carbon dioxide flux in soil respiration and its
relationship to vegetation and climate Tellus Series B, 44 81-99
References 99
RASTETTER, E B, RYAN, M G
, SHAVER, G R, MELILLO, J M
, NADELHOFFER, K J, HOBBIE, J E
& ABER, J D (1991) A general biogeochemical model describing the responses of the C and N
cycles in terrestrial ecosystems to changes in C02, climate, and N deposition Tree Physiol, 9
101-126
RENNER, F (1982) Beitrage zur Gletschergeschichte des Gotthardgebietes und dendrokltmatologtscheAnalysen an fossilen Holzern Diss Umv of Zurich, Department of Geography, Zurich,
Switzerland, 180 pp
RICHARDSON, C W (1981) Stochastic simulation of daily precipitation, temperature, and solar radia¬
tion Water Resour Res,17 182-190
RIEDO, M , GYALISTRAS, D, GRUB, A , ROSSET, M & FUHRER, J (1997b) Modelling grassland
responses to climate change and elevated C02 Acta Oecol Appl ,18 (in press)
ROBOCK, A , TURCO, R P, HARWELL, M A
, ACKERMAN, T P, ANDRESSEN, R, CHANG, H S &
SIVAKUMAR, M V K (1993) Use of general circulation model output in the creation of climate
change scenarios for impact analysis Clim Change, 23 293 335
ROEDEL, W (1992) Physik unserer Umwell Die Atmosphare Springer, Berlin a o,457 pp
ROTACH, M W , MARINUCCI, M R, WILD, M , TSCHUCK, P
, OHMURA, A & BENISTON, M
(1997) Nested regional simulation of climate change over the Alps for the scenano of a doubled
greenhouse forcing 7yieor Appl Climatol, 57 209-227
ROTH, O , DERRON, J O, FISCHLIN, A , NEMECEK, T & ULRICH, M (1995) Implementation and
parameter adaptation of a potato crop simulation model combined with a soil water subsystemIn Kabat, P , Marshall, B ,
van den Broek, B J, Vos, J & van Keulen, H (eds ), Modelling and
parameterization of the soil plant atmosphere system - a comparison ofpotatio growth models
Wageningen Press, Wageningen, The Netherlands, pp 371-399
SACHS, L (1984) Applied statistics - a handbook oftechniques 2nd ed, Springer, New York a o
,707
PP
SAMPSON, DA, COOTER.E J, DOUGHERTY, PM & LEE ALLEN, H (1996) Comparison of the
UKMO and GFDL GCM climate projections in NPP simulations for southern loblolly pinestands Clim Res
,7 55-69
SANTER, B D, WlGLEY, T M L
, SCHLESINGER, M E & MITCHELL, J F B (1990) Developingclimate scenarios from equilibrium GCM-results Max-Planck-Institut fur Meteorologie,Hamburg, Report No 47, 29 pp
SANTER, B D, TAYLOR, K E
, WlGLEY, T M L, JOHNS, T C
, JONES, P D, KAROLY, D J
,
MITCHELL, J F B, OORT, A H , PENNER, J E
, RAMASWAMY, V, SCHWARZKOPF, M D
,
STOUFFER, R J & TETT, S (1996) A search for human influences on the thermal structure of
the atmosphere Nature, 382 39-46
SCHAR, C, FREI, C
, LUTHI, D & DAVIES, H C (1996) Surrogate climate change scenarios for
regional climate models Geophys Res Lett, 23 669-672
SCHAR, C, DAVIES, T, FREI, C, WANNER, H, WIDMANN, M , WILD, M & DAVIES, H C (1998)Current Alpine climate In Cebon, P
, Dahinden, U , Davies, H C, Imboden, D & Jager, C
(eds ), Views from the Alps regional perspectives on climate change MIT Press, Cambridge,Massachusetts a o, 28 pp (in press)
SCHIMEL, D S, PARTON, W J
, KITTEL, T F G, OJIMA, D S & COLE, C V (1990) Grassland
biogeochemistry links to atmospheric processes Clim Change, 17 13-25
SCHRODER, W (1976) Zur Populationsokologie und zum Management des Rothirsches (Cervuselaphus L ) dargestellt an einem Simulationsmodell und der Reproduktionsleislung des
Rothirschbestandes im Han. Habtlttationsschnft, Ludwig Maximilians-Univ, Munchen, 198 pp
SCHUEPP, M (1968) Kalender der Wetter- und Witterungslagen im zentralen Alpengebiet VeroffentlSchweiz Meteorol Zanstalt No 11, 43 pp
100 References
SCHUEPP, M &SCHIRMER, H (1977) Climates of Central Europe In Wallen, CC (ed), Climates
of Central and Southern Europe World Survey of Climatology, Elsevier, Amsterdam-Oxford-New
York, pp 3-74
SCHWEINGRUBER, F H (1988) Climatic information for the past hundred years in width and density of
conifer growth nngs In Wanner, H & Siegenthaler, U (eds ), Long and short term variability
of climate. Lecture notes in earth sciences, 16 Springer, Berlin a o, pp 35-56
SEMTNER, A J (1995) Modeling ocean circulation Science, 269 1379-1385
SHUGART, H H (1984) A theory offorest dynamics - The ecological implications offorest successionmodels Spnnger, New York a o
,278 pp
SHUGART, H H, ANTONOVSKY, M Y
, JARVIS, P G & SANDFORD, A P (1986) C02, climatic
change, and forest ecosystems In Bolin, B , Doos, B R, Jager, J & Warrick, R A (eds), 77ie
greenhouse effect, climatic change, and ecosystems Wiley, Chichester a o, pp 475-522
SHUGART, H H (1990) Using ecosystem models to assess potential consequences of global climatic
change Trends Ecol Evol, 5 303-307
SHUGART, H H , LEEMANS, R & BONAN, G B (eds ) (1992) A systems analysis of the global boreal
forest Cambndge Umv Press, Cambridge a o, 565 pp
SMA (1901-1980) Annalen der Schweizenschen Meteorologtschen Anstalt Schweiz Meteorol Anstalt,
Zurich
SMITH, T M , SMITH, J B & SHUGART, H H (1992) Modeling the response of terrestrial vegetation
to climate change in the tropics In Goldammer, J G (ed) Tropical forests in transition - Ecology
ofnatural and anthropogenic disturbance processes Birkhauser Verlag, Basel-Boston, pp 253-268
SMITH, T M , LEEMANS, R & SHUGART, H H (1994) The application ofpatch models of vegetation
dynamics to global change issues - a workshop summary Kluwer, Dordrecht, 1-22 pp
SOLOMON, A M , DELCOURT, H R, WEST, D C & BLASING, T J (1980) Testing a simulation
model for reconstrucUon of prehistoric forest-stand dynamics Quaternary Res, 14 275 293
SOLOMON, A M, WEST, D C & SOLOMON, J A (1981) Simulating the role of climate change and
species immigration in forest succession In West, D C , Shugart, H H & Botkin, D B (eds),Forest succession concepts and application Springer, New York a.o, pp 154-177
SOLOMON, A M & WEBB-III, T (1985) Computer-aided reconstruction of late-quaternary landscapedynamics Annu Rev Ecol Syst, 16 63-84
SOLOMON, AM (1986) Transient response of forests to C02-induced climate change simulation
modeling expenments in eastern North America Oecologia, 68 567-579
SOLOMON, A M & WEST, D C (1987) Simulating forest ecosystem responses to expected climate
change in eastern North America Applications to decision making in the forest industry In
Shands, W E & Hoffmann, J S (eds), The greenhouse effect, climate change, and U S Forests
The Conservation Foundation, Washington, pp 189-217
SOLOMON, A M (1988) Ecosystem theory required to identify future forest responses to changing C02
and climate In Wolff, W , Soeder, C -J & Drepper, F R (eds), Ecodynamics contributions to
theoretical ecology Springer, Berlin a o, pp 258-274
SOLOMON, AM & BARTLEIN, PJ (1992) Past and future climate change response by mixed
deciduous-coniferous forest ecosystems in northern Michigan Can J For Res, 22 1727-1738
SPITTERS, C J T, VAN KEULEN, H & VAN KRAALINGEN, DWG (1989) A simple and universal
crop growth simulator SUCROS87 In Rabbinge, R,Ward, S A & Van Laar, H H (eds ),
Simulation and systems management in crop protection, Pudoc, Wageningen, pp 147-181
References 101
STADLER, D, BRUENDL, M , SCHNEEBELI, M , MEYER-GRASS, M ,
FLUEHLER, H (1997)
Wasserregime m der Schneedecke und tm Boden an bewaldeten Gebirgsstandorten
Schneeinterzeption, Schmelzwasser- und Oberflaechenabfluss vdf Hochschulverlag AG an der
ETH Zunch
STEPHENSON, N L (1990) Climatic control of vegetation distribution the role of die water balance
Amer Nat, 135 649-670
STOCKTON, C W, BOGGESS, WR & MEKO, DM (1985) Climate and tree nngs In Hecht, A D
(ti), Paleoclimate analysis and modeling John Wiley, New York a o , pp 71-161
SUPIT, I (1986) Manual for generation of daily weather data Centre for Agrobiological Research,
Wageningen, Simulation Reports CABO-TT No 7, 46 pp
SWARTZMAN, GL & KALUZNY, S P (1987) Ecological simulation primer Macmillan PublishingCo, New York, 370 pp
TAUBES, G (1995) Is a warmer climate wilting the forests of the North'' Science, 267 1595
THOENY, J, FISCHLIN, A & GYALISTRAS, D (1994) RASS Towards bridging the gap between
interactive and off-line simulation In Halin, J, Karplus, W & Rimane, R (eds), CISS First
Joint Conference of International Simulation Societies, August 22-25, 1994, Zurich SwitzerlandThe Society for Computer Simulation International, San Diego, California, pp 99-103
THOENY, J, FISCHLIN, A & GYALISTRAS, D (1995) Introducing RASS - the RAMSES simulation
server Institute of Terrestrial Ecology, Swiss Federal Institute of Technology ETH, Zurich,
Switzerland, Systems Ecology Report No 21, 32 pp
THORNTHWAITE, C W & MATHER, J R (1957) Instructions and tables for computing potentialevapotranspiration and the water balance Publications in Climatology, 10 183-311
TSONIS, A A (1996) Widespread increases in low-fequency variability of precipitation over the pastcentury Nature, 382 700-702
UTT1NGER, H (1970) Niederschlag Kltmatologie der Schweiz, E, 5 bis 8 Teil, Schweiz Meteorol
Zentralanstalt, Zurich, 162 pp
VINNIKOV, K Y , GROISMAN, P Y, LUGINA, K M & GOLUBEV, A A (1987) The mean surface air
temperature vanations in the Northern Hemisphere over 1841-1985 Meteor Hydrol, 1 45-55
(in Russian)
VON STORCH, H, ZORITA, E & CUBASCH, U (1993) Downscaling of global climate change estimates
to regional scales an application to Iberian rainfall in wintertime J Clim,6 1161-1171
VON STORCH, H (1995) Inconsistencies at the interface of climate impact studies and global climateresearch Meteorol Z, 4 72-80
VOSE, R S, SCHMOYER, R L
, STEURER, P M, PETERSON, T C
,HEIM, R
, KARL, T R &
EISCHEID, J (1992) The Global Historical Climatology Network long term monthlytemperature, precipitation, sea-level pressure, and station pressure data Report ORNL/CDIAC-53, NDP-041 (available from Carbon Dioxide Information Analysis Center, Oak Ridge National
Laboratory, Oak Ridge, Tennessee)
WALTER, H &BRECKLE, SW (1983) Okologie der Erde Band I Okologische Grundlagen in globalerSicht GusUv Fischer, Stuttgart, 238 pp
WALTER, H & BRECKLE, S W (1984) Okologie der Erde Band 2 Speztelle Okologie der tropischenund Subtropischen Zonen Gustav Fischer, Stuttgart, 461 pp
WALTER, H & BRECKLE, SW (1986) Okologie der Erde Band 3 Speztelle Okologie der Gemassigtenund Arktischen Zonen Euro-Nordasiens Gustav Fischer, UTB Grosse Reihe, Stuttgart, 587 pp
102 References
WALTER.H &BRECKLE.SW (1991) Okologie der Erde Band 4 Speztelle Okologie der Gemassigtenund Arknschen Zonen ausserhalb Euro-Nordasiens Gustav Fischer, UTB Grosse Reihe, Stuttgart,586 pp
WASHINGTON, W M & PARKINSON, C L (1986) An introduction to three-dimensional climate
modeling Oxford University Press, Oxford a o, 422 pp
WEBB III, T (1992) Past changes in vegetation and climate lessons for the future In Peters, RL &
Lovejoy, T E (eds ), Global Warming and Biological Diversity Yale University Press, New
Haven & London, pp 59-75
WERNER, P,VON STORCH, H (1993) Interannual vanability of Central European mean temperature in
January/February and its relation to the large-scale circulation Clim. Res, 3 195-207
WEST, DC , SHUGART, H H & BOTKIN, D B (eds ) (1981) Forest succession concepts and
application , Spnnger, New York a o, 517 pp
WlGLEY, T ML, JONES, P D.KELLY, PM (1980) Scenano for a warm, high C02-world Nature,
283 17-21
WlGLEY, T M L, JONES, P D
, KELLY, P M (1986) Empirical climate studies In B Bohn et al
(eds ), The greenhouse effect, cimatic change and ecosystems John Wiley, New York, pp 271-
322
WlGLEY, T M L, JONES, P D
, BRIFFA.K R , SMITH,G (1990) Obtaining sub-grid scale information
from coarse-resolution general circulation model output J Geophys Res, 95 1943-1953
WILKS, D S (1989) Statistical specification of local surface weather elements from large-scaleinformation Theor Appl Climatol,40 119-134
WILKS, D S (1992) Adapting stochastic weather generation algorithms for climate change studies
Clim Change, 22 67-84
WOODWARD, FI (1987) Climate and plant distribution Cambndge University Press, Cambridge a o,
174 pp
WRIGHT, H E J, KUTZBACH, J E
, WEBB-III, T, RUDDIMAN, W F,STREET-PERROTT, F A &
BARTLEIN, P J (eds) (1993) Global climates since the last glacial maximum University of
Minnesota Press, Minneapolis, 569 pp
ZORITA, E , HUGHES, J P, LETTEMAIER, D P & VON STORCH, H (1995) Stochastic characterization
of regional circulation patterns for climate model diagnosis and estimation of local precipitationJ Clim,8 1023-1042
103
Curriculum Vitae
I was born in Athens, Greece, on Sept 17th, 1964 In 1975 my family emigrated to
Munich were I visited the secondary school In 1983 I graduated from high-school, and
in the same year I began studying Electrical Engineering at the Swiss Federal Institute of
Technology, ETH Zurich In 1987 I did a semester work on modelling the response of
glaciers to climatic change under the supervision of Prof A Ohmura, Geographical
Institute ETHZ After receiving my Diploma in 1988 I worked until 1991 part time as a
software engineer, and part time as as a research associate and teaching assistant at the
Systems Ecology Group ETHZ, directed by Dr A Fischlin From 1991 to 1992 I
worked as a full-time research scientist, and from 1993 to 1995 as a selfemployed scien¬
tific contractor associated with the Systems Ecology Group ETHZ From 1991 on to the
present I have been working on the derivation of climatic scenarios and their application
to assess possible impacts of climatic change on forests, grasslands and snow cover in
the Alpine region Since eight years I live with my partner Barbara Davies in Zurich