the frequency of catastrophic wind …ufdcimages.uflib.ufl.edu/uf/e0/04/99/40/00001/rifai_s.pdf ·...
TRANSCRIPT
THE FREQUENCY OF CATASTROPHIC WIND DISTURBANCES IN THE NORTHWEST
AMAZON, AND THEIR IMPACT UPON TREE MORTALITY AND DIVERSITY
By
SAMI WALID RIFAI
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
2016
© 2016 Sami W. Rifai
To my family for exposing me to nature as a child, commiserating with me over habitat loss, and
showing me the hope of conservation
4
ACKNOWLEDGMENTS
First and foremost I would like to thank my family, Susan, Abdalla, and Randa Rifai for
their relentless support of my decision to endeavor on this fairly unprofitable career path. Next, I
would like to thank my many supervisors over the years for inspiring me and giving me the
opportunities to continue in research: Stephanie Bohlman, Jeffrey Chambers, Daniel Markewitz,
Michael Goulden, Greg Winston, Guaciara dos Santos, and John Southon. I am extremely
grateful to the many students, faculty, and materos of the Facultad de Ciencias Forestales at La
Universidad Nacional de la Amazonía Peruana - for without their help, this work would not have
been possible. This work would also not have been possible without the financial support of
NASA for a Biodiversity grant (NNX09AK21G) awarded to Jeffrey Chambers, and an Earth and
Space Science graduate fellowship (NNX13AN56H). I also acknowledge my collaborators over
much of this project, José David Urquiza Muñoz, Robinson Negrón-Juarez, Daniel Magnabosco
Marra, and Liana O. Anderson. As well, I thank the many faculty at UF who have suffered my
endless questioning on ecology, forestry, statistics, and the academy: Wendell Cropper, Jeremy
Lichstein, Jack Putz. Finally, I would like to thank the many friends I have made over the course
of my many years in the Ph.D. program, of which I don't have enough space to list them all, but
here are some I would like to give a heartfelt shout-out to: Craig Bateman, Ana Carolina, Trevor
Caughlin, Nathan Cooper, Sarah Graves, Jenny Hazelhurst, Rebecca Hazen, Sebastian Palmas,
Anand Roopsind, Bryan Tarbox, Martijn Slot, Thales West, and Lucia Zarba.
5
TABLE OF CONTENTS
page
ACKNOWLEDGMENTS ...............................................................................................................4
LIST OF TABLES ...........................................................................................................................8
LIST OF FIGURES .........................................................................................................................9
LIST OF ABBREVIATIONS ........................................................................................................12
ABSTRACT ...................................................................................................................................14
CHAPTER
1 INTRODUCTION ..................................................................................................................16
2 LANDSCAPE-SCALE CONSEQUENCES OF DIFFERENTIAL TREE MORTALITY
FROM CATASTROPHIC WIND DISTURBANCE IN THE AMAZON ............................18
Introduction .............................................................................................................................18 Methods ..................................................................................................................................20
Study Site .........................................................................................................................20
Blowdown Identification and Satellite Image Analyses .................................................21 Field Data Collection .......................................................................................................22
Modeling Tree Mortality Rates .......................................................................................23 Model 1: pixel-scale mortality depends on ΔNPV ...................................................24
Model 2: individual-tree mortality depends on tree size and wood density
within ΔNPV classes ............................................................................................25
Model 3: Individual-tree mortality depends on ΔNPV, elevation, tree size, and
wood density .........................................................................................................25 Model 4: Individual-tree mortality with spatial dependence ...................................26
Model selection and assessment ......................................................................................26 Landscape simulation of blowdown necromass ..............................................................27
Results.....................................................................................................................................28 Overview .........................................................................................................................28
Plot-based tree mortality estimates ..................................................................................29
Differential tree mortality models ...................................................................................29
Spatial versus non-spatial GLM ......................................................................................29 Goodness of Fit of Model Necromass Predictions ..........................................................30 Landscape necromass estimation from simulation modeling ..........................................30
Discussion ...............................................................................................................................31 Differential tree mortality resulting from catastrophic wind disturbance .......................31
Implications of spatial autocorrelation for estimating mortality effects and
landscape-scale necromass...........................................................................................34 Conclusions.............................................................................................................................35
6
3 SUSTAINED REDUCTION OF TREE SPECIES RICHNESS IN NORTHWESTERN
AMAZONIAN FORESTS FOLLOWING CATASTROPHIC WIND
DISTURBANCES ..................................................................................................................44
Introduction .............................................................................................................................44
Methods ..................................................................................................................................47 Site selection by remote sensing, forest plot installation, inventory, and tree
identification ................................................................................................................47 Classification into disturbance classes ............................................................................48 Quantifying niche partitioning with disturbance intensity ..............................................49
Diversity accumulation curves ........................................................................................50 Results.....................................................................................................................................51
Niche partitioning by pioneer species with disturbance intensity ...................................51 Reductions in species richness and evenness with disturbance intensity and time .........52
Discussion ...............................................................................................................................53 Wind disturbance, differential mortality and niche partitioning by pioneer species .......54
Relation to Canopy Gap Partitioning and the Intermediate Disturbance Hypotheses .....55 An inconsistent forest succession trajectory ....................................................................57
Remaining questions of disturbance and diversity through time in a changing
climate ..........................................................................................................................58 Conclusions.............................................................................................................................59
4 SEA SURFACE TEMPERATURE ANOMALIES AS THE UNDERLYING CAUSE
OF INTERANNUAL VARIATION OF CATASTROPHIC WIND DISTURBANCE IN
THE NORTHWEST AMAZON ............................................................................................66
Introduction .............................................................................................................................66
Methods ..................................................................................................................................69 Study area description .....................................................................................................69
Use of multispectral imagery to map blowdowns ...........................................................69 The visible blowdown:forest area ratio ...........................................................................70 Remote sensing workflow ...............................................................................................71
Description of false positive blowdown correction .........................................................72 Climate data sources ........................................................................................................73 Power-law estimation of blowdown sizes .......................................................................74 Statistical analyses ...........................................................................................................75
Results.....................................................................................................................................75
Discussion ...............................................................................................................................77
The spatial, temporal, and size variation of forest blowdowns .......................................77 Regional scale climatic drivers of forest wind disturbance .............................................80 Macroscale driver of forest wind disturbance .................................................................81
Conclusions.............................................................................................................................82
5 OVERALL CONCLUSIONS .................................................................................................90
APPENDIX
7
A ADDITIONAL TABLES AND FIGURES FOR CHAPTER 2 .............................................92
B ADDITIONAL TABLES AND FIGURES FOR CHAPTER 3 ...........................................108
C ADDITIONAL TABLES AND FIGURES FOR CHAPTER 4 ...........................................116
LIST OF REFERENCES .............................................................................................................129
BIOGRAPHICAL SKETCH .......................................................................................................141
8
LIST OF TABLES
Table page
2-1 β coefficients for fixed effects in a non-spatial Generalized Linear Model (GLM)
(Model 3), and a spatial GLM (Model 4) ..........................................................................37
2-2 Goodness of fit between model necromass predictions and observed necromass from
transect plot data ................................................................................................................38
3-1 Site descriptions .................................................................................................................60
A-1 Mean landscape simulation estimate across 1000 parameter samples of the sum of
predicted dead trees and necromass across the different models. ......................................92
A-2 Logistic regression estimates for Model 2 – GLMM .........................................................93
A-3 Estimated individual dead trees (> 10 cm DBH) and associated necromass produced
by a 300 ha blowdown in the landscape (~500.5 ha) .........................................................94
A-4 Relative abundance of families comprising more than 1% of total individuals. ...............95
B-1 Landsat 5 TM images of blowdown sites before and after disturbance. .........................108
C-1 Blowdowns by Landsat scene location. ...........................................................................116
C-2 Relation of regional climate variables towards blowdown activity .................................117
C-3 Correlations between annual ratio of median visible blowdown to visible forest
(BD:F) in the northwest Amazon and SST indices ..........................................................118
C-4 Linear model estimates of SST anomalies on the ratio of visible blowdowns to
visible forest .....................................................................................................................119
9
LIST OF FIGURES
Figure page
2-1 2009 blowdown study location roughly 100-km south of Iquitos, Perú in the
department of Loreto..........................................................................................................39
2-2 The fraction of observed adult tree mortality vs. the change in pre- and post-
disturbance non-photosynthetic disturbance (ΔNPV) with regressions through the
origin (Model 1) .................................................................................................................40
2-3 Estimated probability of mortality from a catastrophic wind disturbance of a tree
with median tree characteristics, as predicted by the spatial logistic regression model
with the SRE (Model 4).. ...................................................................................................41
2-4 The ratio of estimated mortality probabilities between small and large (2.5 and
97.5% quantiles of the transect data, respectively) for tree characteristics as a
function of disturbance intensity ........................................................................................42
2-5 Total estimated tree mortality and corresponding necromass from the simulated
landscape using four different models ...............................................................................43
3-1 Blowdown site locations in Loreto, Peru. ..........................................................................61
3-2 ΔNPV derived from pre- and post-disturbance Landsat 5 images (30 × 30 m2
resolution) for blowdown sites...........................................................................................62
3-3 Pioneer tree occurrence with disturbance at the Orosa blowdown (22 yrs time since
disturbance; TSD) ..............................................................................................................63
3-4 Species accumulation curves (Hill number: q = 0) rarefied by sample coverage for
blowdown sites...................................................................................................................64
3-5 Simpson’s evenness accumulation curves rarefied by sample coverage for blowdown
sites ....................................................................................................................................65
4-1 The 20 Landsat WRS Path/Row tiles of the northwest Amazon study region. .................84
4-2 Remote sensing workflow..................................................................................................85
4-3 Kernel density estimation of large blowdowns (> 25 ha) ..................................................86
4-4 Ratio of blowdown : forest area by Landsat location from 1984 to 2011 .........................87
4-5 Temporal rainfall trend and blowdown trend ....................................................................88
4-6 Predictive relationship between SSTs and BD:F ...............................................................89
A-1 Moran’s I statistic for model predictions residuals ............................................................96
10
A-2 Polynomial (3rd degree) and LOESS functions to estimate the spatial random effect
(SRE) as a function of ΔNPV. ...........................................................................................97
A-3 The ratio of the probability of mortality of different sized trees across disturbance
intensity to when spatial autocorrelation is included versus when the spatial random
effect is excluded. ..............................................................................................................98
A-4 Estimated mortality probability of a tree with median characteristics with increasing
disturbance intensity as predicted by a spatial logistic regression model (SPDE
GLM) .................................................................................................................................99
A-5 The differential probability of death (Model 4) for trees with different diameters to
wood density ratios with (bottom) and without (top) the SRE, estimated from the
spatial GLM .....................................................................................................................100
A-6 The estimated multiple of probability of death without the SRE across a gradient of
disturbance intensity (ΔNPV) ..........................................................................................101
A-7 The distribution of mean necromass per tree by simulation run of the different
models ..............................................................................................................................102
A-8 Distribution of pixel ΔNPV in the blowdown encompassing landscape. ........................103
A-9 Ratio of model predicted to observed necromass by ΔNPV class ...................................104
A-10 Elevation of transect subplots by associated ΔNPV. .......................................................105
A-11 The mean Global Wood Density Database values of 46 species compared to
measured wood density ....................................................................................................106
A-12 The fraction of observed adult tree mortality in the calibration plots (green) and the
transect plots (blue) vs. the change in pre- and post- disturbance non-photosynthetic
disturbance (ΔNPV). ........................................................................................................107
B-1 Percentage of pioneer stems along ΔNPV gradient at the Nauta blowdown (1 yr
TSD). ................................................................................................................................109
B-2 Nauta pioneer species contribution, 1 yr TSD. ................................................................110
B-3 Mishana pioneer species contribution ..............................................................................111
B-4 Napo pioneer species contribution. ..................................................................................112
B-5 Alpahuayo pioneer species contribution ..........................................................................113
B-6 Species richness accumulation curves (Hill number: q = 0) rarefied by individuals for
blowdown sites.................................................................................................................114
11
B-7 Simpson’s evenness accumulation curves rarefied by individuals for blowdown sites
..........................................................................................................................................115
C-1 Annual rainfall map of northwest Amazon. .....................................................................120
C-2 Fraction of cloud-free observations from MODIS Terra at 1 km pixel scale ..................121
C-3 Blowdown and the algorithm derived blowdown polygon ..............................................122
C-4 Map of 28 years of northwest Amazon blowdowns ........................................................123
C-5 The probability of different sized blowdown events occurring with α exponent
estimates from different studies in the Amazon ..............................................................124
C-6 BD:F and individual SST indices ....................................................................................125
C-7 The annual max precipitation rate increased between 1984 - 2012 for all northwest
Amazon scenes in the lowland. ........................................................................................126
C-8 High rainfall event frequency across the northwest Amazon. .........................................127
C-9 Annual standard deviation between 20 Landsat WRS scene locations of the ratio of
blowdown area:visible forest area (BD:F). ......................................................................128
12
LIST OF ABBREVIATIONS
AMO Atlantic Multidecadal Oscillation, an index of a sea surface temperature
anomaly in the north Atlantic.
BD:F The ratio of visible blowdown to visible forest.
CGP Canopy gap partitioning hypothesis, a hypothesis of tree species
coexistence.
DBH Diameter at breast height, a standard height at which trees are measured
during forest inventory.
ENSO El Niño Southern Oscillation, characterized by the Southern Oscillation
Index which is another index to characterize a sea surface temperature
anomaly in the equatorial Pacific.
GLM Generalized linear model, a class of statistical model that relies on a link
function to linearize the response variable.
GLMM Generalized linear mixed model, a class of statistical model like the GLM,
although with random (mixed) effects.
GV An abbreviation for green vegetation, an endmember in spectral mixture
analysis.
IDH Intermediate disturbance hypothesis, a hypothesis of species coexistence.
ITCZ Inter tropical convergence zone.
MCMC Markov Chain Monte Carlo, a method for fitting statistical models.
NPV An abbreviation for non-photosynthetic vegetation, an endmember used in
spectral mixture analysis.
∆NPV An abbreviation for the change in non-photosynthetic vegetation between
two time periods.
PDO Pacific Decadal Oscillation, an index that characterizes the sea surface
temperature anomaly in the north Pacific.
SOI Southern Oscillation index, an index used to characterize the strength of
the El Niño Southern Oscillation.
SRE Spatial random effect, used in spatial generalized mixed models.
SST Sea surface temperatures.
13
TM Thematic Mapper, a multispectral sensor on the Landsat 5 satellite
platform.
TSD Time since disturbance.
WRS Worldwide Reference System.
14
Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy
THE FREQUENCY OF CATASTROPHIC WIND DISTURBANCES IN THE NORTHWEST
AMAZON, AND THEIR IMPACT UPON TREE MORTALITY AND DIVERSITY
By
Sami Walid Rifai
May 2016
Chair: Stephanie A. Bohlman
Major: Forest Resources and Conservation
Catastrophic wind disturbances can cause aggregations of tree mortality, called
blowdowns, in the Amazon. I address three broad themes relating blowdown disturbances to
forest structure, diversity, and global climate. Clearly strong winds kill trees in the Amazon, but
it has not been documented whether it kills trees selectively or randomly. Therefore in Chapter 1
I ask, “which trees are most susceptible to die in a blowdown?” We found that large trees are the
most likely to die, which is relevant for quantifying landscape scale forest necromass from wind
disturbance because large trees carry disproportionately large amounts of total forest biomass.
The effect of the post blowdown disturbance on the successional dynamics tree diversity
has had only limited exploration. In Chapter 2, I examine the effect of blowdowns on tree
diversity. The Intermediate Disturbance and the Canopy Gap Partitioning hypotheses both
predict that disturbance promotes diversity. Results strongly showed that wind disturbance did
not promote diversity at any level of disturbance intensity or time since disturbance.
The representation of the global carbon cycle in climate models is hindered by
uncertainty regarding the strength of the Amazon carbon sink. Quantifying the frequency of
extreme forest disturbance events is one way to reduce this uncertainty. In Chapter 3, I address
the underlying causes of spatial and temporal variation in blowdown frequency. I found only
15
weak correlations with regional scale climate variation, yet blowdown frequency strongly
corresponded with the sea surface temperature anomaly indices of the Atlantic Multidecadal
Oscillation, the El Niño Southern Oscillation, and the Pacific Decadal Oscillation.
16
CHAPTER 1
INTRODUCTION
Large aggregations of wind thrown and broken trees, called blowdowns, are a common
form of natural disturbance in the northwest Amazon. Squall line storms and convection fed
storm systems produce downbursts with sufficient velocity to create blowdowns (Garstang et al.,
1998), that can span areas in excess of 30 km2 (Nelson et al., 1994). Despite their large size,
these disturbances are clustered in space (Fisher et al., 2008), which has resulted in optical
remote sensing imagery, such as Landsat, in becoming the easiest method to map these
disturbances (Nelson et al., 1994; Chambers et al. 2008; Espírito-Santo et al., 2010; Negrón-
Juarez et al., 2011). Yet the northwest Amazon also has an everwet climate, where much of the
region does not experience any month with less than 100 mm of rainfall (Sombroek 2001). This
high precipitation regime results in higher cloud cover, which translates to few usable Landsat
image acquisitions in wet years (Asner 2001). Thus despite their large size (Nelson et al., 1994),
and high potential to kill large quantities of trees (Negrón-Juarez et al., 2011) - these
disturbances have most often gone unmeasured. This knowledge gap of the natural disturbance
regime contributes to the uncertainty regarding the Amazon's carbon balance (Davidson et al.,
2012; Trumbore et al. 2015), but the lack of well quantified disturbance regimes also hinders the
representation of the global carbon cycle in Land Surface Model components of Earth System
Models (Running 2008).
Blowdown size is easily measurable from space when cloud-free imagery is available, yet
it is hard to tell who dies. The identity of killed trees is important for scaling the total necromass
(i.e., dead wood that is committed carbon emissions). The carbon emission contribution due to
blowdowns is greater if the largest trees are more vulnerable to die, because the largest
individuals often constitute the majority of Neotropical forest biomass (Nascimento and
17
Laurance, 2002; Rifai et al. 2015). For this reason, in Chapter Two I address how blowdowns
affect forest structure and cause differential mortality of adult trees. Disturbance has been
invoked to explain species coexistence by multiple hypotheses of biodiversity such as the
Intermediate Disturbance Hypothesis (Connell 1978) and the Canopy Gap Partitioning
Hypothesis (Denslow 1980), yet empirical evidence in tropical forests has been mixed with
evidence for (Molino and Sabatier, 2001; Bongers et al., 1999) and against these hypotheses
(Hubbell et al., 1999). In Chapter Two, I examine the role of wind disturbance in promoting
niche partitioning of canopy gaps by pioneer species, and how disturbance intensity and time
since disturbance affect tree species diversity in the dynamic and hyperdiverse northwest
Amazon forest. Finally, the frequency of blowdown events is poorly constrained over most of the
Amazon basin. There is evidence showing catastrophic wind disturbance can be drastically
higher in certain years (Negrón-Juarez et al., 2010), although appreciation of the interannual
variation in the Amazon’s wind disturbance regime has not been incorporated into basin wide
estimates of necromass attributed to wind disturbance (Espírito-Santo et al. 2014). In Chapter
Three, I quantify the spatial and temporal variation in blowdown activity, as well as investigate
the underlying climate variables that drive year to year variation in total blowdown activity.
18
CHAPTER 2
LANDSCAPE-SCALE CONSEQUENCES OF DIFFERENTIAL TREE MORTALITY FROM
CATASTROPHIC WIND DISTURBANCE IN THE AMAZON
Introduction
The magnitude of tree mortality occurring in a tropical forest can determine its
community dynamics, where periods of low mortality (“background”) promote gap phase
replacement, and a disturbance event resulting in high mortality can trigger a “reset” of the
community followed by colonization of pioneer species. Much of the published research on
tropical tree mortality has focused on background mortality, which is pragmatic considering that
tropical forest plot networks, a main source of forest dynamics data, are not well suited to sample
spatially aggregated disturbances (Fisher et al. 2008; Chambers et al. 2013). Catastrophic
disturbances (i.e., events with greater than 5% mortality as defined by Lugo and Scatena, 1996)
in tropical forests from wind have largely focused on hurricane-prone areas (Brokaw and Grear,
1991; Lugo and Scatena, 1996; Curran et al. 2008). However, the role of non-hurricane
catastrophic wind disturbance within tropical forests is increasingly recognized to alter both
species composition (Chambers et al. 2009b; Marra et al. 2014) and create sudden losses of live
biomass (Negrón-Juarez et al. 2010) which will decompose and be returned to the atmosphere as
carbon emissions. In the Amazon basin in particular, wind storms can create blowdowns with
spatial extents in excess of 30 km2 (Nelson et al. 1994). Amazonian blowdowns are thought to
emanate from two meteorological mechanisms: downbursts from convective systems, and squall
line storms (Garstang et al. 1998). A recent display of the widespread mass mortality induced by
a squall line storm was seen in 2005, where approximately 500 million trees were estimated to
have died (Negrón-Juarez et al. 2010). Yet not all trees die within a blowdown, so the identities
of which trees are killed has remained an open question until now. The individual variation in
biomass of adult trees can span orders of magnitude, so understanding which trees die is of
19
special importance for reducing the large uncertainties regarding disturbance and its role in the
Amazon’s large contribution to global land-atmosphere carbon exchange (Davidson et al. 2012).
Previous research to quantify tropical forest blowdowns has mostly focused on
identifying the shape of size frequency distributions of blowdowns (Fisher et al. 2008; Chambers
et al. 2009; Kellner and Asner 2009; Chambers et al. 2013; Asner et al. 2013). However, no
studies to date have quantified how mortality rates vary between trees within tropical forest
blowdowns, and how this differential mortality will affect estimates of landscape-scale carbon
losses. On the contrary, because this was unknown - previous efforts to estimate mortality and
necromass from wind disturbance events have had to model mortality as a random uniform
process (e.g., Negrón-Juarez et al. 2010). Accurately quantifying the effects of episodic wind
disturbances on community and ecosystem dynamics requires information not only on the size
frequency distribution of blowdowns, but also on tree mortality rates within blowdowns in
relation to the properties of individual trees (e.g., their size and functional traits, such as wood
density) and the physical environment (e.g., topographic position). A key concern related to
differential mortality (i.e., non-random, or selective mortality) is that large trees contain
disproportionate amounts of forest biomass (Clark and Clark, 1996). Therefore, if large trees
suffer disproportionately high mortality during blowdowns, ignoring differential mortality will
lead to underestimation of carbon lost to the atmosphere from the disturbance.
There are several reasons to expect differential tree mortality (i.e., dependence of
mortality probability on microsite or individual tree properties) in large blowdowns in the
tropics. First, background (non-catastrophic) tree mortality rates in tropical forests depend on the
characteristics of individual trees, such as high wood density (Chao et al. 2009; Kraft and Metz
et al. 2010) and − less consistently − large size (Clark and Clark, 1996; Laurance et al. 2000;
20
Thomas et al. 2013). Second, fire, drought, and hurricanes have all been shown to cause
differential mortality (Nepstad et al. 2007; Barlow et al. 2003; Curran et al. 2008). Third, studies
of wind disturbance in temperate forests have found that increasing tree size (Rich et al. 2007),
low wood density (Canham et al. 2001; Rich et al. 2007), and higher topographic position
(Sinton et al. 2000) promote probability of death. Finally, tree mortality from wind disturbance
has a high potential for spatial autocorrelation because the probability of a tree dying is not
independent from its neighbor’s fate. A large windthrown tree may directly kill or damage
neighbors as it falls, or indirectly affect neighboring tree mortality probability due to increased
wind exposure. Consequently, accurately predicting landscape mortality will likely require
accounting for spatial autocorrelation.
In this study, we consider three questions: 1) How much do landscape and tree structural
characteristics predispose a tree to die in an Amazon blowdown? 2) Does quantifying spatial
autocorrelation alter the estimated degree of differential mortality? 3) How does incorporation of
individual tree-based differential mortality and spatial autocorrelation alter necromass estimation
across a forest blowdown landscape?
Methods
Study Site
Using 30 m spatial resolution Landsat TM imagery, we located a blowdown that occurred
in late 2009 in the Peruvian Amazon (4.389° S, 73.602° W), roughly 100 km south of Iquitos in
the department of Loreto. The vegetation is terra firme forest and receives approximately 3100
mm rainfall per year (Sombroek, 2001). In this study area, elevation ranged from 130 m to 160
m, where higher elevation strongly corresponds to hill top topographic positions, intermediate
elevations correspond to slopes and lesser ridgelines, and the lower elevations correspond with
depressions and valleys.
21
Blowdown Identification and Satellite Image Analyses
Landsat 5 Thematic Mapper (TM) acquisitions with scene identifiers
“LT50060632009245CUB00” (September 2, 2008) and “LT50060632009341CUB00”
(December 7, 2009) were used to assess the pre- and post- disturbance landscape, respectively.
The images were atmospherically corrected using the ENVI Fast Line-of-sight Atmospheric
Analysis of Hypercubes (FLAASH) prior to analysis. Chambers et al. (2007) demonstrated that
forest wind disturbance can be detected and quantified from Landsat images due to a significant
increase in dead woody material, or non-photosynthetic vegetation (NPV) in the images, and the
change in NPV can be used as a metric of disturbance intensity. Multiple endmember spectral
mixture analysis (MESMA) (Roberts et al, 1998), a variant of linear spectral unmixing, was used
to estimate the pre- and post- disturbance NPV fraction of the blowdown landscape. For the
unmixing, we used three endmembers: NPV, green vegetation (GV), and shade. Soil, which is
often used as a fourth endmember in forested landscapes, was not used as an endmember because
while blowdown events produce a high fraction of windthrown trees, relatively little soil is
visible from above. Constrained reference endmember selection (Roberts et al. 1998) was used to
derive spectral libraries of the target endmembers from the post-disturbance images. Shade
normalization was applied to produce an image of just %NPV and %GV for each pixel. The pre-
disturbance %NPV data was then subtracted from the post- disturbance data to yield ΔNPV
image (Chambers et al. 2007; Negrón-Juarez et al. 2010 & 2011; Marra et al. 2014). A very
similar implementation of the ΔNPV method has been demonstrated to be sensitive to subpixel-
scale clustered treefalls composed of as few as six trees (Negrón-Juarez et al. 2011).
To generate landscape topographic variables, we downscaled the AST14DEM digital
elevation model (available on LP DAAC) from 30 m to 10 m spatial resolution through cubic-
convolution. Elevation and ΔNPV values were extracted for plot units from the three hectare
22
transect and the ΔNPV calibration plots (explained in section 2.2). Image analysis was performed
with ENVI (Exelis Visual Information Solutions, Boulder, Colorado) and the VIPER Tools
ENVI plugin (www.vipertools.org).
Field Data Collection
Three intersecting transects, consisting of 103 30 x 10 m contiguous plots (3.1 ha total
area), were installed inside and outside the blowdown affected area (Fig. 2-1; hereafter called
“transect data”). Disturbed and non-disturbed plots were chosen to have comparable topographic
characteristics. Elevation ranged between 132 m and 164 m across the entire transect, with mean
elevation of 144 m in the undisturbed portion of the transect and 152 m in the disturbed portion.
Disturbance within individual plots ranged from no disturbance-induced tree mortality to near
complete tree mortality. Diameter at breast height (DBH, 1.3m) was measured on all live and
dead individuals greater than 10 cm DBH. Diameters of some dead, fallen trees were measured
above 1.3 m from the base (breast height) because their boles were made inaccessible by other
windthrown trees.
Approximately 95% of live trees were identified to species, and no attempt was made to
identify dead trees. Dominant families in the transect plot were Lecythidaceae, Euphorbiaceae,
Sapotaceae, Myristicaceae, and Fabaceae (Table A-4). Wood density was sampled for one live
individual of each species using a gas powered drill (Tanaka TED-270PFR) by collecting all
wood chips produced by drilling into the tree from the cambium to the pith with an 18 mm wide
ship auger bit. The hole’s volume produced by drilling was measured with vernier calipers, and
the mass of the oven dried wood shavings was divided by the hole volume to determine wood
density (Francis, 1994). These wood density values were compared with matching species (when
available) that had replicated values from the Global Wood Density Database (Chave et al. 2009;
Zanne et al. 2009) to ensure the method was comparable to previously published measurements
23
(Fig. A-11). Wood density samples were also collected via chainsaw for all dead individuals.
Dead wood samples were submerged for three days to approximate their pre-death green volume,
after which the volume was determined using the water balance submersion method. Samples
were then oven-dried and weighed.
A separate but closely located set of thirty 20 x 20 m plots (hereafter referred to as the
“ΔNPV calibration plots”) was installed in the same blowdown to calibrate the remote sensing-
derived disturbance intensity metric (ΔNPV) to tree mortality. Unlike the transect plots, each
calibration plot was installed so its center would be as close as possible to the corresponding
post-disturbance Landsat 5 TM pixel (Fig. 2-2, inset). Each tree greater than 10 cm DBH with its
base in the plot was counted, and categorized as live or dead. Because our aim was to quantify
blowdown mortality, we only counted dead trees that were snapped or windthrown. Snapped but
resprouted trees, as well as live trees that were partially uprooted, were counted as being alive
despite being likely to die within a few years. Thus, the ΔNPV metric in this study only describes
immediate post-disturbance mortality, and does not incorporate delayed tree mortality resulting
from the disturbance.
Modeling Tree Mortality Rates
Here, we briefly introduce four different models used to quantify how both differential
mortality from the blowdown event and spatial autocorrelation affect estimates of landscape
necromass produced by the blowdown. Details of each model are presented in the subsequent
sections. Model 1 predicted the overall fraction of tree mortality within a pixel by regressing the
overall fraction of tree mortality solely against the pixel’s ΔNPV (Chambers et al. 2007). Model
2 predicted individual tree level probability of death from the blowdown event by fitting a
logistic regression of mortality probability as a function of tree size and wood density within
each of six categories of disturbance intensity (ΔNPV). Model 3 was similar to Model 2, but also
24
allowed for differential mortality in relation to topographic position, in addition to ΔNPV, tree
size and wood density. Model 4 was similar to Model 3, but also accounted for spatial
autocorrelation. Model 1 was fit twice, once using the transect data (also used to fit Models 2 -
4), but also fit with the ΔNPV calibration plot data to generate an estimate of the ΔNPV-
mortality relationship with reduced uncertainty in the ΔNPV predictor (see Fig. 2-2 inset).
Models 1 - 4 were fit to the transect data in a Bayesian modeling framework using either Markov
Chain Monte Carlo (MCMC) (for Models 1 - 3) or nested Laplace Approximation (for Model 4)
to estimate the model parameters. The joint posterior distribution of parameters for each model
was then used to simulate the landscape necromass produced by the blowdown while
propagating model uncertainty.
Model 1: pixel-scale mortality depends on ΔNPV
Following Chambers et al. (2007), tree mortality from the blowdown was estimated from
ΔNPV for each pixel using a simple linear regression (eq. 1):
% 𝑚𝑜𝑟𝑡𝑎𝑙𝑖𝑡𝑦 = 𝛽 ∗ 𝛥𝑁𝑃𝑉 + 𝜖 (2-1)
No intercept term was used to ensure that the predicted tree mortality was zero when ΔNPV = 0
as we were only trying to capture mortality related to disturbance (and thus ΔNPV) rather than
background mortality. Forcing the relationship through the origin was also necessary to reduce
bias in the lower portion of the ΔNPV and to constrain the over-prediction of necromass (section
2.5). For purposes of comparison of the relationship with an intercept, the model fit with
intercept is included in the Supplementary Material (Fig. A-12). As explained above, the model
was fit separately to each of two different datasets: the transect data and the ΔNPV calibration
plots. Each ΔNPV calibration plot was spatially located at the center of a Landsat pixel, whereas
the transect plots intercepted one to four Landsat pixels (Fig. 2-2). The lack of a one-to-one
correspondence between a transect plot and a Landsat pixel is expected to increase uncertainty in
25
the explanatory variable (ΔNPV); we explore the implications of this uncertainty in the
Discussion section. The posterior distribution of the slope and standard deviation were obtained
using MCMC methods implemented in the R-language ‘FilzbachR’ package (Lyutsarev and
Purves, 2011).
Model 2: individual-tree mortality depends on tree size and wood density within ΔNPV
classes
Model 2 assumed that the probability of mortality (PM) for an individual tree depended on
tree size (DBH) and wood density (WD) as:
PM = 1
1 + 𝑒−(𝛽0+ 𝛽1∗𝐷𝐵𝐻 + 𝛽2∗𝑊𝐷)
(2-2)
where β0 is the intercept term. The model was fit as a generalized linear mixed model (GLMM)
with intercept and slope parameters treated as random effects that varied across six groups of
transect plots partitioned by ΔNPV class (< 0, 0-0.04, 0.041-0.15, 0.151-0.4, 0.41-0.7, and >
0.71, with higher values indicating greater disturbance severity). The ranges of the ΔNPV groups
were set to more evenly distribute observations of live and dead trees. Including elevation did not
improve fit in Model 2, likely because elevation was correlated with ΔNPV (r = 0.468) (Fig A-
10). The GLMM was fit using the Gibbs Sampling MCMC package, JAGS (version 3.4.0) in
conjunction with the R package, “rjags” (Plummer 2014).
Model 3: Individual-tree mortality depends on ΔNPV, elevation, tree size, and wood density
Model 3 assumed that the probability of mortality (PM) for an individual tree depended on
DBH and WD, as well as a transect plot’s ΔNPV and topographic position (elevation), according
to a logistic function generalized linear model (GLM) (eq. 3):
Pm = 1
1 + 𝑒−(𝛽0 +𝛽1∗∆𝑁𝑃𝑉+𝛽2∗𝐷𝐵𝐻+𝛽3∗𝑊𝐷+𝛽4∗𝐸𝐿𝐸𝑉+𝛽5∗∆𝑁𝑃𝑉∗𝐷𝐵𝐻+ 𝛽6∗∆𝑁𝑃𝑉∗𝑊𝐷)
(2-3)
26
where β0 is the intercept, and ELEV is elevation. Eq. 3 was fit using the R-language ‘FilzbachR’
package. Model selection (i.e., inclusion/exclusion of certain interaction terms) is explained
below.
Model 4: Individual-tree mortality with spatial dependence
As explained in the introduction, tree mortality within a blowdown is an inherently
spatial process, because the survival of an individual tree partially depends on whether its
neighbors have been killed. To account for this spatial dependence, we used the stochastic partial
differential equation GLM in the ‘INLA’ package for the R-language (Rue et al. 2009; Lindgren
et al. 2011; Beguin et al. 2012). In this model (hereafter “spatial GLM”), a spatial random effect
(SRE; constrained to have a mean of 0) is used to model a continuous random spatial process.
The SRE term is an additive model term (for each spatial location) that accounts for spatial
dependence between observations, with the value of the SRE term varying among spatial
locations in a spatially autocorrelated manner (Beguin et al. 2012). The degree and scale of
autocorrelation is not prescribed, but rather is determined by the structure of the data as part of
the model fitting procedure. Model 4 has the same form as Model 3, but with the SRE term
included to account for spatially autocorrelated variation in mortality (eq. 4):
Pm = 1
1 + 𝑒−(𝛽0 +𝛽1∗∆𝑁𝑃𝑉+𝛽2∗𝐷𝐵𝐻+𝛽3∗𝑊𝐷+𝛽4∗𝐸𝐿𝐸𝑉+𝛽5∗∆𝑁𝑃𝑉∗𝐷𝐵𝐻+ 𝛽6∗∆𝑁𝑃𝑉∗𝑊𝐷+𝑆𝑅𝐸)
In addition to sharing a common form with Model 3, this model form was confirmed by a formal
selection procedure (see below).
Model selection and assessment
For the first three models, the selection of covariate combinations was guided by
Akaike’s Information Criteria, whereas the Deviance Information Criteria (Spiegelhalter et al.
2002) was used for the Spatial GLM. To compare the performance of the different models, we
27
calculated two goodness-of-fit indices for each model: the ratio of predicted to observed
necromass in the transect plots, and the coefficient of determination (R2) where R2 = 1 – (Sum of
SquaresResidual)/(Sum of SquaresTotal). R2 was calculated for individual trees and individual
transect plots. The pixel ΔNPV of 27% of the landscape encompassing the blowdown, and
15.5% of the transect subplots’ were less than 0 (Fig. A-8). Negative ΔNPV occurs when pixels
in the post-disturbance image exhibit less NPV than the corresponding pixels in the pre-
disturbance image. The linear pixel-scale ΔNPV model (Model 1) predicts negative necromass
when ΔNPV is negative. Therefore, for purposes of comparison across all models, only trees in
pixels with positive ΔNPV were used to calculate R2 for individual trees or by plot. Spatial
autocorrelation in individual tree model residuals (which would indicate a model’s failure to
capture spatial dependence patterns in mortality) was assessed with the Moran’s I statistic using
the correlog function in the ‘pgirmess’ R package (Giraudoux, 2013). Residuals were considered
to be spatially autocorrelated if Moran’s I was significantly different from zero (p < 0.05).
Landscape simulation of blowdown necromass
We used results from the four statistical models with a simple simulation model to
compare the estimated total landscape-scale necromass generated by the blowdown. For each
model, we combined the estimated parameters (and uncertainties; see below) with ΔNPV and
ASTER DEM values to simulate necromass in 30 × 30 m pixels across a 500.5 ha area (5,561
pixels) that includes the blowdown. In each pixel, we simulated the mortality (or survivorship) of
50 trees, which was the mean number of trees present in a pixel area as estimated from the non-
disturbed transect plots (ΔNPV < 0). The DBH and WD of these 50 trees per pixel were chosen
at random from trees inventoried in the non-disturbed transect plots. The aboveground biomass
of each tree was estimated from its DBH and wood density using an allometric equation
(equation 2.1 from Chave et al. 2005). For Model 1, the estimated mortality fraction of each
28
pixel was drawn from a normal distribution with a mean of the pixel ΔNPV multiplied by the
slope β, with standard deviation σ being drawn from its posterior distribution. For Models 2-4,
the posterior distributions from the model fits were used to parameterize the probability of death
functions (eq. 2 - 4) for each individual tree in the simulated landscape. The spatial GLM (Model
4) indicated that the spatial random effect increased with disturbance intensity (Fig. A-2), so the
simulation required predictions of SRE across the landscape, which were obtained by fitting the
SRE as a third degree polynomial predicted from the ΔNPV (see Fig. A-2 for details). To
propagate model uncertainty, 1000 sets of parameter values were drawn from each model’s joint
posterior distribution and used for each of 1000 simulations to generate a distribution of
landscape necromass predictions for each model. For Model 4, each of the 1000 parameter sets
included a random draw from the sampling distribution of the polynomial SRE model fit to
ΔNPV (Fig. A-2); these randomly generated SRE values accounted for parameter uncertainty in
the polynomial model, but did not account for residual variation in the polynomial model
because doing so would add a source of residual variation to the Model 4 simulations that was
absent in other model simulations.
Results
Overview
The probability of individual tree mortality in the blowdown increased with both tree size
and elevation, while it decreased with wood density (Fig. 2-3). The choice of model as assessed
by goodness of fit had a large impact on simulated landscape necromass. The linear ΔNPV
model (Model 1), which did not include differential tree mortality, most underpredicted the
amount of necromass in the transect data (Table 2-2), while the Spatial GLM (Model 4), which
included differential tree mortality and accounted for spatial autocorrelation in mortality, most
closely predicted the observed necromass in the transect data.
29
Plot-based tree mortality estimates
The plot-based model predicting the fraction of mortality linearly from ΔNPV using the
transect inventory data (Model 1) yielded a slope of 0.776 with 𝜖 ~ 𝑁(𝜇 = 0.04, 𝜎2 = 0.032), and
R2 = 0.632. When fit with the calibration plots, the slope was estimated to be 1.156 with
𝜖 ~ 𝑁(𝜇 = 0.05, 𝜎2 = 0.039), and R2=0.845 (Fig. 2-2). Both ΔNPV slope estimates are similar
with prior studies (slope of 0.996 in Chambers et al. 2007; 1.03 in Negrón-Juarez et al. 2010).
Differential tree mortality models
Models 2-4 all indicated that tree mortality was non-random, with large trees, low wood
density trees, and trees at higher elevations being the most likely to die (Fig. 2-3). When
probability of mortality was determined with the spatial model, the largest tree in the field data
(98.7 cm) was predicted to be up to five times more likely to have died than the trees with
minimum diameter measured in this study (10 cm DBH ) at high ΔNPV (Fig. 2-4). Trees with
the lowest wood density (~ 0.18 g/cm3) were up to four times more likely to die than trees with
the highest wood density (~ 1.1 g/cm3) at high levels of disturbance intensity. Model 2, the
ΔNPV-partitioned GLMM, generally showed effect size of DBH to increase with higher levels
ΔNPV (Table A-2). The interaction between DBH and ΔNPV shows that large diameter trees are
even more likely to die when disturbance intensity is high (Table 2-1; Fig. 2-3 & 2-4). The
interaction between WD and ΔNPV suggests that the effect of wood density towards promoting
survival is stronger as disturbance intensity increases (Table 2-1; Fig. 2-3 and 2-4; Fig. A-4 and
A-6). Large trees with low wood density (i.e. high DBH:WD ratio) are the most prone to die,
even in the lower quarter of the ΔNPV spectrum (Fig. A-5).
Spatial versus non-spatial GLM
The spatial GLM was the only model of those tested that reduced spatial autocorrelation,
as measured by Moran’s I, to a non-significant (p > 0.05) level across all tested distance classes
30
(30 - 1450 m) (Fig. A-1). The coefficient estimates for the covariates in the spatial GLM and the
non-spatial version were similar (Table 1), but 95% credible intervals were narrower for the non-
spatial GLM, as expected when spatial autocorrelation is ignored (Lichstein et al. 2002). The
probability of mortality of individual trees was generally overpredicted by the non-spatial
individual tree-based models (models 2 and 3) (Table 2-2; Fig. A-9).
Goodness of Fit of Model Necromass Predictions
When comparing cumulative predicted necromass from the different models, the linear
ΔNPV model (Model 1) most underpredicted the necromass (59.3% predicted:observed), while
the non-spatial GLM (Model 3) most over predicted necromass (111.2%) (Table 2-2). The
spatial GLM (Model 4) predicted 91.2% of the observed necromass, and the ΔNPV-partitioned
GLMM (Model 2) predicted 111.1% (Table 2-2). The linear ΔNPV model (Model 1) performed
best at predicting necromass when R2 was calculated by individual tree, while the Spatial GLM
(Model 4) performed best when R2 was calculated by plot (Table 2-2).
Landscape necromass estimation from simulation modeling
The linear ΔNPV model (Model 1) that used the transect data rather than the calibration
data predicted the least necromass across the landscape (Fig. 2-5). Compared to this model, the
ΔNPV-partitioned GLMM (Model 2), the spatial GLM (Model 4), and the non-spatial GLM
(Model 3) models predicted 67.9%, 51%, and 93.6% more necromass, respectively (Table A-3).
The linear ΔNPV model (Model 1) parameterized from the ΔNPV calibration plots predicted
31.5% more landscape-level necromass than the same model parameterized from the transect
data. The uncertainty of cumulative landscape necromass estimates varied considerably among
the different models (Fig. 2-5; Table A-3). The ΔNPV-partitioned GLMM (Model 2) and the
spatial GLM allocated considerably less necromass into pixels with negative ΔNPV than the
other models (Table A-1).
31
Discussion
Differential tree mortality resulting from catastrophic wind disturbance
We found that tree structural attributes that enhance survival during catastrophic wind
disturbance in the Peruvian Amazon are different than attributes commonly cited to promote
survival during periods of background mortality. For example, large Neotropical canopy
emergent trees have been observed to exhibit approximately half the annual mortality of the
landscape average (Thomas et al. 2013), while other observations have suggested increasing
diameter to be an inconsistent predictor of death under background mortality conditions (Chao et
al. 2008). In contrast, we observed drastically higher probabilities of mortality for large trees
from a catastrophic wind disturbance event. This contrast in predictive mortality of a trait
highlights the importance of differentiating periods of background mortality from episodes of
catastrophic mortality. Similarly, increased mortality has been reported for large trees near edges
from fragmented forests in the central Amazon (Laurance et al. 2000), where increased exposure
to wind has been suggested to be one of the root causes. This contrast in size effects between
background and episodic mortality is not unique to wind disturbance, as a recent study also
found a size mediated increase in mortality rates for large trees exposed to drought (Bennett et al.
2015). Forest simulators and ecosystem models may produce more accurate mortality predictions
if different trait-based mortality algorithms are implemented to decide which trees die under
background conditions vs. during catastrophic disturbance events. As in Neotropical forests,
blowdowns are also well documented in northern temperate forests (Canham and Loucks, 1984),
where the size dependent effect upon tree mortality has also been observed from storm
windthrow events in the Adirondack mountains of New York (Canham et al. 2001), a 236,000 ha
blowdown in the southern boreal forests of Minnesota, USA (Rich et al. 2007), and in tree
mortality from tornado impacted forests (Peterson, 2007). In our study, the probability of small
32
tree mortality only increased after larger trees had most likely died (Fig. 2-3a). This has also
been observed in the central Amazon during periods of elevated wind disturbance (Toledo et al.
2012), as well as in other studies covering periods of background levels of mortality (Clark et al.
1991; Chao et al. 2009).
Wood density has often been cited as a predictor of background tree mortality in tropical
forests (Chao et al. 2008), explaining ~31% of the interspecific variation in tree mortality rates
across a network of large forest inventory plots (Kraft and Metzger et al. 2010). Unlike tree size,
wood density likely promotes survival through both periods of background and catastrophic
mortality as wood density is highly correlated with strength (Niklas, 1992) – which is important
because branch-fall is a major source of damage for understory and sub-canopy trees (Clark and
Clark, 1991). Tree diameter, wood density, elevation and disturbance intensity interacted to alter
the probability of mortality within a blowdown. For example, the difference in probability of
death between low and high wood density trees is not as large as differences in mortality due to
size or elevation (Figs. 2-3 & 2-4), yet when wood density is coupled with tree size, the
differences in probability of death are at their most extreme (Fig. A-5), in that large trees with
low wood density are especially likely to die in a wind disturbance.
For this blowdown, individual trees at higher elevation were more prone to die (Table 2-
1; Figs. 2-3 & 2-4). The blowdown epicenter was located on a ridge, so this elevation effect on
mortality may not be translatable to wind disturbances with epicenters on floodplains or
depressions. Yet in a study from the central Amazon, approximately one quarter of the variation
in tree mortality across a time interval with multiple wind storms could be attributed to variation
in topography and soil fertility (Toledo et al. 2011). Potentially this could be due to increased
exposure to high wind speeds on ridges that remove large individuals (Everham and Brokaw,
33
1996), enhanced drought stress on ridges (Silva et al. 2013), or some combination of the two
factors.
At very high levels of disturbance intensity, individual differences in diameter and
elevation position become less important in differentiating which trees will die (Fig. 2-4). This
may be explained by two inter-related causes: 1) as disturbance intensity increases and the
fraction of trees that die approaches 1, mortality (by definition) becomes less discriminating, and
(2) as disturbance intensity increases, each tree is under greater risk of being crushed or damaged
by a neighboring tree. We cite two results to support the 2nd claim. First, the SRE term of the
spatial GLM, which captures the influence of neighboring trees on an individual’s probability of
mortality, increased the probability of mortality when ΔNPV > 0.065 and plateaued when ΔNPV
> 0.3 (Fig. A-2). Second, we found the difference in probability of mortality between high and
low WD trees increased with ΔNPV (Fig. 2-3), which is consistent with the idea that high WD
should promote survival from damage caused by falling neighbors.
Our estimated landscape necromass is most likely an underestimate of the actual longer-
term necromass produced by the blowdown because only the number of trees killed within one
year of the blowdown is estimated, while more died in the years following. Second, linear
modeling of the ΔNPV metric is not robust to detecting treefall events composed of fewer than
six trees (Negrón-Juarez et al. 2011), so many smaller treefall events were undetected. Finally,
coarse woody debris that results from mechanisms other than tree mortality, including crown
breakage and branchfall, likely represent a significant contribution to the true necromass pool
created in the blowdown. For example, Palace et al. (2008) found that coarse necromass was
underestimated by 45% when only derived from tree mortality rates.
34
Implications of spatial autocorrelation for estimating mortality effects and landscape-scale
necromass
Landscape necromass simulations using the spatial GLM predicted 51% more necromass
than the linear ΔNPV fit with random-mortality using the transect data, yet only 15% more
necromass than the random mortality linear ΔNPV fit using the calibration data. The difference
in necromass estimates was a direct result of the difference in the slope parameters between the
linear regression models, where the slope of the transect fit (0.78) was lower than the calibration
plot fit (1.16). This discrepancy was likely the result of the rectangular transect plots spanning
multiple ΔNPV pixels (thereby introducing error in the explanatory variable and reducing its
estimated effect size), whereas the square calibration plots were located within the center of a
single pixel. The ΔNPV error induced by crossing multiple pixels in the transect data is likely
what caused the reduced slope in the ΔNPV-mortality relationship fit with the transect data (Fig.
2-2). It is likely this misalignment also served to reduce the effect size of the ΔNPV term in the
differential mortality models, and reduced their overall ability to explain variation in mortality
between individuals.
The clarity of the linear mortality-ΔNPV relationship was also obscured by the study
region’s occasional complex topography. Slopes in excess of 45° (30 m scale) are not uncommon
in the study region, producing a complex pattern of illumination and shading on top of the
canopy, which in turn affects ΔNPV estimates. Two of the three largest model residuals in the
relationship from the ΔNPV calibration plots were from areas where the field measured sub-plot
slope was in excess of 35°. Despite the complex topography, the estimated slopes of the linear
mortality-ΔNPV relationships are fairly consistent with what has been estimated from studies in
other tropical forest landscapes (Negrón-Juarez et al. 2010 & 2011).
35
Whereas the overall heterogeneity of disturbed forest patches across the blowdown was
captured by mapping ΔNPV, addressing spatial dependence between pixels and between
individual tree-level observations was crucial to constraining the predictions for a tree’s
probability of death. The coefficients for DBH, wood density and elevation in the non-spatial and
spatial GLMs were generally similar (Table 2-1), yet the inclusion of the spatial random effect in
the spatial GLM better constrained the estimation of necromass in the lower range of the ΔNPV
spectrum. For example, the simulated average necromass per killed tree was similar between
both the non-spatial and the spatial GLMs (Fig. A-7), yet the non-spatial GLM simulation of
landscape disturbance estimated 26% more tree death and > 27% more necromass than the
spatial GLM. This discrepancy was due to the non-spatial GLM estimating more necromass in
the lower ΔNPV range of pixels (Table A-1 & Fig. A-9), whereas the SRE of the spatial GLM
increased survival when ΔNPV was under 0.065, and increased mortality when ΔNPV was
higher (Figs. A-2 & A-3).
It was essential that in scaling up necromass predictions from the plot to the landscape
level that mortality was not overpredicted in the lower portion of the disturbance intensity
spectrum because the majority of pixels in the landscape encompassing the blowdown have low
ΔNPV values (Fig. A-8). Because the spatial GLM model’s residuals were not spatially
autocorrelated, had the smallest error of predicted:observed necromass, did not overpredict
mortality in the low ΔNPV pixels, and produced the highest R2 for predicted necromass by
individual plot, we suggest that the spatial GLM provides the best model to estimate landscape
necromass.
Conclusions
We found that two factors - differential mortality and the spatial structure of mortality -
acted independently to affect total necromass on the landscape. Simple relationships relating tree
36
mortality to disturbance metrics in tropical forests can oversimplify the complex processes that
create important variation in tree mortality related to tree and landscape characteristics. Here we
show evidence of differential mortality within a northwestern Amazon forest affected by a
catastrophic wind disturbance event. While estimating landscape necromass with the linear
ΔNPV relationship coupled with random mortality is by far the easiest to implement, we
conservatively estimate that this method under-predicts total blowdown necromass by 15-51%
for the blowdown documented here. Probabilistic estimation of tree mortality is better
approached when underlying spatial processes are included in models of tree mortality. Failure to
account for spatial autocorrelation may produce undue confidence in parameter estimates, of
which even small differences can result in large estimation discrepancies when scaled to the
landscape. Whereas tree death from wind disturbance and differential probability of death have
sometimes been incorporated into local scale gap dynamic models, this has yet to be
incorporated into larger regional scale forest carbon models. Incorporating differential mortality
from wind disturbance is likely to augment estimates of forest biomass lost because of the strong
predisposition of large trees to die.
37
Table 2-1. β coefficients for fixed effects in a non-spatial Generalized Linear Model (GLM)
(Model 3), and a spatial GLM (Model 4). 2.5, 50, and 97.5% Bayesian credible
intervals are presented. ΔNPV – difference in non-photosynthetic vegetation before
and after disturbance event, DBH – diameter at breast height (cm), WD – wood
density (g oven dry/cm3 green volume), Elevation (m), ΔNPV:DBH – ΔNPV
interaction with diameter, ΔNPV:WD – ΔNPV interaction with wood density.
Non-Spatial GLM Spatial GLM
Effect 2.5% 50% 97.5% 2.5% 50% 97.5%
Intercept -15.784 -12.907 -9.626 -20.503 -11.617 -2.762
ΔNPV 0.882 3.404 6.029 -1.564 2.021 5.648
DBH -0.006 0.010 0.025 -0.007 0.012 0.031
WD -1.597 -0.446 0.789 -2.097 -0.633 0.824
Elevation 0.050 0.071 0.091 -0.001 0.059 0.118
ΔNPV:DBH 0.067 0.123 0.183 0.050 0.110 0.176
ΔNPV:WD -7.931 -4.272 -0.688 -7.494 -3.401 0.595
38
Table 2-2. Goodness of fit between model necromass predictions and observed necromass from transect plot data. The model number
is listed in parentheses. Traits included diameter at breast height (DBH) and wood density (WD). Q1-3 reference questions
from last paragraph of introduction.
Model General description Spatial Traits Question 𝑅𝑡𝑟𝑒𝑒2
𝑅𝑝𝑙𝑜𝑡2
𝑃𝑟𝑒𝑑.
𝑂𝑏𝑠.
Linear ΔNPV -
calibration data (1)
Predicts fraction of adult tree mortality from
blowdown; fit with ΔNPV calibration plot data No No baseline 0.45 0.57 0.88
Linear ΔNPV-
transect data (1)
Predicts fraction of adult tree mortality from
blowdown; fit with transect plot data No No baseline 0.38 0.47 0.59
ΔNPV partitioned
GLMM (2)
Hierarchical model that predicts individual
mortality using DBH & WD; partitioned by
disturbance intensity using ΔNPV (binned)
No Yes Q1 0.21 0.54 1.11
Non-spatial GLM (3) Predicts individual mortality using ΔNPV
(continuous), DBH, WD & Elevation No Yes Q1 0.29 0.55 1.11
Spatial GLM (4) Same as Non-spatial GLM, but accounts for
spatial dependence between observations Yes Yes Q2 & Q3 0.34 0.58 0.91
39
Figure 2-1. 2009 blowdown study location roughly 100-km south of Iquitos, Perú in the
department of Loreto. Top middle: The post-disturbance Landsat 5 TM (RGB:Bands
5,4,3) with yellow dotted line indicating the position of transect plots. Top right:
Variation in disturbance intensity indicated by different categories of ΔNPV (increase
in exposed wood detected by pre- and post-disturbance Landsat images). Bottom: a
photo from near the center of the blowdown, taken less than one year after the
disturbance event.
40
Figure 2-2. The fraction of observed adult tree mortality vs. the change in pre- and post-
disturbance non-photosynthetic disturbance (ΔNPV) with regressions through the
origin (Model 1). The dotted and dashed lines represent the 95% credible interval for
the calibration plots (green squares) and transect plots (closed blue circles).
Calibration plots were carefully located to be centered on the Landsat pixels, while
transect plots crossed multiple pixels.
41
Figure 2-3. Estimated probability of mortality from a catastrophic wind disturbance of a tree with
median tree characteristics, as predicted by the spatial logistic regression model with
the SRE (Model 4). Shaded regions represent 95% confidence interval as calculated
by the Delta Method (Jackson 2011). The median tree characteristics were 17.3 cm
DBH, 0.64 wood density, and 149 m elevation.
42
Figure 2-4. The ratio of estimated mortality probabilities between small and large (2.5 and 97.5%
quantiles of the transect data, respectively) for tree characteristics as a function of
disturbance intensity. Tree size (A), wood density (B) and elevation (C), illustrate the
interaction between disturbance intensity and tree/site characteristics. Large and small
values were: 55 and 10.3 cm DBH; 0.97 and 0.33 wood density; and 162 and 134 m
elevation. The model used was the spatial GLM (Model 4). The dashed line indicates
no difference in the mortality probability.
43
Figure 2-5. Total estimated tree mortality (left column) and corresponding necromass (right
column) from the simulated landscape using four different models. For model 1, the
dark line represents the simulation from the model fit from the transect data and the
dotted line represents the simulation generated from the model fit with the calibration
data. For all other models, only the transect data was used. Solid vertical lines
represent the 50% quantile, while dotted lines represent the 2.5 and 97.5% quantiles
from 1000 runs of each model.
44
CHAPTER 3
SUSTAINED REDUCTION OF TREE SPECIES RICHNESS IN NORTHWESTERN
AMAZONIAN FORESTS FOLLOWING CATASTROPHIC WIND DISTURBANCES
Introduction
Downbursts and squall lines create catastrophic forest disturbances in the Amazon basin
(Garstang et al. 1998) that can span orders of magnitude in size (Nelson et al. 1994). The
blowdowns created by these wind disturbances result in a heterogeneous mix of forest patches
with highly varied levels of tree mortality. These large canopy gap disturbances have been noted
to promote the occurrence of pioneer species that are otherwise rare in undisturbed forests
(Chambers et al. 2009; Marra et al. 2014). The role of disturbance in maintaining species
diversity of tropical forests has been invoked by multiple hypotheses of species coexistence. The
most famous of these is the Intermediate disturbance hypothesis (IDH) (Connell 1978), which
predicts that diversity will peak at intermediate disturbance frequencies, disturbance intensities,
and time since disturbances (TSD) (Sheil and Burslem, 2003) due to the delay of competitive
exclusion by environmental heterogeneity created by the disturbance. The Canopy gap
partitioning hypothesis (CGP; Denslow 1980) also predicts that environmental heterogeneity
(light, temperature, humidity) created by disturbance (canopy gap formation) facilitates the
maintenance of pioneer and other gap specialists in the landscape. Both the IDH and CGP
hypotheses rely on equalizing effects (Chesson, 2000) in the sense that a species that is
competitive in a canopy gap is less competitive against other species in shade, and vice versa.
Neutral theory (Hubbell 2001) has spurred an extensive body of research, yet it does not make
specific predictions about diversity-disturbance relationships, but it does downplay the
importance of niche partitioning in favor of the roles of dispersal limitation, functional
equivalence between individuals, and stochastic processes in structuring tropical forest
community assemblages. Hubbell et al. (1999) did not find evidence for differential colonization
45
by pioneer tree species in a study of 1284 canopy gaps on Barro Colorado Island in Panama as
predicted by the IDH or CGP. However, it is likely that the canopy gaps studied were of
insufficient size to facilitate colonization by the pioneer trees. Brokaw (1985) demonstrated a
marked increase in successful pioneer colonization at gap sizes > 500 m2, yet the dataset of
Hubbell et al. (1999) included only four gaps > 400 m2. The role of species with intermediate
light demand further obscures the role of canopy gaps in promoting diversity. The recruitment of
species with intermediate light demand is unclear because of the relatively few studies that have
quantified gap size colonization dynamics, most have focused on pioneers (e.g. Putz et al. 1983;
Brokaw 1985 & 1987; Rose 2000). This lack of well-defined species regeneration requirements
has been highlighted as a key hindrance to testing the IDH (Shea et al., 2004).
Despite several studies conducted on the diversity-disturbance relationships of tropical
forest, the role of disturbance in affecting forest diversity is still unclear. A meta-analysis of IDH
studies found support for the IDH predictions in 22 of 48 studies (Kershaw and Mallik, 2013)
although most of these were extra-tropical. Sheil (2001) found support for the IDH with
disturbance induced by tree poisoning in Uganda, Molino and Sabatier (2001) found strong
support for the IDH in forests of the Guiana Shield 10 years after gap creation from silviculture,
while Bongers et al (2009) only found marginal support for the IDH in Ghanaian tropical forests.
Species richness was higher in forest succession plots 10 years after a hurricane strike in
Nicaragua, than nearby forest plots that were not disturbed (Vandermeer et al., 2000).
Theoretical work has suggested that the pattern of a hump shaped curve along a gradient of
disturbance (frequency, intensity, TSD) can actually be explained by alternative mechanisms
other than the IDH such as density dependence, poorly determined life history attributes of the
component species, or fluctuating strength of competitive interactions (Shea et al. 2004). Connell
46
(1978) posits that disturbance interrupts competition which effectively delays competitive
exclusion. However, through microcosm experimentation on protists, Violle et al. (2010) found
strong evidence that interspecific competitive interactions increase along a gradient of increasing
disturbance intensity. In contrast, Neutral theory largely disregards differential competitive
interactions and has successfully predicted the rank abundance distribution of species at local
spatial scales. Hubbell et al. (1999) attribute their negation of the IDH to “a random-thinning
process in stems >1 cm DBH” in canopy gaps, and strong dispersal limitation that is common
among Neotropical trees. However, the strength of these neutral processes in determining species
composition has been suggested to give way to niche partitioning as canopy gap size/disturbance
increases (Chambers et al. 2009) – where pioneer colonization of large canopy gaps would be
considered an example of niche partitioning.
Clearly, environmental conditions in canopy gaps greatly differ from the understory of
forest patches with intact canopies (Denslow 1987). Light availability (Canham et al., 1990), soil
moisture (Vitousek and Denslow, 1986), and soil nutrients increase with gap size (Denslow et al.
1998), as do soil temperature (Denslow 1980), air temperature and vapor pressure deficit
(Fetcher et al. 1985). Environmental conditions of canopy gaps are sufficiently different that it
could be expected to filter the species composition within the gap, but canopy gaps create
complex light environments (Canham et al., 1990) and pioneer colonization is often unsuccessful
in small (< 500 m2) canopy gaps (Brokaw 1985; Hubbell et al., 1999). Moreover, spatially
extensive wind disturbances do not necessarily guarantee large pioneer colonization in the
Neotropics (Vandermeer et al., 2000), and the effect of extensive wind disturbance such as
hurricanes on promoting a more diverse species composition is not clear in general, with some
47
evidence for both positive (Vandermeer et al., 2000) and negligible (Brokaw and Walker, 1991)
effects.
Here we examine how catastrophic wind disturbance alters tree species diversity through
time since disturbance and across levels of disturbance intensity, and we specifically predict the
following:
1. Niche partitioning by pioneer species grows stronger with disturbance intensity.
2. Intermediate levels of wind disturbance promote local-scale diversity.
3. Diversity increases beyond pre-disturbance levels before the expected lifespan of the
dominant gap-colonizing pioneer species (Cecropia sciadophylla ~ 20 yrs; Zalamea et
al., 2008).
We use the northwest Amazon as our study region because it experiences a higher frequency of
blowdown events than what has been reported elsewhere in the Amazon Basin (Rifai et al., in
prep). The northwest Amazon is also ideal because the forests are hyperdiverse (Gentry 1988a &
1988b; ter Steege et al. 2006), the soils are comparatively more fertile than elsewhere in the
Basin (Quesada et al. 2011a), and it experiences an everwet climate (Sombroek 2001) – all of
which are thought to increase the dynamic rates of these forests (Malhi et al. 2006; Quesada et al.
2011b) and indirectly reduce wood density (Chave et al. 2006; ter Steege et al. 2006). For these
reasons, we expect the northwest Amazon to recover from disturbance faster than elsewhere in
the Basin.
Methods
Site selection by remote sensing, forest plot installation, inventory, and tree identification
Five blowdowns that formed between 1988 and 2009 were located in the Department of
Loreto, Peru with Landsat 5 & 7 imagery (Fig. 3-1) and sampled between the years 2009 - 2011.
The youngest blowdown occurred in 2009 (1 yr TSD) at the Nauta site, where terra firme forest
dominates the landscape. The next youngest blowdown site, Mishana, was disturbed in 2006 (6
yrs TSD) and was located in a seasonally flooded forest, near the Nanay River. The blowdown at
48
the Napo site occurred in 1998 (12 yrs TSD), and was located in a terra firme forest. The
Alpahuayo blowdown occurred between 1992-1993 (17 yrs TSD) and was located in the
Alpahuayo-Mishana Reserve in mixed terra firme & white sands (aka varillal) forest. The Orosa
blowdown occurred in 1988 (22 yrs TSD) in terra firme forests. The elevation for each subplot
was extracted from the freely available Shuttle Radar Topography Digital Elevation Model
resampled to 30 m (SRTMGL1_003). The topographic positions within each site varied, except
for Mishana. The blowdown epicenter was located on a ridge at Nauta (1 yr TSD), in a
floodplain at Mishana (6 yrs TSD), and in between ridges at Napo (12 yr TSD), Alpahuayo (17
yrs TSD), and Orosa (22 yrs TSD). The Mishana blowdown was comparatively less intense and
more dispersed than the other blowdowns (Fig. 3-2). Elevation was correlated with disturbance
intensity (see following section) at all of the blowdowns, except the Mishana site (Table 3-1).
Additional site characteristics are listed in Table 3-1. Forest inventory plots of 3.0 to 4.5 ha were
installed at each site in the form of straight-line or intersecting transect subplots (Fig. 3-2), in
which all trees > 10 cm diameter at breast height (DBH) were measured. Approximately 95% of
the trees at each site were identified to species or morphospecies; unidentified trees were
disregarded.
Classification into disturbance classes
The subplots from each blowdown were grouped into disturbance intensity classes of:
none, intermediate and high for most sites based on remote sensing sensed disturbance intensity
metrics, and none, peripheral, and interior for one site (Alpahuayo) as described below. Subplots
were also divided into two broad elevation classes: ridge and valley, using the median elevation
from each site to specify the dividing line, except Mishana which did not exhibit sufficient
elevational change to warrant subdivision. Disturbance intensity was assessed at all but
Alpahuayo with the ∆NPV metric, which corresponds linearly to the fraction of adult tree
49
mortality (Chambers et al., 2007; Negrón-Juarez et al., 2011; Rifai et al. in review). In brief, the
pixel fractions of green vegetation and non-photosynthetic vegetation (NPV) were extracted
using linear spectral unmixing (Roberts et al., 1993) from Landsat imagery at each site before
and after the disturbance. The NPV fraction from the pre-disturbance image was then subtracted
from the post-disturbance image to yield the ∆NPV metric (Fig. 3-2). Further information
regarding the method is detailed in Negrón-Juarez et al. (2011). The corresponding Landsat
images used in the analysis are listed in Table A-1. We did not calculate the ∆NPV metric for
Alpahuayo because no cloud-free Landsat imagery was available for the two years following
disturbance, and the NPV signal attenuates to pre-disturbance levels within 1-2 years following
the disturbance. However, the new growth following a blowdown creates a spectral signature
that is distinct from older growth forests. So while estimation of tree mortality at Alpahuayo was
not possible, we were able to delineate the boundaries of the blowdown because of the distinct
spectral signature (higher reflectance in Band 4) created by the young vegetation. The
disturbance intensity at Alpahuayo was then categorized as ‘none’ if it was outside of the
blowdown, ‘peripheral’ if it was within 100 m of the periphery of the blowdown, and ‘inside’ if
inside the blowdown. The ∆NPV thresholds for designating disturbance levels at the other sites
for the disturbance classes ‘none’, ‘intermediate, and ‘high’ were <0.05, 0.05-0.3, and >0.3
which correspond to adult tree mortality rates of <5%, 5-35%, and >35% when using the ΔNPV-
mortality relationship derived at the Nauta blowdown site (Rifai et al. in review).
Quantifying niche partitioning with disturbance intensity
To test the prediction that niche processes strengthen with disturbance intensity, we used
the colonization of highly gap-specialized pioneer species as a proxy measure for the strength of
niche partitioning by pioneer species. We selected the dominant pioneer genera of the region:
Cecropia, Pourouma, Miconia, and Vismia (Pennington et al. 2004), that are not considered
50
long-lived pioneers, like Carapa, Goupia, or Parkia (Pennington et al. 2004), which might have
colonized and persisted from disturbances prior to the blowdown events. We calculated the
relative abundance of pioneers per subplot. However, pioneer stem density declines through time
as colonizers close the canopy (Brokaw 1985). To compensate for this, we also calculated the
fractional contribution of pioneers to total sub-plot basal area. These measures were regressed
against ΔNPV with generalized linear models either using a Gaussian (for both relative
abundance and basal area) or binomial (for relative abundance) link function. We used the Nauta
data (1 yr TSD) to estimate the removal of pioneers by disturbance. We used the inventory data
from the older blowdowns (Mishana, Napo, Alpahuayo, and Orosa) to estimate pioneer
colonization after disturbance. In the case of Alpahuayo where we lacked ΔNPV, we used an
ANOVA with Tukey’s range test to examine differences in pioneers across the disturbance
classes (outside, peripheral, and inside blowdown).
Diversity accumulation curves
We used species accumulation curves to examine the effect of disturbance intensity and
time since disturbance to test the prediction that diversity peaks at intermediate levels of
disturbance, and before the presumed end of Cecropia sciadopyhlla’s approximately 20 year
lifespan (Zalamea et al., 2008). We rarefied diversity curves for species richness (q = 0; as
explained below) and evenness (q = 2) by sample coverage as opposed to individual or sampling
unit, as explained below, and as suggested by Chao and Jost (2012). Sample coverage can be
defined as the fraction of individuals within a community that are observed within a sample
taken from the community (Chao and Lee, 1992; Chao and Jost, 2012). Rarefying by sample
coverage allows for a less biased comparison of communities of different inherent richness, and
for better comparison of communities where the asymptotic limit of the diversity accumulation
curve has not been reached (Chao and Jost, 2012). This approach is ideal for comparing
51
undersampled, hyperdiverse communities because the accumulation curve can be extrapolated by
up to half of the empirical sample with little bias (Chao et al. 2014). Hill numbers (Hill 1973)
represent a unified spectrum of diversity to entropy (Jost, 2006), where a Hill number of q = 0 is
the effective number of species (species richness), whereas q = 2 is equivalent to the inverse of
Simpson’s concentration index, which we hereafter refer to as evenness (Hill 1973; Chao et al.
2014). To allow future comparison to other studies, diversity curves were also rarefied by
individuals. The results of this analysis are provided in the supplementary information.
Confidence intervals were produced through 500 bootstrap simulations using the ‘iNEXT’
package (Hsieh et al. 2015) for the R language. The diversity accumulation curve of each group
was interpolated to the limit of its sample size. To further allow comparison between unequal
sample sizes of different disturbance classes, each accumulated curve was extrapolated by up to
50% of the empirical sample. Preliminary results from a clustering analysis showed that
elevation differences significantly structured community assemblages (Rifai et al. unpublished),
so sub-plots were also divided into either ‘valley’ or ‘ridge’ classes based on the median
elevation of subplots within each site. Subdivision at the Mishana site was unnecessary because
elevation differences were small within the floodplain and the clustering analysis did not indicate
community assemblage varied with the limited range of elevation.
Results
Niche partitioning by pioneer species with disturbance intensity
The relationship between pioneer abundance and disturbance intensity varied greatly
among sites. The recently disturbed Nauta blowdown (1 yr TSD) showed large reductions in
pioneer basal area and relative abundance with increasing disturbance intensity (∆NPV) (p <
0.001; Fig. B-1 & B-2) with a steep decline in the pioneer stem fraction when ∆NPV exceeded
0.2 (Fig. B-1). However, the effect of ∆NPV on reducing either pioneers was non-significant
52
when fit with only the subplots considered either outside or intermediate to the blowdown (p =
0.273; subplot ∆NPV < 0.3), which may indicate that lower levels of disturbance intensities are
insufficient to remove pioneer species. Disturbance intensity (∆NPV) had no effect on pioneer
basal area or relative abundance at the Mishana blowdown (6 yrs TSD) (p > 0.5; Fig. B-3),
whereas pioneer abundance and basal area increased with ∆NPV at the Napo blowdown (12 yrs
TSD) (p < 0.03; Fig. B-4). The subplots from the blowdown interior at the Alpahuayo site (18
yrs TSD) had more pioneers than the peripheral or outside the blowdown sub-plots (p < 0.002;
Fig B-5), yet the difference between the peripheral and the outside sub-plots was non-significant
(p > 0.44; Fig. B-5). The fraction of total basal area occupied by pioneer species and the relative
abundance of pioneer species increased with ∆NPV at the Orosa (22 yrs TSD sites (p < 0.001;
Fig. 3-3).
Reductions in species richness and evenness with disturbance intensity and time
We did not find evidence of wind disturbance promoting local-scale diversity, or
diversity increasing beyond pre-disturbance levels within the 22 yr time frame. Disturbed
subplots had reduced species richness at all blowdown sites (Fig. 3-4). But, time since
disturbance did not promote species richness or evenness within the range (1 - 22 yrs; Fig. 3-4)
sampled. Species richness rarefied by sample coverage does not decline when individuals are
removed at random, yet the forest that experienced high disturbance intensity at the Nauta site (1
yr TSD) had lower species richness than the intermediate or undisturbed forest, which suggests
species filtering from the disturbance. This is further evidenced by the lower pioneer abundance
in the high disturbance sub-plots (Figs. B-1 & B-2). Previous investigation at the Nauta
blowdown concluded mortality is higher for species of low wood density and/or large size (Rifai
et al. in review).
53
Surprisingly, there was a reduction of species richness in high disturbance parts of the
Mishana (6 yrs TSD), Alpahuayo (17 yrs TSD), and Orosa (22 yrs TSD) blowdowns. The
sustained reduction of species richness through 22 yrs TSD refutes our prediction that diversity
increases within the 20 yr time span after disturbance. The intermediate disturbance intensity
area within each blowdown site generally did not reach higher species richness than areas outside
of the blowdown with corresponding topographic position (Fig. 3-4), which also nullifies our
prediction that diversity would increase under intermediate disturbance intensity. One notable
anomaly was that the species richness of the high disturbance area of the Napo blowdown (12 yrs
TSD) was similar to that from outside the blowdown, which may have been due to the unusually
high presence of Cecropia outside the blowdown. This may suggest the “no-disturbance” sub-
plots at Napo were recovering from a previous disturbance within the lifespan of Cecropia (e.g.,
20 years), although we were unable to find evidence of a large disturbance after 1985 using the
Landsat 5 record. The consequences of disturbance for evenness were generally consistent with
the reductions in species richness across the blowdown sites (Fig. 3-5). Differences between the
high disturbance intensity, intermediate, and undisturbed forest were not significant at Nauta (1
yr TSD) (Fig. 3-4A), and the ridge sites at Napo exhibited greater evenness than the valley sites,
but disturbance did not have a consistent effect on either topographic position (Fig. 3-4C).
Otherwise, disturbance consistently reduced evenness at the Mishana, Alpahuayo, and Orosa
sites (Fig. 3-4B, D & E).
Discussion
We examined the response of niche partitioning and diversity to canopy gap disturbances
of far greater size than previously tested. We did not find evidence that disturbance promoted
diversity when disturbance was measured either spatially (disturbance intensity) or temporally
(time since disturbance). Rather, disturbance consistently reduced species richness and most
54
often evenness as well. While disturbance intensity generally increased the abundance and basal
area of pioneer species, the increase in pioneers was highly variable among the four oldest plots
that had sufficient time for pioneer recruitment to occur. Surprisingly, pioneers were frequently
observed in forest patches that had not been disturbed by the blowdowns (Figs. 3 & S1-S5),
suggesting that non-catastrophic canopy gap dynamics were sufficient to maintain pioneer
species diversity in the landscape. In the plot with the shortest post disturbance time (1 year), the
presence of pioneer species strongly decreased with disturbance intensity.
Wind disturbance, differential mortality and niche partitioning by pioneer species
The low shade tolerance of pioneer species in tropical forests is thought to result in their
absence from late successional forests (Connell 1978; Finegan 1996), with the exception of
certain long-lived pioneer genera than can become canopy emergents such as Ceiba and Parkia.
Contrary to this viewpoint, we found many pioneers in undisturbed sub-plots adjacent to all of
the blowdowns (Figs. 3 & S1-S5). The continued presence of pioneers in undisturbed forest
suggests that the background rate of canopy gap formation is sufficiently high to maintain the
presence of pioneers in the landscape. Quesada et al. (2012) hypothesized that higher tree
turnover in the northwest Amazon may be attributable to the higher fertility soils promoting
rapid growth life history strategies with less structural investment. The reduction of pioneer trees
observed at the 1 yr TSD blowdown disturbance (Fig. B-1 & B-2), and the selective mortality
associated with large size and low wood density (Rifai et al. in review) may explain the rapid
reduction of species richness in the intermediate and blowdown areas at Nauta (Fig. 3-4).
Further, pioneer tree colonization at the four oldest blowdowns was highly inconsistent where
many subplots with ∆NPV indicative of greater than 50% mortality were not successfully
colonized by the dominant pioneer genera (Figs. 3-3 and B-3, B-4, B-5), which was likely due to
rapid canopy closure by the advanced regeneration of surviving trees and lianas.
55
Relation to Canopy Gap Partitioning and the Intermediate Disturbance Hypotheses
Our results do not support either the Intermediate Disturbance or Canopy Gap
Partitioning hypotheses. We acknowledge the IDH, as proposed by Connell (1978), also predicts
increased species richness in areas with intermediate levels of disturbance frequency - which we
did not test, nor would be easily accomplished in a tropical forest setting. It could also be argued
we did not see diversity peak with disturbance because the ‘undisturbed’ forest patches were
already experiencing sufficient disturbance to allow for peak diversity. However, if failures of
the IDH or CGP are conditioned upon inappropriate baseline conditions it becomes difficult to
conceive of how the IDH and CGP could be tested or negated. Rather than heed a recent call to
outright abandon the IDH (Fox 2013), we suggest our results may highlight some mechanistic
shortcomings of IDH and CGP for application in hyperdiverse forest ecosystems.
We suggest the subset of species that require large canopy gaps to grow is small
compared to the total regional species pool. Many species of tropical trees were assumed to
require at least a small canopy gap to germinate when the CGP was proposed (Denslow 1980a &
1980b). Late successional community assemblages of tropical forests in this region are species
rich (Gentry 1988b) and the majority are thought to be dispersal limited (Hubbell 2001), whereas
the dominant Neotropical pioneer genera are small seeded and well dispersed by fruit bats and
birds in comparison to most shade tolerant species (Turner 2001; Lobova et al., 2003). After
disturbance, pioneer seeds are often already abundant in the seed bank (Putz 1983; Dalling et al.
1998) and rapidly colonize the pit and mound topography left by windthrown trees (Putz 1983).
We hypothesize that pioneers in undisturbed northwest Amazon forests persist through a storage
effect (Warner and Chesson, 1985), where occasional non-catastrophic canopy gap creation
creates the necessary conditions for pioneers to colonize, reach sexual maturity, and continue
seeding the landscape for subsequent colonization of canopy gaps. So while hundreds of species
56
can be removed from the canopy by catastrophic wind disturbance, only a handful of pioneer
species are adapted to exploit the conditions created by the disturbance, and temporarily
outcompete the shade tolerant but slower growing species for canopy dominance. Consequently,
wind disturbance does not increase tree species diversity at the local scale because pioneer
species are frequently already present, often dominate the newly opened spaces, and because
wind disturbances act as a filter to reduce the species pool of survivors by differential mortality
with respect to the traits of wood density and large size (Rifai et al. in review).
Our results suggest that the common practice of using pioneers as a quantitative measure
of previous disturbance history may be an unreliable method for quantifying actual forest
disturbance (i.e., killed trees and the reduction of live aboveground biomass). Some tropical
forest studies that found support for the IDH in tropical forests measured disturbance through the
presence of pioneer species (Molino and Sabatier, 2001; Bongers et al. 2009) or have assumed a
successional sequence of forest recovering from clearing (Shiel 2001; Shiel and Bongers, 2003).
In contrast, we quantified disturbance by the change in the amount of visible non-photosynthetic
vegetation, which is an effective predictor of adult tree mortality (Chambers et al. 2007; Negrón-
Juarez et al. 2011; Rifai et al. in review). Preferentially sampling the neighborhood of forest
patches where pioneers are present (presumably disturbed patches) may confound the diversity
response if these ‘disturbed’ plots are being compared to ‘undisturbed’ samples that are
preferentially collected from areas that are presumed to be undisturbed because pioneers are
absent. Pioneer colonization following disturbance can be highly irregular (e.g., the Mishana
blowdown, Fig. B-3). An alternative example is found in the post-hurricane disturbance forest
succession plots of Vandermeer et al. (2000), where pioneer colonization was minimal. This was
attributed to the disturbance being both intense and spatially extensive enough to have removed
57
the pioneer propagule supply (Vandermeer et al., 2000), which is at odds with other evidence
suggesting pioneer species are well represented in the seed bank (Putz 1983; Dalling et al. 1998).
In regards to pioneer trees, “the absence of evidence is not the evidence of absence”, that no
disturbance has occurred.
An inconsistent forest succession trajectory
A mounting body of evidence shows patterns of Neotropical forest succession to be
dependent upon the source of initial disturbance type (Mesquita et al. 2001), and generally less
predictable (Norden et al. 2015) than described in successional theory (Finegan, 1996). The
reduction of species richness with increasing disturbance intensity was substantial at the Mishana
blowdown site (6 yrs TSD), where more than a third of the species pool was still missing from
the highly disturbed forest patches. But unlike other blowdown sites, pioneers were absent. We
suggest two potential causes for this unexpected successional pattern. First, even though
Cecropia is well known to colonize river banks and floodplains, these locations are different than
the darker interior of seasonally flooded Neotropical forests. Also, liana density at the Mishana
site may affect post-disturbance regeneration. We did not quantify liana density, but we observed
the Mishana site to host a more lianas than any of the other blowdown sites. Because the liana
density was also high in the surrounding forest outside the blowdown, it is likely that liana
density in the canopy was very high before the blowdown occurred. We speculate that the large
majority of pre-disturbance lianas would have survived and continued to generate enough shade
to suppress post-disturbance pioneer colonization. Second, the Mishana blowdown occurred
between August 6 and August 22, 2005. The dry season precipitation in 2005 was average in the
study region, while both the wet and dry seasons of 2006 were anomalously wet with 1,100 mm
more rain than average (precipitation data from the Tropical Rainfall Measuring Mission). The
unusually high rainfall of 2006 likely led to a greatly extended period of forest inundation, which
58
may have prevented successful colonization by pioneer trees or others that did not successfully
extend their crowns above the water line. Seven of the ten most common (and perhaps most
competitive) species in the most disturbed parts of the blowdown (ΔNPV > 0.3) were also among
the ten most common species outside of the blowdown (ΔNPV < 0.05), which does not indicate a
strong compositional shift. The decline in evenness could be attributed to a reduction of less
common species (Fig. 3-5). This blowdown did not promote pioneer recruitment, it reduced
species richness and evenness, and did not cause notable differences in the most common species
between the blowdown and regions adjacent.
Remaining questions of disturbance and diversity through time in a changing climate
The longer-term legacy of catastrophic wind disturbances in high diversity tropical forest
remains unknown. We were limited in this study to examination of the consequences of
catastrophic wind disturbance up to 22 years. With this limited timescale, we were unable to
determine the time needed for diversity to recover to pre-disturbance levels. A prolonged period
of decreased species richness after mega-disturbance may carry implications for management
and conservation as the pace of climate change increases and forest degradation becomes more
widespread. Forests across the tropics are rapidly being fragmented by agriculture and
development, which has made trees in forest fragments more susceptible to wind disturbance
(Laurance et al. 1998). Convection-fed storm cells are thought to be the primary source of
blowdown causing downbursts (Garstang et al. 1998). Observational evidence has shown an
increase in rainfall across the northwest Amazon (Gloor et al., 2013), and an increase in
convective activity across the tropics in the 21’st Century (Tan et al. 2015), which may suggest
increasing wind disturbance in the tropics. Climate forecasts from general circulation models
regarding precipitation extremes are highly uncertain (Rummukainen, 2012), yet analysis of an
ensemble of CMIP5 models under “business as usual” carbon emissions scenarios predicts
59
increases in extreme precipitation events, especially in the tropics (Roman et al. 2015). There is
also non-meteorological evidence of an increasing wind disturbance regime, where analysis from
28 years of Landsat imagery in the northwest Amazon has shown a significant increasing
temporal trend in the amount of forest being disturbed by blowdowns (Rifai et al. in prep). The
growing body of evidence showing an increase in anthropogenic and natural disturbance regimes
of tropical forests, combined with decreased tree diversity following disturbance shown here,
should warrant concern for managing and maintaining Neotropical forest biodiversity and carbon
stocks.
Conclusions
Here we present evidence that blowdowns do not promote diversity at either intermediate
or high levels of disturbance intensity at the local scale, nor within the 22 year time-since-
disturbance time span of this study, in addition to three key points. First, niche partitioning by
pioneers after wind disturbance is highly variable even in patches of high tree mortality. Second,
reduction of tree species richness by wind disturbance is not always explained by large increases
of a relatively few pioneer species. Instead, wind disturbance may filter the species pool through
differential tree mortality, with some species more likely to die than others. Finally, we suggest
that neither the IDH nor the CGP hypotheses are applicable to Neotropical forest recovery from
wind disturbance because when many individuals are removed from a hyperdiverse forest where
most species are represented by few individuals, many species are also likely to be removed. The
subsequent colonization of spaces opened by the killed trees is dominated by advanced
regeneration and relatively few pioneer species, which persist for decades and delay the recovery
of diversity to pre-disturbance levels.
60
Table 3-1. Site descriptions. TF is terra firme, FP is floodplain forest, and WS is white sands
forest. Cor(Elev, ΔNPV) – correlation between elevation and disturbance ΔNPV.
Site Name Time since
disturbance
(years)
Blowdown
size (ha)
Forest
type
Elevation
range (m)
Cor(Elev,ΔNPV) Total Tree
Species
Richness
Nauta 1 300 TF 134 - 166 0.41 424
Mishana 6 500 FP 97 - 115 0.045 429
Napo 12 300 TF 111 - 137 -0.69 534
Alpahuayo 17 50 TF &
WS
114 - 144 -0.46 639
Orosa 22 800 TF 100 - 123 -0.60 547
61
Figure 3-1. Blowdown site locations in Loreto, Peru.
62
Figure 3-2. ΔNPV derived from pre- and post-disturbance Landsat 5 images (30 × 30 m2
resolution) for blowdown sites. Nauta (A; 1 yr TSD), Mishana (B; 6 yrs TSD), Napo
(C, 12 yrs TSD), and Orosa (D; 22 yrs TSD). White pixels represent areas masked
from analysis.
63
Figure 3-3. Pioneer tree occurrence with disturbance at the Orosa blowdown (22 yrs time since
disturbance; TSD). Ratio of pioneer basal area to total sub-plot basal area (A) and (B)
the relative abundance of pioneers (B) versus subplot ΔNPV. The 95% confidence
interval is shown in grey.
64
Figure 3-4. Species accumulation curves (Hill number: q = 0) rarefied by sample coverage for
blowdown sites. (A) Nauta, (B) Mishana, (C) Napo, (D) Alpahuayo, (E) Orosa. Time
since disturbance (TSD) is given at the top of each panel. Shaded regions represent
the 95% confidence interval from 500 bootstrapped simulations. Solid lines represent
interpolation, and dashed lines show extrapolations.
65
Figure 3-5. Simpson’s evenness accumulation curves rarefied by sample coverage for blowdown
sites. (A) Nauta, (B) Mishana, (C) Napo, (D) Alpahuayo, (E) Orosa. Time since
disturbance (TSD) is given near the top of each panel. Shaded regions represent the
95% confidence interval from 500 bootstrapped simulations. Solid lines represent
interpolation, and dashed lines show extrapolations.
66
CHAPTER 4
SEA SURFACE TEMPERATURE ANOMALIES AS THE UNDERLYING CAUSE OF
INTERANNUAL VARIATION OF CATASTROPHIC WIND DISTURBANCE IN THE
NORTHWEST AMAZON
Introduction
Wind disturbances that cause blowdowns affect tropical forest biomass (Negrón-Juarez et
al. 2011; Rifai et al. in review) and light availability (Canham et al. 1990), the size distribution of
trees (Farrior et al. 2016), biodiversity (Chambers et al. 2009; Marra et al. 2014), and regional
carbon balance (Negrón-Juarez et al. 2011; Chambers et al. 2013). Blowdowns catalyze many
important ecological processes within tropical forests such as tree mortality (Rifai et al. in press),
stand regeneration (Rifai et al. in prep), and species co-occurrence through patch dynamics
(Chambers et al. 2013). Convection fed storm cells and squall line storms in the Amazon
occasionally produce downbursts with high enough speeds to cause blowdowns (Garstang et al.
1998), and the tree mortality and loss of biomass caused by these storms can be very large. For
example, in 2005 a single squall line storm produced blowdowns dispersed across much of the
Amazon Basin and is estimated to have killed more than 500 million trees (Negrón-Juarez et al.,
2010).
Despite the importance of blowdowns to the structural dynamics of forests, there is no
consensus regarding Amazon blowdown frequency, the degree to which blowdowns are
clustered in space, the size distribution of blowdowns, or why blowdowns seem more common in
some years. Disagreements over the size distribution of blowdowns and how they are clustered
(Fisher et al. 2008; Lloyd et al. 2009; Chambers et al. 2009; Espírito-Santo et al. 2010; Chambers
et al. 2013; Espírito-Santo et al. 2014) has spurred debate whether the inferred Amazon carbon
sink (Phillips et al. 1998) is real, or whether it is a sampling artifact brought about by an
insufficient quantity of forest plots (Fisher et al. 2008; Chambers et al. 2013; Meyer et al. 2013;
67
Di Vittorio et al. 2014; but see Gloor et al. 2009) and/or the bias induced by the haphazard
spatial configuration of the plot network (Marvin et al. 2014). Numerical analyses have also
demonstrated the plot network to be a suboptimal data source for the estimation of canopy gap
disturbances larger than 0.1 ha (Fisher et al. 2008; Chambers et al. 2013; di Vittorio et al. 2014).
These knowledge gaps in the frequency of moderate to large scale canopy gap disturbances
contribute to the overall uncertainty of the carbon balance of the Amazon Basin (di Vittorio et al.
2014).
The Amazon Basin comprises 5.5 million km2 of tropical forest, yet nearly all of the data
used to estimate the frequency of wind disturbance originated from the central portion (Nelson et
al. 1994; Espírito-Santo et al. 2014), even though meteorological data suggests that the northwest
Amazon has a higher occurrence of wind disturbance (Garstang et al. 1998). Geographic
variation in the wind disturbance regime may also contribute to spatial gradients in forest
structure. For example, the shorter biomass turnover times (Malhi et al. 2006; Aragão et al.,
2009; Quesada et al., 2012) and lower wood density (Baker et al. 2004) in the northwest Amazon
have been attributed to cross-basin differences in soil fertility (Quesada et al. 2012). However, it
is also possible that wind disturbance may contribute to cross-basin differences in biomass
turnover, but this has not been investigated. Here we quantify the natural disturbance regime of a
large area of the northwest Amazon over a 28 year time series to understand how it has changed
through time and in relation to changing weather patterns. Quantifying the wind disturbance
regime may also contribute to an understanding of why forest turnover rates vary across the
basin (Quesada et al., 2012), and may help reconcile contrasting views on whether or not
biomass gains on inventory plots are a sampling artifact (Fisher et al. 2008; Gloor et al. 2009;
68
Chambers et al. 2013; Vittorio et al. 2014) or if these gains indeed indicate a widespread increase
in growth (Phillips et al. 1998; Gloor et al. 2009).
Large blowdowns can be highly clustered in space, which requires observations of large
areas to accurately estimate the proportion of the landscape being disturbed. Optical remote
sensing has proven effective for mapping blowdowns because of its continuous coverage over
large areas and the large data catalogues that have been generated by the Landsat missions
(Nelson et al. 1994; Chambers et al. 2009; Negrón-Juarez et al. 2010 & 2011). A difficulty with
applying optical remote sensing based methods to mapping disturbance in tropical forests is the
presence of cloud cover and the limited time the disturbance remains visible from space. The
northwest Amazon experiences high annual rainfall (Fig. C-1) and experiences near constant
cloud cover for much of the year, with any given square kilometer only having a 10-20%
probability of being clearly observed from space (Fig. C-2). Dry seasons in much of the
northwest Amazon typically experience no month with less than 100-mm rainfall (Sombroek
2001). Lack of cloud free images makes it difficult to observe the short time window when
blowdowns can be detected from satellite images by the spectral signal created from the large
quantities of exposed, dead wood. After large canopy gap disturbances, lianas and colonizing
pioneer species grow quickly and obscure the spectral signal of dead wood. Previously, only
Landsat images with minimal cloud cover were used to detect blowdowns because manually
masking clouds greatly increases the time to process each image. This substantially reduced the
number of images available to map blowdowns. Here we use automated cloud masking
algorithms, incorporate machine learning techniques, and distributed computing to greatly
increase the quantity of images that can be used to map the time series of blowdowns in this
especially cloudy area of the Amazon basin. To do this, we leveraged the computing power of
69
Google Earth Engine (Hansen et al. 2013) to analyze a far larger collection of Landsat data than
previously used to map blowdowns across the duration of the Landsat 5 and 7 missions to
characterize the spatiotemporal variation in blowdown activity. In doing so, we are also able to
quantify the frequency distribution of blowdown sizes with the most exhaustive blowdown
dataset ever compiled. Most of all, we utilize this time series to address our main study
objectives of investigating how the interannual variation in blowdown activity varies with
regional changes in climate in addition to global scale climatological patterns driven by sea
surface temperature (SST) anomalies.
Methods
Study area description
The study region was comprised of 20 Landsat Worldwide Reference System (WRS)
scene locations in the northwest Amazon that in total encompassed an area of approximately
587,000 km2 (Fig. 4-1; Table C-1). Annual rainfall varies between approximately 2,500 - 4,500
mm/year, with an interannual standard deviation ranging between 200 - 600 mm, depending on
location (Fig. C-1). The drier half of the year lasts from early July to December, although the
onset starts earlier in the southern half of the study region. Only the southern half of the study
region experiences any month (1-2 months on average) with < 100 m of rainfall (Sombroek
2001). Due to relatively little deforestation in comparison to the rest of the Amazon (Hansen et
al. 2013), the study region still retains large, contiguous tracts of forest (Hansen et al. 2013).
Use of multispectral imagery to map blowdowns
We initially attempted to overcome persistent cloud cover by mapping blowdown events
with both the MODIS Nadir-Bidirectional Reflectance Distribution Function and Enhanced
Vegetation Index 16-day composited products, but these approaches ultimately failed. Spatially
variable cloud and aerosol contamination that could not be resolved or removed resulted in too
70
much spectral variability between adjacent forest pixels to differentiate blowdowns from non-
blowdown area with confidence. The coarse 250-500 m pixel size of MODIS data also prevented
correction for false-positive blowdown identifications (falsely identified forest disturbances due
to forest clearing, agriculture, seasonal wetlands, or cloud/aerosol contamination) in all but the
largest size classes (> 5 km2). Conversely, the higher spatial and spectral resolution of Landsat 5
TM and 7 ETM+ imagery allows forest blowdowns to be visually distinguished from other
landscape features by the characteristic shape of blowdowns, in addition to their unique spectral
signature. The blowdown shape typically resembles a broken fan, where the epicenter is located
at the base and extends radials of decreasing disturbance intensity. The spectral signature of a
blowdown is largely dominated by the signal of non-photosynthetic vegetation, with some
inclusion of green vegetation, as opposed to undisturbed forest that is mostly comprised of green
vegetation.
The visible blowdown:forest area ratio
Blowdown occurrences have been suggested to be more frequency in areas with high
precipitation (Nelson et al. 1994; Espírito-Santo et al. 2010). However years with high rainfall
tend to exhibit high cloud cover in the Amazon, and by extension, fewer usable Landsat
acquisitions (Asner, 2001). Therefore, years with high rainfall may have both higher blowdown
activity, but decreased probability of detecting blowdowns due to higher cloud cover, which may
have induced a bias in earlier blowdown analyses that relied on nearly cloud-free imagery.
Instead of estimating the level of blowdown disturbances from just cloudless imagery, we
quantified the ratio of visible blowdown to visible forest (BD:F) for cloud-free sections of all
available imagery. Blowdowns are not well distributed across the landscape, so any individual
Landsat scene acquisition could be biased high or low depending upon the cloud placement. A
way to overcome this bias is to process many Landsat scenes, so that positive and negative biases
71
may partially cancel each other out when estimating the overall median of BD:F across the
northwest Amazon.
Remote sensing workflow
We used Google Earth Engine (Hansen et al. 2013) to expedite blowdown mapping of a
large number of Landsat images. As shown in the image analysis workflow (Fig. 4-2), we
collected 975 Landsat 5 TM and Landsat 7 ETM+ images from the 20 WRS scene locations that
were acquired between July and February and contained less than 30% cloud cover. Landsat
scenes between March and June over this area were too cloudy to use. The Landsat ecosystem
disturbance adaptive processing system (LEDAPS; Schmidt et al., 2013) algorithm was applied
to both Landsat collections to convert the images from radiance to surface reflectance. Landsat 5
imagery spanned a time period from 1984 - 2012, whereas Landsat 7 imagery was only used
from 1999 - 2003 because the scanline corrector failure in 2003 resulted in a striping pattern of
missing data within each image.
Rivers, swamps, non-forest outcrops of savanna, apparent logging activity, human
settlements, and agriculture were masked by hand from the analysis with visual verification from
the higher resolution imagery accessible from Google Earth. Clouds and atmospheric haze were
removed from each image through a dynamic masking procedure. Spectral unmixing procedures
(Adams and Gillespie, 2006) were applied to masked images to estimate the per-pixel fractions
of the following endmembers: green vegetation (GV), non-photosynthetic vegetation (NPV),
haze contamination, and shade. Haze over forests will artificially inflate estimates of NPV, so a
haze endmember was included to reduce this artifact. We then used the surface reflectance bands
and the fraction of endmembers per pixel in a land cover classification. A training set of samples
was collected from the Landsat 5 and 7 scenes “LT50050602001008AAA02” and
“LE70040641999324AGS00”, respectively, for the following land-cover categories: upland
72
terra-firme forest, seasonally flooded forest, secondary forest, blowdowns, haze contaminated
forest, and cloud edges that were missed by the dynamic cloud mask. The spectral bands and the
unmixed endmember fractions were stacked, sampled with the training samples, and used to train
a Random Forests algorithm (Breiman, 2001) for classifying the Landsat imagery.
Description of false positive blowdown correction
A binary blowdown/non-blowdown raster was created from the classification and was
despeckled by filtering out clusters of less than four (0.36 ha) blowdown pixels. The binary
blowdown/non-blowdown raster was then smoothed with a morphological filter and turned into
polygons (Fig. 4-2) using the GDAL geospatial abstraction library (Warmerdam, 2008). Several
landscape features that were not part of the training classes used, such as recently logged forests
or seasonal swamps, were classified into the blowdown polygons, so an extensive removal
process of false-positive blowdown identifications was necessary. The frequency of false-
positive identifications decreased with blowdown size, therefore a size threshold of 25 ha for the
polygonized blowdowns was set as the minimum size blowdown so as to minimize the rate of
false-positive blowdown identifications. Blowdown polygons were then visually checked for
being false-positives with the corresponding Landsat images until the false-positive rate across
scenes was under 10%. Next, the ratio of visible blowdown area to visible forest area (BD:F) was
calculated for each processed Landsat scene. The Landsat image of a given WRS scene location
that exhibited the highest visible BD:F was selected as the observation for that year. For the
analysis of the blowdown size distribution, all blowdown polygon geometries across the 28 year
time series were merged so that double observations of the same blowdown were not double
counted.
73
Climate data sources
Amazon downbursts are thought to be produced when hot air masses are rapidly cooled
by precipitating convective storm cells, where the rapid increase in air density from cooling
causes the air to sink at high velocities creating the downburst (Garstang et al., 1998). In terms of
regional climate variation, blowdowns are thought to be most common in areas with higher
rainfall (Nelson et al., 1994) and more frequent high-precipitation events (Espírito-Santo et al.
2010). There has also been increasing evidence that macro-scale SST anomalies produce climate
teleconnections in the Amazon - which has proven effective in forecasting the fire disturbance
regime of the southern Amazon (Fernandes et al. 2011; Chen et al., 2011). The El Niño Southern
Oscillation (ENSO) alters regional precipitation regimes across the globe (Ropelewski and
Halpert, 1987), and causes tree-killing droughts across large swaths of the Amazon basin
(Williamson et al., 2000; Nepstad et al. 2004), while the Pacific Decadal Oscillation (PDO) and
the Atlantic Multidecadal Oscillation (AMO) are thought to exacerbate or moderate the impact
of ENSO-driven precipitation impacts depending on their phase (teleconnections) (Kayano and
Andreoli, 2007; Kayano and Capistrano, 2014).
To examine the correlation between temporal variation in regional climate and
blowdowns, we averaged climate data over each WRS scene on an annual basis. We used annual
rainfall from the Climate Hazards Infrared Precipitation with Station data (CHIRPS; Funk et al.
2014). We also used precipitation rate for the years 1998-2011 from the Tropical Rainfall
Measuring Mission (TRMM; product version 3B42). Temperature, humidity, wind speed, latent
and sensible heat fluxes were extracted from the National Center for Environmental Prediction’s
Climate Forecast System Version 2 (NCEP - CFSV2; Saha et al. 2014). For macro-scale
variation in climate, we used SST index anomalies that were the monthly time series of the
AMO, the PDO, and the Southern Oscillation Index (SOI; a commonly used index of El Niño)
74
from the National Oceanic and Atmospheric Association’s Earth System Research Laboratory
climate index data archive (http://www.esrl.noaa.gov/psd/data/climateindices/list/). The AMO
and PDO time series were generated from the Kaplan extended Sea Surface Temperatures
(Kaplan et al. 1998), whereas the Southern Oscillation Index is derived by differencing the mean
sea level air pressure between Tahiti and Darwin, Australia (Ropelewski and Jones, 1987).
Power-law estimation of blowdown sizes
To provide comparison between this study and previous blowdown and canopy gap
disturbance studies, the size frequency distribution was assumed to follow a power-law (Fisher et
2008), where the number of blowdowns (η) of a given size (z) is characterized by ηz=Az-α, where
A is the minimum blowdown size and α controls the frequency of large blowdowns (or the
degree of clustering). A large α exponent indicates most blowdowns were smaller, whereas a
small α indicates that more blowdown area was clustered in very large blowdowns. We estimated
α by fitting the blowdown sizes to a continuous truncated Pareto distribution, where 25 ha was
used as the minimum blowdown size (A), because we could not confidently discriminate
blowdowns < 25 ha. α was calculated by solving for the analytical solution from the probability
density function of the Pareto distribution. The confidence interval for α was estimated by
maximum likelihood using the mle2 function from the ‘bbmle’ package (Bolker, 2015) for the R
language. Estimates of α are sensitive to the methods used to delineate blowdown area, yet
because our analysis delineated blowdowns conservatively and often dissected single blowdowns
into multiple parts (Fig. C-3), our estimates of α are very likely higher than the true α. Also,
contiguous blowdowns were often interrupted by clouds and non-target landscape features -
which reduced individual blowdowns into multiple smaller blowdowns, and should also
upwardly bias α.
75
Statistical analyses
Spatial differences in blowdown frequency were evaluated first by comparing the BD:F
ratio across the 20 Landsat WRS scene locations, and a 2-D kernel density estimate using the full
28 year blowdown record using the kde.points function from the ‘GISTools’ package (Brunsdon
and Chen, 2014) for the R language. We applied regression techniques to examine the
relationship between the BD:F ratio and climate variables. Specifically, we evaluated temporal
trends with random-intercept linear models using the lmer function of the 'lme4' R language
package (Bates et al. 2014), where the intercept was allowed to vary with each Landsat WRS
scene location. We calculated a Pseudo R2 of the random intercept models according to
Nakagawa and Schielzeth (2013) using the r.squaredGLMM function of the ‘MuMIn’ package
(Barton 2016) for the R language.
We used linear regression models fit by Ordinary Least Squares to examine the region’s
overall interannual variation of BD:F with SST indices. We tested the annual maximum, annual
minimum, and each monthly value of each SST index for the maximum explanatory power of
interannual variation of BD:F. The optimum time window for the AMO was the annual
maximum, the December value for the SOI, and the July value for the PDO. We then tested all
possible permutations of these three covariates, and ranked them with Akaike’s Information
Criterion. The top-ranked model was then used to evaluate the impact of the SST anomalies on
the annual median BD:F of the 20 northwest Amazon Landsat WRS scene locations.
Results
Blowdown activity was spatially variable (Fig. 4-3 & 4-4) across the 20 Landsat WRS
scenes. When analyzed by location, the estimate of α varied between 1.196 and 1.586 (Table C-
1). The alpha exponent of the overall region over the 28 year time series was estimated to be
1.299 (95% CI: 1.256 - 1.344). This estimate was considerably lower than estimates from
76
previous research and suggests that blowdown area is more aggregated in very large blowdowns
than previously published estimates from elsewhere in the Amazon (Fig. C-5). The median
amount of blowdown activity varied by a factor of 8.4 across the 20 scenes. Similarly, the
standard deviation of BD:F across all 20 WRS scene locations varied from 9.2e-05 to 8.7e-04
through the 28 year time series, or by a range of 9.5x (Fig. C-10). While all of the 20 WRS
scenes locations experienced blowdowns during the 28 year time series, blowdowns were most
common in the central scenes (6-60, 6-61, 7-60, 7-61; Fig. 3), and least common in the most
southern scenes (8-63, 9-63; Fig. 4-3). The scene locations with the highest mean annual rainfall
and highest mean annual frequency of rainfall events in excess of 10 mm/hr (Fig. C-9) did not
correspond with the scenes of highest blowdown activity (Fig. 4-3).
There was significant meteorological change in the northwest Amazon during the 28
years of the study. Between 1984 and 2011, mean daytime air temperatures decreased by more
than 0.8℃, nighttime temperatures increased by 0.5℃. In the scenes located entirely in the
lowland Amazon (16 scenes, see Fig. 4-1), surface humidity increased by 8.5% while Latent
Heat Flux increased by 12.8%, both of which promote thunderstorms. Annual rainfall across all
scenes significantly increased at a rate of 6.2 mm/yr during the 1984 - 2011 interval, but did not
change or decreased slightly in two locations (WRS: 5-61, and 7-60, respectively; Fig. 4-5). The
maximum annual precipitation rate increased through time in all scene locations (Fig. C-8).
Despite these changes in regional climate, none of them explained much variation (R2 > 0.05) in
BD:F (Table C-2). Also when compared by WRS scene locations, the mean annual frequency of
high precipitation events (Fig. C-9) did not correspond with the Landsat WRS scene location's
median BD:F (Table C-2).
77
Blowdown activity increased through time (p < 0.001 & R2 = 0.046), but was also highly
variable over time (Fig. 4-4). Interannual variation of the overall median of BD:F varied by a
factor of 16.9. The standard deviation of BD:F across all 20 WRS Landsat scene locations for a
given year varied from a minimum of 4.26e-04, to a maximum of 7.52e-04, or by a factor of 17.6
(Fig. C-9). Daytime temperature, mean wind speed, were all positively correlated with BD:F,
while the dry season rainfall negatively correlated, but the corresponding R2 was negligible to
low (< 0.05) for all tested climate variables (Table C-2).
The overall variation in BD:F across the northwest Amazon positively corrleated with the
annual maximum of the Atlantic Multidecadal Oscillation index (adjusted R2 = 0.302 & p <
0.002; Fig. C-6), the December value of the El Niño/Southern Oscillation Index (adjusted R2 =
0.363, p < 0.001; Fig. C-6), and the July value of the Pacific Decadal Oscillation (adjusted R2 =
0.137, p < 0.03; Fig. C-6). The ENSO was correlated with both the AMO and PDO (Table C-3).
The best fit model explained 64% of the interannual variation in BD:F (p < 0.001) and followed
the form:
BD:F = -3.3e-2 + 2.54e-04×AMOmax - 4.96e-02×SOIDec + 3.22×PDOJul + 4.91e-05×SOIDec×PDOJul (3-1)
where BD:F was the annual median of visible blowdown area to visible forest area across all 20
Landsat scene locations, AMOmax was the yearly maximum of the Atlantic Multidecadal
Oscillation index, SOIDec was December value of the El Niño/Southern Oscillation Index, and
PDOJul was the July value of the Pacific Decadal Oscillation Index (Fig. 4-6). A summary of the
SST anomalies effect on BDA:VF is presented Table C-4.
Discussion
The spatial, temporal, and size variation of forest blowdowns
We found blowdowns are more clustered in space (with a lower α exponent) than
previously reported. Kellner and Asner (2009) estimated an exponent of 1.84 - 2.33 across five
78
tropical forests in Costa Rica and Hawaii, while Asner et al. (2013) estimated a mean exponent
of 1.83 across different forests in the southern Peruvian Amazon. Our estimate of 1.299 (95% CI:
1.256 - 1.344) is considerably lower (Fig. C-5), which likely reflects the difference in
disturbance mechanism, where canopy gaps in this study were formed by downbursts as opposed
to canopy gaps formed by falling trees from background mortality. The α exponent estimate of
1.5 (Chambers et al. 2009) for the Nelson et al. (1994) dataset of 330 blowdowns, derived from
137 Landsat scenes taken from the central Amazon over a time period from 1987-1989 is still
higher than ours for the northwest Amazon. This discrepancy may be partially explained because
the Nelson et al. (1994) dataset is considerably smaller than ours (330 vs. 3,524 blowdown
observations), which can result in an upwardly biased estimation of the exponent (White et al.
2008). Also, the Nelson (1994) blowdowns were delineated by hand, in contrast to our
automated approach. Next, the Nelson (1994) blowdowns were mapped over a three year time
series, which may have been too short a time interval to capture a high blowdown activity year.
Finally, the annual rainfall of the northwest Amazon study area is 500 - 1000 mm higher than in
Nelson’s central Amazon study area, which translates to more convective storms in the northwest
Amazon, and most likely more blowdowns.
We documented high temporal variation in blowdown activity, where the median of the
overall northwest Amazon BD:F varied by a factor 16.9 across years. There was also
considerable spatial variability, where when averaged across the full 28 year time series the
BD:F varied by a factor of 8.4 across the 20 Landsat WRS scene locations. The high interannual
and spatial variation of blowdown activity documented here indicates that characterizations of
forest disturbance regimes from short time intervals, or limited spatial extents should not be
extrapolated to large areas.
79
The results of this study are most likely underestimating the true rates of blowdown
disturbance for two reasons. First, processing partially clouded imagery improves the detection
of blowdowns - but for many years and scene locations there are no images with acceptable
cloud cover available. Missing data was especially true for the Landsat WRS scene locations
flanking the Andes where the probability of cloud-free observations was lowest (Fig. C-2). The
bias from cloud cover could explain the lack of the expected relationship between rainfall
intensity (Table C-2) and blowdown activity (Fig. 4-3), because the scenes with the highest
precipitation (Fig. C-1) and storm frequency (Fig. C-8) were also the cloudiest (Fig. C-2).
Hierarchical modeling techniques that explicitly deal with observation bias have become
common in species occupancy modeling where observations are typically biased by sampling
effort (Royle et al., 2008; Hobbs and Hooten, 2015). We did not correct for observation bias
induced by the varying amounts of cloud cover and the number of available observations
(Landsat acquisitions) across the 20 Landsat WRS scene locations, although future investigations
could estimate BD:F by treating it as a latent variable within a hierarchical model, where an
appropriate function relating bias in BD:F to cloud cover and Landsat scene sample size could be
derived to improve the estimation of BD:F and propagate the surrounding uncertainty.
Overall blowdown activity was underestimated because we did not map blowdowns < 25
ha. The frequency of blowdowns between 0.1 - 25 ha is still poorly described via remote sensing
methods because while Landsat can discriminate forest gaps on even the sub-pixel scale
(Negrón-Juarez et al, 2011), it becomes difficult to differentiate natural canopy gap disturbances
< 20-30 ha from anthropogenic disturbances such as logging activity and slash-and-burn
agriculture with Landsat 30 m2 imagery. To remove the possibility of including human
disturbance as a blowdown, we removed a large buffer around roads, rivers, and settlements
80
from the analysis, altogether excluding 29,764 km2 from the study area, which may have also
included blowdowns.
Regional scale climatic drivers of forest wind disturbance
Of the many changing regional climatic variables in the northwest Amazon, maximum
precipitation rate was the best predictor of blowdown activity across the 20 Landsat scene
locations, although it explained little variation (Table C-2). Only two of the 20 Landsat scene
locations did not exhibit a positive trend of increasing blowdown formations through the 1984 -
2011 study period. These were also the only two locations that did not exhibit a positive trend for
annual rainfall (Fig. 4-6; scene locations 5-61 & 7-60). The effect of both total dry season
rainfall and precipitation rate on blowdown activity was statistically significant, yet did not
explain much of the underlying variation (Table C-2). The negative trend of total dry season
rainfall with BD:F, coupled with the positive trend of max day time temperatures may suggest
downbursts are more likely to occur in hot years, with less rainfall. These trends may be
consistent with the understanding that downbursts are more likely to form when hot air masses
are rapidly cooled by a passing convective thunderstorm (Garstang et al., 1998). The variation in
blowdown activity was also not well explained by any of the other changing regional climatic
variables. It is worth noting that the duration of downbursts lasts from only a few to 15 minutes
(Garstang et al. 1998), whereas the climatological data used in this study was measured over
much coarser time intervals (3 hrs TRMM, 5 days CHIRPS), on much coarser spatial scales (5 -
55 km versus 30 m2 for the Landsat pixels), or produced by satellite based remotely sensing
(TRMM & CHIRPS) or derived from a hybridized circulation model (NCEP - CFSV2) rather
than empirically observed. Thus it is possible that we could not resolve an appreciable amount of
the BD:F variation with regional meteorological variables because there was no available data at
appropriate temporal and spatial scales.
81
Macroscale driver of forest wind disturbance
The temporal variation of the northwest Amazon’s overall BD:F corresponded most with
the macroscale climate SST indices of the AMO, SOI, and PDO. The magnitude of the AMO
anomaly varies on an annual basis, yet the long interval of roughly 55 - 70 years between
minimum and maximum phases (Schlesinger and Ramankutty, 1994; Knudsen et al. 2011) has
hindered investigation of its effects on the Amazonian climate. When the North Tropical Atlantic
is anomalously warm (positive phase AMO), it slows the southward drift of the Inter Tropical
Convergence Zone (ITCZ). The delay in arrival of the ITCZ consequently reduces rainfall over
the Amazon (Nobre and Shukla, 1996). The reduction of rainfall over the northwest Amazon is
further exacerbated in years when a warm AMO is joined by La Niña conditions (Kayano and
Capistrano, 2014). Also, large differences between dry and wet season monthly rainfall in the
Amazon can be largely explained by a linear combination of the AMO and low frequency
variations of the El Niño Southern Oscillation (Garcia-Garcia and Ummenhofer, 2015), where it
has also been suggested the PDO and El Niño Southern Oscillation modulate the strength of the
AMO driven precipitation anomaly (Garcia-Garcia and Ummenhofer, 2015).
The AMO's effect on blowdown activity and Amazon rainfall may seem counterintuitive,
because most evidence thus far has suggested blowdown activity to correspond with total
precipitation (Nelson et al. 2014; Fig. 4-5) or high storm frequency (Espírito-Santo, et al. 2010).
Yet we also found a negative correlation between total dry season rainfall and BD:F, and a
positive correlation with maximum daytime temperatures and BD:F (Table C-2). The conditions
for downburst to occur can arise when hot air is rapidly cooled by a precipitating storm cell.
Garstang et al. (1998) suggested that atmospheric instabilities during the dry to wet season
transition may promote conditions for downbursts to occur in the Amazon. The delayed onset of
the ITCZ by the AMO, coupled with warmer dry season temperatures and exacerbated
82
differences in dry/wet season rainfall may promote greater atmospheric instability during the
dry/wet season transitions - although it is beyond the scope of the current study to speculate
further on how these systems interact or promote downbursts. However, to our knowledge this is
the first observation of the AMO, SOI, and PDO promoting wind disturbance in the Amazon.
Finally, the relatively simple linear model with AMO, SOI, and PDO covariates already explains
more than half of the interannual variance of BD:F (Fig. 4-6), yet this relationship could be
improved to capture the geographic differences in SST-precipitation anomalies that exist
between the northwest and central Amazon (Kayano and Capistrano 2014, García-García and
Hummenhofer, 2015). To account for these sub-regional differences, future studies might
optimize the SST index time lags for different spatial regions, as was done by Chen et al. (2011)
in their predictions of Amazon fire frequencies with SST indices.
Conclusions
In this study we document the Landsat observable blowdowns over a 570,000 km2 swath
of the northwest Amazon, by which we use this data record to make the following points. First,
blowdown activity varies tremendously through space (8.4x) and time (16.9x), although the true
differences are obscured by high cloud cover resulting in limited Landsat acquisitions over areas
(primarily scenes, 9-60 and 9-61; Fig. C-2) that have been predicted to withstand high levels of
wind disturbance because of high storm frequency (Fig. C-9). Second, the observed size
distribution of blowdowns in this study (Fig. C-5) shows far more large blowdowns than prior
studies on Amazon blowdowns - suggesting blowdown area may be more clustered in space than
previously thought. Third, the most probable proximate cause of blowdowns seems to be related
to precipitation rate and perhaps the amount of dry season precipitation, although the strength of
these regional climate relationships is weak - perhaps because of insufficient resolution in
meteorological data to capture the downburst events. Yet the strongest relationship with
83
blowdown activity in the northwest Amazon most corresponds with the AMO, ENSO, and PDO
SST indices (Fig. 4-6 & Fig C-6).
It is important to recognize that the rates of wind disturbance reported in this study are
underestimates because of under-detection of blowdowns by the remote sensing methodology,
but also because of observation biases induced by nearly ever-present cloud cover in areas with
high storm frequency. With the exception of some LiDAR-based studies with limited spatial and
temporal extent, nearly all published disturbance frequencies of blowdowns in the Amazon Basin
have relied upon passive optical remote sensing with similar limitations and biases, and to the
best of our knowledge - only the Landsat array has proven successful. However, there are
promising advances in remote sensing that will improve the detectability of catastrophic forest
disturbances. The increasing availability of cloud-penetrating radar satellite products (ex: ALOS-
1 PALSAR & Sentinel-1a) will open a new frontier in detection of forest disturbances, where the
frequent cloud cover of the northwestern Amazon basin should not be a major issue for detecting
disturbance. The advent of the Sentinel-2a mission will also greatly augment the number of
annual multispectral acquisitions, and perhaps allow wind disturbances < 25 ha to be mapped
because of its four spectral bands acquired at 10 m spatial resolution (as compared to Landsat’s
30 m). In total, we expect these advancements will result in less biased detection of forest
disturbances, and an increasingly accurate quantification of the Amazon wind disturbance
regime.
84
Figure 4-1. The 20 Landsat WRS Path/Row tiles of the northwest Amazon study region. The
background shows a Landsat 7 composites with RGB represented by bands 5, 4 & 3,
where the white areas have been masked from the study region. Also plotted are
national boundaries.
85
Figure 4-2. Remote sensing workflow. The (top) diagram illustrates the conversion process of
blowdown features in Landsat imagery to polygons (bottom).
86
Figure 4-3. Kernel density estimation of large blowdowns (> 25 ha). The probability density per
degree cell that a blowdown occurred over the 28 year study period is represented by
the color scale. Colored circles show all the specific occurrences of blowdowns > 25
ha over the study period.
87
Figure 4-4. Ratio of blowdown : forest area by Landsat location from 1984 to 2011. Overall
linear trend (p < 0.001) with surrounding 95% confidence interval is overlaid.
88
Figure 4-5. Temporal rainfall trend and blowdown trend. (A) Linear trend of annual rainfall for
1984 - 2012 . The spatial location where the two outliers in terms of BD:F , (panel B)
are located highlighted in a map of rainfall trend (panel A). (B) The temporal trend of
annual blowdown:forest area (BD:F) for each of the Landsat WRS scene locations.
The two scene locations with negative trends in blowdown:forest area, were also the
only two scene locations without an increasing precipitation trend (marked with red
stars).
89
Figure 4-6. Predictive relationship between SSTs and BD:F. (Top) Time series of the annual median of
blowdown:forest area (BD:F) of northwest Amazon with the yearly maximum of the Atlantic
Multidecadal Oscillation index (AMO), the Southern Oscillation Index during the month of
December (SOI), and the Pacific Decadal Oscillation index during the month of July (PDO).
(Bottom) Predicted versus observed annual median of BD:F of northwest Amazon, with 1:1
line (blue).
90
CHAPTER 5
OVERALL CONCLUSIONS
Blowdowns in the northwest Amazon affect forest structure, demographics of tree
survival and colonization, and tree species diversity. Yet the wind disturbance regime has been
difficult to quantify because of nearly year-round cloud cover, which has hindered observation
by optical remote sensing. In Chapter 1, I have shown that catastrophic wind disturbance has a
proclivity to remove the largest trees in the forest, and selectively kill trees of lower wood
density. Incorporating differential mortality when scaling the biomass loss from disturbance
resulted in a 51% increase of necromass, compared to previous scaling methods that used
random tree mortality.
In Chapter 2, I demonstrate the sustained reduction in both species richness and evenness
following blowdown disturbances. Contrary to the predictions of the Intermediate Disturbance
Hypothesis and the Canopy Gap Partitioning Hypothesis, any level of disturbance reduced tree
species richness and evenness for decades. Blowdowns also promoted the niche partitioning of
canopy gaps by pioneer tree species, however this relationship was highly variable. Forest
patches that were outside the blowdown also hosted many pioneer tree species, which is in
contrast to previous theory on Neotropical forest succession that posits short-lived pioneer tree
species should be absent from ‘undisturbed’ forest. The reason for the continued presence of
pioneer tree species in the landscape is unknown, but presumably the non-catastrophic mortality
rates are high enough to produce enough canopy gaps for pioneers to occasionally colonize.
Finally, blowdowns act as a link between global climate and major forest structural
change across the northwest Amazon. In Chapter 3, I show through a large remote sensing
analysis that the annual amount of blowdown disturbance is both spatially and temporally
variable. Next I demonstrate that it has significantly increased between the years 1984 - 2011.
91
These changes were only weakly related to regional climate variation, although I caution that
meteorological data of finer temporal and spatial resolution is required to accurately investigate
these relationships. Finally, I show how global scale sea surface temperature anomalies can
predict more than half of the inter-annual variation in blowdown activity. The sea surface
temperature anomaly of the Atlantic Multidecadal Oscillation promotes blowdown activity. The
El Niño Southern Oscillation also promotes blowdowns during the La Niña phase, yet this effect
is partially mediated by an interaction with the Pacific Decadal Oscillation. Linking sea surface
temperature anomalies with climate driven forest disturbance regimes has only recently been
recognized, yet these relationships could be used to improve the understanding of forest
dynamics, as well as reduce the uncertainty surrounding the Amazon’s carbon balance.
92
APPENDIX A
ADDITIONAL TABLES AND FIGURES FOR CHAPTER 2
Table A-1. Mean landscape simulation estimate across 1000 parameter samples of the sum of
predicted dead trees and necromass across the different models.
Model Dead Trees Necromass (Mg)
∆NPV<0 ∆NPV<0.05 All ∆NPV<0 ∆NPV<0.05 All
sum %total sum %total sum sum %total sum %total sum
Linear
∆NPV –
calib. data
4511 10.6 11240 26.4 42559 1769 10.6 4411 26.4 16709
Linear
∆NPV –
transect
data
4160 12.9 9765 30.3 32236 1632 12.9 3235 30.3 12655
GLMM -
∆NPV
partitioned
1579 4.6 5594 16.4 34034 648 3.1 4256 20.1 21175
Non-spatial
GLM
7677 17.5 16276 37.0 43945 3420 13.9 7766 31.6 24564
Spatial
GLM
3125 9.0 8534 24.5 34793 1475 7.6 4525 23.2 19472
93
Table A-2. Logistic regression estimates for Model 2 – GLMM. The median estimates are
presented for logistic regression β coefficients for the GLMM using the transect data
partitioned by ΔNPV categories (Model 2) and the sample size of live and dead trees
used to fit the model.
ΔNPV range β.intercept β.DBH β.WD Live Trees Dead Trees
<0 -2.994 -0.005 -1.273 262 5
0-0.04 -3.508 0.041 -1.099 296 12
0.04-0.15 -1.821 0.025 -1.002 460 69
0.15-0.4 -0.759 0.041 -0.986 202 87
0.4-0.7 -0.761 0.070 -1.502 121 90
>0.7 -0.989 0.100 -1.299 30 40
94
Table A-3. Estimated individual dead trees (> 10 cm DBH) and associated necromass produced
by a 300 ha blowdown in the landscape (~500.5 ha). Simulation runs drew from the
joint-posterior distributions of associated model parameters. Linear ΔNPV-
necromass estimates were generated by random tree mortality.
Model Individual dead from blowdown Necromass (Mg) from blowdown
2.5% 25% 50% 75% 97.5% 2.5% 25% 50% 75% 97.5%
Linear ΔNPV-
calibration data 37367 40466 42377 44532 48690
14674 15848 16602 17493 19117
Linear ΔNPV-
transect data 29200 31092 32173 33270 35803
11465 12207 12619 13080 14122
ΔNPV-part.
GLMM 28519 32789 33969 35314 40409
16929 19609 21189 22726 25534
Non-spatial
GLM 37661 41718 43735 46059 50775
19491 22581 24432 26501 30345
Spatial GLM 10656 25022 34587 42460 65319 7032 15051 19159 23302 33634
95
Table A-4. Relative abundance of families comprising more than 1% of total individuals.
Family Individuals Relative Abundance (%)
Annonaceae 22 1.60
Apocynaceae 28 2.04
Arecaceae 63 4.60
Burseraceae 55 4.01
Chrysobalanaceae 61 4.45
Clusiaceae 28 2.04
Elaeocarpaceae 19 1.39
Euphorbiaceae 154 11.23
Fabaceae 111 8.10
Lauraceae 27 1.97
Lecythidaceae 186 13.57
Melastomataceae 15 1.09
Moraceae 83 6.05
Myristicaceae 116 8.46
Myrtaceae 14 1.02
Olacaceae 26 1.90
Rubiaceae 15 1.09
Sapotaceae 131 9.56
Sterculiaceae 37 2.70
Urticaceae 40 2.92
96
Figure A-1. Moran’s I statistic for model predictions residuals. (A) Linear ΔNPV-calibration
data fit, (B) Linear ΔNPV – transect data fit, (C) ΔNPV partitioned GLMM, (D) the
Non-spatial GLM, and (E) the Spatial GLM. Note difference in scale of Moran’s I
among model types. Red circles indicate the model residuals are spatially
autocorrelated (p>0.05).
97
Figure A-2. Polynomial (3rd degree) and LOESS functions to estimate the spatial random effect
(SRE) as a function of ΔNPV. SRE = -0.554 + 10.13*ΔNPV – 19.59*ΔNPV2 +
11.723*ΔNPV3. The probability of death increases with SRE. The mean mortality
rate, conditional on the values of other predictors in Model 4 (equation 4), occurs
when SRE = 0. Positive and negative values of SRE, respectively, indicate greater
and lesser probability of death compared to the conditional mean.
98
Figure A-3. The ratio of the probability of mortality of different sized trees across disturbance
intensity to when spatial autocorrelation is included versus when the spatial random
effect is excluded. The y-axis is the ratio of mortality probability of the model with
spatial autocorrelation (Model 4) divided by the mortality probability of the model
when the spatial random effect is constrained to zero. The ratio is plotted for three
diameters (10, 50, 100 cm), with median WD (0.64) and elevation (149).
99
Figure A-4. Estimated mortality probability of a tree with median characteristics with increasing
disturbance intensity as predicted by a spatial logistic regression model (SPDE
GLM). Shaded regions represent 95% confidence interval as calculated by the Delta
Method. Unless otherwise specified, plotted predictions are for a tree of 17.3 cm, 0.64
wood density, and situated at 149 m elevation. Probability of death is only estimated
by the fixed effects here, meaning the spatial random effect is constrained to zero
across the range of ΔNPV.
100
Figure A-5. The differential probability of death (Model 4) for trees with different diameters to
wood density ratios with (bottom) and without (top) the SRE, estimated from the
spatial GLM. Shaded regions represent 95% confidence interval as calculated by the
Delta Method. Plotted predictions are for trees situated at 149 m elevation with DBHs
of 75 cm, 50 cm, 22.5 cm, and 10 cm, respectively, with wood densities of 0.25, 0.5,
0.75, and 1, respectively.
101
Figure A-6. The estimated multiple of probability of death without the SRE across a gradient of
disturbance intensity (ΔNPV). A tree of (A) large diameter (55 cm), (B) low wood
density (0.32), and (C) high elevation (162 m) are compared to a tree of low diameter
(10.3 cm), high wood density (0.97), and low elevation (134 m), respectively. The
dashed line indicates no multiplicative difference in the probability of death.
Differences in diameter, wood density, and elevation represent the bounds of 2.5 and
97.5% quantiles of the inventory data.
102
Figure A-7. The distribution of mean necromass per tree by simulation run of the different
models. ΔNPV linear model (Model 1 - black), the ΔNPV partitioned GLMs (Model
2 - green), the non-spatial GLM (Model 3 - red), and the spatial GLM (Model 4 -
blue) are plotted. Dotted line indicates the median value of the associated model.
103
Figure A-8. Distribution of pixel ΔNPV in the blowdown encompassing landscape (green) and
associated ΔNPV of 30*10 m transect sub-plots of the inventory data (blue).
104
Figure A-9. Ratio of model predicted to observed necromass by ΔNPV class. The linear ΔNPV
model fit with the calibration plots (gray) and transect plots (yellow), the non-spatial
GLM (red), the spatial GLM (blue), and the ΔNPV partitioned GLMM (green) are
plotted. The horizontal gray line indicates a ratio of 1:1
105
Figure A-10. Elevation of transect subplots by associated ΔNPV.
106
Figure A-11. The mean Global Wood Density Database values of 46 species compared to
measured wood density. Only species with more than five observations and a
standard deviation ≤ 0.1 were included. The 1:1 ratio is shown by the red line. Only
139 of the 413 species encountered in the 3 ha forest inventory area had at least one
observation in the Global Wood Density Database. Intraspecific variation was high
for species in the Global Wood Density Database that were also encountered in the 3
ha plot. Of the 90 species that had at least two observations in the Global Wood
Density Database, 72 had a range in wood density values greater than 0.1.
107
Figure A-12. The fraction of observed adult tree mortality in the calibration plots (green) and the
transect plots (blue) vs. the change in pre- and post- disturbance non-photosynthetic
disturbance (ΔNPV). Unlike Model 1, an intercept was allowed in these models –
although this reduced the R2 values and let to over-prediction of mortality when
ΔNPV was near 0.
108
APPENDIX B
ADDITIONAL TABLES AND FIGURES FOR CHAPTER 3
Table B-1. Landsat 5 TM images of blowdown sites before and after disturbance.
Name Location Pre-disturbance Post-disturbance
Nauta -73.60, -4.39 LT50060632009245CUB00 LT50060632009341CUB00
Mishana -73.48, -3.88 LT50060632005218CUB00 LT50060632005234CUB01
Napo -73.09, -3.14 LT50060621998231XXX02 LT50060621998295XXX02
Alpahuayo -73.45, -3.95 LT50060631991004CUB00 LT50060631994204CUB02
Orosa -72.43, -3.69 LT50060631988108CUB00 LT50060631988268CUB02
109
Figure B-1. Percentage of pioneer stems along ΔNPV gradient at the Nauta blowdown (1 yr
TSD).
110
Figure B-2. Nauta pioneer species contribution, 1 yr TSD. Total pioneer basal area (A) and
pioneer relative abundance (B) are plotted.
111
Figure B-3. Mishana pioneer species contribution. Total basal area (A) and relative abundance
(B) 6 yrs TSD are plotted.
112
Figure B-4. Napo pioneer species contribution. Total basal area (A) and relative abundance (B)
12 yrs TSD are plotted.
113
Figure B-5. Alpahuayo pioneer species contribution. Total basal area (A) and relative abundance
(B) 17 yrs TSD are plotted.
114
Figure B-6. Species richness accumulation curves (Hill number: q = 0) rarefied by individuals
for blowdown sites. (A) Nauta, (B) Mishana, (C) Napo, (D) Alpahuayo, (E) Orosa are
plotted. Time since disturbance is listed in upper left corner. Shaded regions represent
the 95% confidence interval from 500 bootstrapped simulations. Solid lines represent
interpolation while dashed lines were extrapolations.
115
Figure B-7. Simpson’s evenness accumulation curves rarefied by individuals for blowdown sites.
(A) Nauta, (B) Mishana, (C) Napo, (D) Alpahuayo, (E) Orosa are plotted. Time since
disturbance is listed in upper left corner. Shaded regions represent the 95%
confidence interval from 500 bootstrapped simulations. Solid lines represent
interpolation while dashed lines were extrapolations.
116
APPENDIX C
ADDITIONAL TABLES AND FIGURES FOR CHAPTER 4
Table C-1. Blowdowns by Landsat scene location. Forest Area was the maximum observed
visible forest by scene location, after masking. α represents the scaling exponent of a
power-law.
WRS
Path/Row
Monitored Forest
Area (km2)
Scenes
Analyzed
Blowdowns α Max Blowdown
Size (ha)
5-60 28890 57 231 1.432 1016
5-61 31491 30 146 1.273 1016
5-62 28992 61 220 1.298 618
5-63 29382 65 234 1.219 1471
6-60 27760 62 344 1.338 1521
6-61 31674 53 254 1.390 2333
6-62 26203 56 241 1.301 1135
6-63 26469 61 280 1.212 1382
7-60 29247 58 371 1.187 4666
7-61 30597 65 294 1.337 1649
7-62 30875 64 342 1.278 1066
7-63 29716 82 235 1.343 975
8-60 23649 21 116 1.196 1317
8-61 30422 35 206 1.247 1317
8-62 31922 55 287 1.255 1066
8-63 26283 65 161 1.339 394
9-60 5875 13 18 1.439 395
9-61 19180 11 115 1.586 797
9-62 21149 29 159 1.340 1469
9-63 11454 32 66 1.234 810
117
Table C-2. Relation of regional climate variables towards blowdown activity. The covariate is
listed in the climate variable, and the dependent variable is the annual ratio of visible
blowdown to visible forest (BD:F).
Climate variable Intercept Effect t-value p-
value
R2
Max day time
temperature
-9.957e-04 (5.089e-04) 3.774e-05 (1.522e-05) 2.479 0.0156
Max wind speed 4.776e-05 (1.289e-04) 9.223e-05 (5.409e-05) 1.705 0.007
Max 5-day
rainfall
1.587e-04 (5.761e-05) 1.760e-06 (8.906e-07) 1.977 0.010
Max
precipitation rate
1.731e-04 (8.698e-05) 3.377e-05 (1.259e-05) 2.683 0.035
Dry Season
Precipitation
8.010e-04 (1.718e-04) -6.932e-07 (2.776e-
07)
-2.497 0.03
Precipitation
trend *
9.312e-06 (3.859e-06) 2.968e-07 (3.013e-07) 0.985 0.05
Max Latent Heat
Flux
2.112e-04 (7.651e-05) 1.384e-07 (1.906e-07) 0.726 0.001
Max Specific
Heat Flux
8.090e-05 (1.355e-04) 5.956e-07 (4.331e-07) 1.375 0.005
MAF 10 mm/hr
*
7.699e-05 (7.839e-05) 4.678e-06 (6.326e-06) 0.739 0.469 0.029
MAF 20 mm/hr
*
1.153e-04 (4.332e-05) 1.318e-05 (2.882e-05) 0.457 0.653 0.011
MAF 30 mm/hr
*
1.297e-04 (2.710e-05) 2.559e-05 (1.344e-04) 0.190 0.851 0.002
Notes: 'Max' climate variables were taken as the maximum value during the year until the
Landsat image acquisition date. Dry Season Precipitation was defined as the cumulative rainfall
between Julian days 250 and 325. MAF is the mean annual frequency between 1998 - 2012 of
rainfall events with precipitation rate greater than 10, 20, or 30 mm/hr. p values are reported for
Ordinary Least Squares (OLS) fit linear models, which are indicated by a *. All other models
were fit as random-intercept linear mixed models, where p values are not commonly reported. R2
represents the Marginal R2 as for the random-intercept models calculated by the
r.squaredGLMM function in the 'MuMIn' package for the R language. R2 represents the Multiple
R2 for OLS models. t-value reflects the climate variable effect, and not the intercept.
118
Table C-3. Correlations between annual ratio of median visible blowdown to visible forest
(BD:F) in the northwest Amazon and SST indices. SST indices include the annual
maximum of the Atlantic Multidecadal Oscillation (AMO), the Southern Oscillation
Index in the month of December (SOI), and the Pacific Decadal Oscillation index in
the month of July (PDO).
BD:F AMO SOI PDO
BD:F *
AMO 0.573 *
SOI 0.622 0.478 *
PDO 0.411 -0.067 0.305 *
119
Table C-4. Linear model estimates of SST anomalies on the ratio of visible blowdowns to visible
forest. Adjusted R2 = 0.652. *p < 0.05; **p < 0.01
Effect Estimate Standard Error Significance.
AMO.max 0.0003 0.0001 *
SOI.Dec -0.05 0.015 **
PDO.Jul 0.00003 0.00002
SOI.Dec:PDO.Jul 0.00005 0.00001 **
Constant -0.033 0.021
120
Figure C-1. Annual rainfall map of northwest Amazon. Mean annual rainfall (left) and the
standard deviation of interannual rainfall (right).
121
Figure C-2. Fraction of cloud-free observations from MODIS Terra at 1 km pixel scale. The
highest rainfall areas indicated in Fig. C-1 are also the most cloudy.
122
Figure C-3. Blowdown and the algorithm derived blowdown polygon. Large blowdown from
Landsat 7 scene: LE70090622002255EDC00 (top) displayed in RGB with bands
5,4,3. The yellow represents the area categorized as blowdown (bottom). Notice the
large amounts of visibly disturbed forest not captured by the algorithm.
123
Figure C-4. Map of 28 years of northwest Amazon blowdowns. Northwest Amazon tiles outlined
in black, with masked areas in yellow. Black polygons represent all blowdowns
mapped between 1984-2011.
124
Figure C-5. The probability of different sized blowdown events occurring with α exponent
estimates from different studies in the Amazon. Also plotted (black dots) are the
probability densities of the blowdowns, binned accordingly: 25-50, 50 - 100, 100 -
200, 200 - 500, 500 - 1000, 1000 - 2000, and 2000 - 5000 hectares.
125
Figure C-6. BD:F and individual SST indices. (A) Annual median estimate of blowdown to
forest ratio (BD:F) against the annual max of the Atlantic Multidecadal Oscillation
index (AMO). (B) BD:F against the December SOI. (C) BD:F against the July PDO.
126
Figure C-7. The annual max precipitation rate increased between 1984 - 2012 for all northwest
Amazon scenes in the lowland.
127
Figure C-8. High rainfall event frequency across the northwest Amazon. Mean frequency of
precipitation rates exceeding 10, 20, and 30 mm/hr from Tropical Rainfall Measuring
Mission data from 1998 - 2011 (top) and associated standard deviations (bottom).
128
Figure C-9. Annual standard deviation between 20 Landsat WRS scene locations of the ratio of
blowdown area:visible forest area (BD:F).
129
LIST OF REFERENCES
Adams, J. B., and A. R. Gillespie. 2006. Remote Sensing of Landscapes with Spectral Images: A
Physical Modeling Approach. Cambridge University Press.
Asner, G. P. 2001. Cloud cover in Landsat observations of the Brazilian Amazon. International
Journal of Remote Sensing 22:3855–3862.
Asner, G. P. 2013. Geography of forest disturbance. Proceedings of the National Academy of
Sciences 110:3711–3712.
Baker, T. R., O. L. Phillips, Y. Malhi, S. Almeida, L. Arroyo, A. Di Fiore, T. Erwin, T. J.
Killeen, S. G. Laurance, W. F. Laurance, S. L. Lewis, J. Lloyd, A. Monteagudo, D. A.
Neill, S. Patiño, N. C. A. Pitman, J. N. M. Silva, and R. Vásquez Martínez. 2004.
Variation in wood density determines spatial patterns in Amazonian forest biomass.
Global Change Biology 10:545–562.
Barlow, J., C. A. Peres, B. O. Lagan, and T. Haugaasen. 2003. Large tree mortality and the
decline of forest biomass following Amazonian wildfires. Ecology Letters 6:6–8.
Bartoń, K. 2016. MuMIn: Multi-Model Inference.
Bates, D., M. Maechler, B. Bolker, S. Walker, R. H. B. Christensen, H. Singmann, B. Dai, and
G. Grothendieck. 2015. lme4: Linear Mixed-Effects Models using “Eigen” and S4.
Beguin, J., S. Martino, H. Rue, and S. G. Cumming. 2012. Hierarchical analysis of spatially
autocorrelated ecological data using integrated nested Laplace approximation. Methods in
Ecology and Evolution 3:921–929.
Bennett, A. C., N. G. McDowell, C. D. Allen, and K. J. Anderson-Teixeira. 2015. Larger trees
suffer most during drought in forests worldwide. Nature Plants 1:15139.
Bolker, B., and R. D. C. Team. 2014. bbmle: Tools for general maximum likelihood estimation.
Bongers, F., L. Poorter, W. D. Hawthorne, and D. Sheil. 2009. The intermediate disturbance
hypothesis applies to tropical forests, but disturbance contributes little to tree diversity.
Ecology Letters 12:798–805.
Breiman, L. 2001. Random Forests. Machine Learning 45:5–32.
Brokaw, N. V. L. 1985. Gap-Phase Regeneration in a Tropical Forest. Ecology 66:682–687.
Brokaw, N. V. L. 1987. Gap-Phase Regeneration of Three Pioneer Tree Species in a Tropical
Forest. Journal of Ecology 75:9–19.
Brokaw, N. V. L., and J. S. Grear. 1991. Forest Structure Before and After Hurricane Hugo at
Three Elevations in the Luquillo Mountains, Puerto Rico. Biotropica 23:386.
130
Brokaw, N. V. L., and L. R. Walker. 1991. Summary of the Effects of Caribbean Hurricanes on
Vegetation. Biotropica 23:442–447.
Brunsdon, C., and H. Chen. 2014. GISTools: Some further GIS capabilities for R.
Canham, C. D., J. S. Denslow, W. J. Platt, J. R. Runkle, T. A. Spies, and P. S. White. 1990. Light
regimes beneath closed canopies and tree-fall gaps in temperate and tropical forests.
Canadian Journal of Forest Research 20:620–631.
Canham, C. D., and O. L. Loucks. 1984. Catastrophic windthrow in the presettlement forests of
Wisconsin. Ecology 65:803.
Canham, C. D., M. J. Papaik, and E. F. Latty. 2001. Interspecific variation in susceptibility to
windthrow as a function of tree size and storm severity for northern temperate tree
species. Canadian Journal of Forest Research 31:1–10.
Chambers, J. Q., J. I. Fisher, H. Zeng, E. L. Chapman, D. B. Baker, and G. C. Hurtt. 2007.
Hurricane Katrina’s Carbon Footprint on U.S. Gulf Coast Forests. Science 318:1107–
1107.
Chambers, J. Q., R. I. Negrón-Juárez, G. C. Hurtt, D. M. Marra, and N. Higuchi. 2009a. Lack of
intermediate-scale disturbance data prevents robust extrapolation of plot-level tree
mortality rates for old-growth tropical forests. Ecology Letters 12:E22–E25.
Chambers, J. Q., R. I. Negron-Juarez, D. M. Marra, A. D. Vittorio, J. Tews, D. Roberts, G. H. P.
M. Ribeiro, S. E. Trumbore, and N. Higuchi. 2013. The steady-state mosaic of
disturbance and succession across an old-growth Central Amazon forest landscape.
Proceedings of the National Academy of Sciences 110:3949–3954.
Chambers, J. Q., A. L. Robertson, V. M. C. Carneiro, A. J. N. Lima, M.-L. Smith, L. C. Plourde,
and N. Higuchi. 2009b. Hyperspectral remote detection of niche partitioning among
canopy trees driven by blowdown gap disturbances in the Central Amazon. Oecologia
160:107–117.
Chao, A., N. J. Gotelli, T. C. Hsieh, E. L. Sander, K. H. Ma, R. K. Colwell, and A. M. Ellison.
2014. Rarefaction and extrapolation with Hill numbers: a framework for sampling and
estimation in species diversity studies. Ecological Monographs 84:45–67.
Chao, A., and L. Jost. 2012. Coverage-based rarefaction and extrapolation: standardizing
samples by completeness rather than size. Ecology 93:2533–2547.
Chao, A., and S.-M. Lee. 1992. Estimating the Number of Classes via Sample Coverage. Journal
of the American Statistical Association 87:210–217.
Chao, K.-J., O. L. Phillips, E. Gloor, A. Monteagudo, A. Torres-Lezama, and R. V. Martínez.
2008. Growth and wood density predict tree mortality in Amazon forests. Journal of
Ecology 96:281–292.
131
Chao, K.-J., O. L. Phillips, A. Monteagudo, A. Torres-Lezama, and R. Vásquez Martínez. 2009.
How do trees die? Mode of death in northern Amazonia. Journal of Vegetation Science
20:260–268.
Chave, J., C. Andalo, S. Brown, M. A. Cairns, J. Q. Chambers, D. Eamus, H. Fölster, F.
Fromard, N. Higuchi, T. Kira, J.-P. Lescure, B. W. Nelson, H. Ogawa, H. Puig, B. Riéra,
and T. Yamakura. 2005. Tree allometry and improved estimation of carbon stocks and
balance in tropical forests. Oecologia 145:87–99.
Chave, J., D. Coomes, S. Jansen, S. L. Lewis, N. G. Swenson, and A. E. Zanne. 2009. Towards a
worldwide wood economics spectrum. Ecology Letters 12:351–366.
Chave, J., H. C. Muller-Landau, T. R. Baker, T. A. Easdale, H. ter Steege, and C. O. Webb.
2006. Regional and Phylogenetic Variation of Wood Density Across 2456 Neotropical
Tree Species. Ecological Applications 16:2356–2367.
Chen, Y., J. T. Randerson, D. C. Morton, R. S. DeFries, G. J. Collatz, P. S. Kasibhatla, L. Giglio,
Y. Jin, and M. E. Marlier. 2011. Forecasting Fire Season Severity in South America
Using Sea Surface Temperature Anomalies. Science 334:787–791.
Chesson, P. 2000. Mechanisms of Maintenance of Species Diversity. Annual Review of Ecology
and Systematics 31:343–366.
Clark, D. B., and D. A. Clark. 1991. The Impact of Physical Damage on Canopy Tree
Regeneration in Tropical Rain Forest. Journal of Ecology 79:447–457.
Clark, D. B., and D. A. Clark. 1996. Abundance, growth and mortality of very large trees in
neotropical lowland rain forest. Forest Ecology and Management 80:235–244.
Connell, J. H. 1978. Diversity in Tropical Rain Forests and Coral Reefs. Science 199:1302–
1310.
Curran, T. J., L. N. Gersbach, W. Edwards, and A. K. Krockenberger. 2008. Wood density
predicts plant damage and vegetative recovery rates caused by cyclone disturbance in
tropical rainforest tree species of North Queensland, Australia. Austral Ecology 33:442–
450.
Dalling, J. W., M. D. Swaine, and N. C. Garwood. 1998. Dispersal Patterns and Seed Bank
Dynamics of Pioneer Trees in Moist Tropical Forest. Ecology 79:564–578.
Davidson, E. A., A. C. de Araújo, P. Artaxo, J. K. Balch, I. F. Brown, M. M. C. Bustamante, M.
T. Coe, R. S. DeFries, M. Keller, M. Longo, J. W. Munger, W. Schroeder, B. S. Soares-
Filho, C. M. Souza, and S. C. Wofsy. 2012. The Amazon basin in transition. Nature
481:321–328.
Denslow, J. S. 1980a. Gap Partitioning among Tropical Rainforest Trees. Biotropica 12:47–55.
132
Denslow, J. S. 1980b. Patterns of plant species diversity during succession under different
disturbance regimes. Oecologia 46:18–21.
Denslow, J. S., A. M. Ellison, and R. E. Sanford. 1998. Treefall gap size effects on above- and
below-ground processes in a tropical wet forest. Journal of Ecology 86:597–609.
Espírito-Santo, F. D. B., M. Gloor, M. Keller, Y. Malhi, S. Saatchi, B. Nelson, R. C. O. Junior,
C. Pereira, J. Lloyd, S. Frolking, M. Palace, Y. E. Shimabukuro, V. Duarte, A. M.
Mendoza, G. López-González, T. R. Baker, T. R. Feldpausch, R. J. W. Brienen, G. P.
Asner, D. S. Boyd, and O. L. Phillips. 2014. Size and frequency of natural forest
disturbances and the Amazon forest carbon balance. Nature Communications 5:3434.
Espírito-Santo, F. D. B., M. Keller, B. Braswell, B. W. Nelson, S. Frolking, and G. Vicente.
2010. Storm intensity and old-growth forest disturbances in the Amazon region.
Geophysical Research Letters 37:L11403.
Everham, E. M., and N. V. L. Brokaw. 1996. Forest damage and recovery from catastrophic
wind. The Botanical Review 62:113–185.
Farrior, C. E., S. A. Bohlman, S. Hubbell, and S. W. Pacala. 2016. Dominance of the suppressed:
Power-law size structure in tropical forests. Science 351:155–157.
Fernandes, K., W. Baethgen, S. Bernardes, R. DeFries, D. G. DeWitt, L. Goddard, W. Lavado,
D. E. Lee, C. Padoch, M. Pinedo-Vasquez, and M. Uriarte. 2011. North Tropical Atlantic
influence on western Amazon fire season variability. Geophysical Research Letters
38:L12701.
Fetcher, N., S. F. Oberbauer, and B. R. Strain. 1985. Vegetation effects on microclimate in
lowland tropical forest in Costa Rica. International Journal of Biometeorology 29:145–
155.
Finegan, B. 1996. Pattern and process in neotropical secondary rain forests: the first 100 years of
succession. Trends in Ecology & Evolution 11:119–124.
Fisher, J. I., G. C. Hurtt, R. Q. Thomas, and J. Q. Chambers. 2008. Clustered disturbances lead to
bias in large-scale estimates based on forest sample plots. Ecology Letters 11:554–563.
Fox, J. W. 2013. The intermediate disturbance hypothesis should be abandoned. Trends in
Ecology & Evolution 28:86–92.
Francis, J. K. 1994. Simple and Inexpensive Method for Extracting Wood Density Samples from
Tropical Hardwoods. Tree Planters’ Notes 45:10–12.
Funk, C., P. Peterson, M. Landsfeld, D. Pedreros, J. Verdin, S. Shukla, G. Husak, J. Rowland, L.
Harrison, A. Hoell, and J. Michaelsen. 2015. The climate hazards infrared precipitation
with stations—a new environmental record for monitoring extremes. Scientific Data
2:150066.
133
García-García, D., and C. C. Ummenhofer. 2015. Multidecadal variability of the continental
precipitation annual amplitude driven by AMO and ENSO. Geophysical Research Letters
42:2014GL062451.
Garstang, M., S. White, H. H. Shugart, and J. Halverson. 1998. Convective cloud downdrafts as
the cause of large blowdowns in the Amazon rainforest. Meteorology and Atmospheric
Physics 67:199–212.
Gentry, A. H. 1988a. Changes in Plant Community Diversity and Floristic Composition on
Environmental and Geographical Gradients. Annals of the Missouri Botanical Garden
75:1–34.
Gentry, A. H. 1988b. Tree species richness of upper Amazonian forests. Proceedings of the
National Academy of Sciences 85:156–159.
Giraudoux, P. 2014. pgirmess: Data analysis in ecology.
Gloor, M., R. J. W. Brienen, D. Galbraith, T. R. Feldpausch, J. Schöngart, J.-L. Guyot, J. C.
Espinoza, J. Lloyd, and O. L. Phillips. 2013. Intensification of the Amazon hydrological
cycle over the last two decades. Geophysical Research Letters 40:1729–1733.
Gloor, M., O. L. Phillips, J. J. Lloyd, S. L. Lewis, Y. Malhi, T. R. Baker, G. López-Gonzalez, J.
Peacock, S. Almeida, A. C. A. De Oliveira, E. Alvarez, I. Amaral, L. Arroyo, G. Aymard,
O. Banki, L. Blanc, D. Bonal, P. Brando, K.-J. Chao, J. Chave, N. Dávila, T. Erwin, J.
Silva, A. Di Fiore, T. R. Feldpausch, A. Freitas, R. Herrera, N. Higuchi, E. Honorio, E.
Jiménez, T. Killeen, W. Laurance, C. Mendoza, A. Monteagudo, A. Andrade, D. Neill,
D. Nepstad, P. N. Vargas, M. C. Peñuela, A. P. Cruz, A. Prieto, N. Pitman, C. Quesada,
R. Salomão, M. Silveira, M. Schwarz, J. Stropp, F. Ramírez, H. Ramírez, A. Rudas, H.
Ter Steege, N. Silva, A. Torres, J. Terborgh, R. Vásquez, And G. Van Der Heijden. 2009.
Does the disturbance hypothesis explain the biomass increase in basin-wide Amazon
forest plot data? Global Change Biology 15:2418–2430.
Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D.
Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini,
C. O. Justice, and J. R. G. Townshend. 2013. High-Resolution Global Maps of 21st-
Century Forest Cover Change. Science 342:850–853.
Hill, M. O. 1973. Diversity and Evenness: A Unifying Notation and Its Consequences. Ecology
54:427–432.
Hobbs, N. T., and M. B. Hooten. 2015. Bayesian Models: A Statistical Primer for Ecologists.
Princeton University Press.
Hsieh, T. C., and K. H. M. and A. Chao. 2015. iNEXT: Interpolation and Extrapolation for
Species Diversity.
Hubbell, S. P. 2001. The Unified Neutral Theory of Biodiversity and Biogeography (MPB-32).
Princeton University Press.
134
Hubbell, S. P., R. B. Foster, S. T. O’Brien, K. E. Harms, R. Condit, B. Wechsler, S. J. Wright,
and S. L. de Lao. 1999. Light-Gap Disturbances, Recruitment Limitation, and Tree
Diversity in a Neotropical Forest. Science 283:554–557.
Jackson, C. 2011. Multi-State Models for Panel Data: The msm Package for R. Journal of
Statistical Software 38.
Kayano, M. T., and R. V. Andreoli. 2007. Relations of South American summer rainfall
interannual variations with the Pacific Decadal Oscillation. International Journal of
Climatology 27:531–540.
Kayano, M. T., and V. B. Capistrano. 2014. How the Atlantic Multidecadal Oscillation (AMO)
modifies the ENSO influence on the South American rainfall. International Journal of
Climatology 34:162–178.
Kellner, J. R., and G. P. Asner. 2009. Convergent structural responses of tropical forests to
diverse disturbance regimes. Ecology Letters 12:887–897.
Kershaw, H. M., and A. U. Mallik. 2013. Predicting Plant Diversity Response to Disturbance:
Applicability of the Intermediate Disturbance Hypothesis and Mass Ratio Hypothesis.
Critical Reviews in Plant Sciences 32:383–395.
Knudsen, M. F., M.-S. Seidenkrantz, B. H. Jacobsen, and A. Kuijpers. 2011. Tracking the
Atlantic Multidecadal Oscillation through the last 8,000 years. Nature Communications
2:178.
Kraft, N. J. B., M. R. Metz, R. S. Condit, and J. Chave. 2010. The relationship between wood
density and mortality in a global tropical forest data set. New Phytologist 188:1124–
1136.
Laurance, W. F., P. Delamônica, S. G. Laurance, H. L. Vasconcelos, and T. E. Lovejoy. 2000.
Conservation: Rainforest fragmentation kills big trees. Nature 404:836–836.
Laurance, W. F., L. V. Ferreira, J. M. Rankin-de Merona, and S. G. Laurance. 1998. Rain forest
fragmentation and the dynamics of Amazonian tree communities. Ecology 79:2032–
2040.
Lichstein, J. W., T. R. Simons, S. A. Shriner, and K. E. Franzreb. 2002. Spatial autocorrelation
and autoregressive models in ecology. Ecological Monographs 72:445–463.
Lindgren, F., H. Rue, and J. Lindström. 2011. An explicit link between Gaussian fields and
Gaussian Markov random fields: the stochastic partial differential equation approach.
Journal of the Royal Statistical Society: Series B (Statistical Methodology) 73:423–498.
Lloyd, J., E. U. Gloor, and S. L. Lewis. 2009. Are the dynamics of tropical forests dominated by
large and rare disturbance events? Ecology Letters 12:E19–E21.
135
Lobova, T. A., S. A. Mori, F. Blanchard, H. Peckham, and P. Charles-Dominique. 2003.
Cecropia as a food resource for bats in French Guiana and the significance of fruit
structure in seed dispersal and longevity. American Journal of Botany 90:388–403.
Lugo, A. E., and F. N. Scatena. 1996. Background and catastrophic tree mortality in tropical
moist, wet, and rain forests. Biotropica 28:585–599.
Lyutsarev, V., and D. Purves. 2013. filzbach: Filzbach MCMC sampler. R package version 0.6-
1.
Malhi, Y., D. Wood, T. R. Baker, J. Wright, O. L. Phillips, T. Cochrane, P. Meir, J. Chave, S.
Almeida, L. Arroyo, N. Higuchi, T. J. Killeen, S. G. Laurance, W. F. Laurance, S. L.
Lewis, A. Monteagudo, D. A. Neill, P. N. Vargas, N. C. A. Pitman, C. A. Quesada, R.
Salomao, J. N. M. Silva, A. T. Lezama, J. Terborgh, R. V. Martinez, and B. Vinceti.
2006. The regional variation of aboveground live biomass in old-growth Amazonian
forests. Global Change Biology 12:1107–1138.
Marra, D. M., J. Q. Chambers, N. Higuchi, S. E. Trumbore, G. H. P. M. Ribeiro, J. dos Santos,
R. I. Negrón-Juárez, B. Reu, and C. Wirth. 2014. Large-scale wind disturbances promote
tree diversity in a central Amazon forest. PLoS ONE 9:e103711.
Marvin, D. C., G. P. Asner, D. E. Knapp, C. B. Anderson, R. E. Martin, F. Sinca, and R.
Tupayachi. 2014. Amazonian landscapes and the bias in field studies of forest structure
and biomass. Proceedings of the National Academy of Sciences 111:E5224–E5232.
Mesquita, R. C. G., K. Ickes, G. Ganade, and G. B. Williamson. 2001. Alternative successional
pathways in the Amazon Basin. Journal of Ecology 89:528–537.
Meyer, V., S. S. Saatchi, J. Chave, J. Dalling, S. Bohlman, G. A. Fricker, C. Robinson, and M.
Neumann. 2013. Detecting tropical forest biomass dynamics from repeated airborne
Lidar measurements. Biogeosciences Discussions 10:1957–1992.
Molino, J.-F., and D. Sabatier. 2001. Tree diversity in tropical rain forests: A validation of the
Intermediate Disturbance Hypothesis. Science 294:1702–1704.
Nakagawa, S., and H. Schielzeth. 2013. A general and simple method for obtaining R2 from
generalized linear mixed-effects models. Methods in Ecology and Evolution 4:133–142.
Nascimento, H. E. M., and W. F. Laurance. 2002. Total aboveground biomass in central
Amazonian rainforests: a landscape-scale study. Forest Ecology and Management
168:311–321.
Negrón-Juárez, R. I., J. Q. Chambers, G. Guimaraes, H. Zeng, C. F. M. Raupp, D. M. Marra, G.
H. P. M. Ribeiro, S. S. Saatchi, B. W. Nelson, and N. Higuchi. 2010. Widespread
Amazon forest tree mortality from a single cross-basin squall line event. Geophysical
Research Letters 37:L16701.
136
Negrón-Juárez, R. I., J. Q. Chambers, D. M. Marra, G. H. P. M. Ribeiro, S. W. Rifai, N. Higuchi,
and D. Roberts. 2011. Detection of subpixel treefall gaps with Landsat imagery in central
Amazon forests. Remote Sensing of Environment 115:3322–3328.
Nelson, B. W., V. Kapos, J. B. Adams, W. J. Oliveira, and O. P. G. Braun. 1994. Forest
disturbance by large blowdowns in the Brazilian Amazon. Ecology 75:853–858.
Nepstad, D. C., I. M. Tohver, D. Ray, P. Moutinho, and G. Cardinot. 2007. Mortality of large
trees and lianas following experimental drought in an Amazon forest. Ecology 88:2259–
2269.
Nepstad, D., P. Lefebvre, U. Lopes da Silva, J. Tomasella, P. Schlesinger, L. Solórzano, P.
Moutinho, D. Ray, and J. Guerreira Benito. 2004. Amazon drought and its implications
for forest flammability and tree growth: a basin-wide analysis. Global Change Biology
10:704–717.
Niklas, K. J. 1992. Plant Biomechanics: An Engineering Approach to Plant Form and Function.
University of Chicago Press.
Nobre, P., and J. Shukla. 1996. Variations of sea surface temperature, wind stress, and rainfall
over the Tropical Atlantic and South America. Journal of Climate 9:2464–2479.
Norden, N., H. A. Angarita, F. Bongers, M. Martínez-Ramos, I. Granzow-de la Cerda, M. Van
Breugel, E. Lebrija-Trejos, J. A. Meave, J. Vandermeer, G. B. Williamson, and others.
2015. Successional dynamics in Neotropical forests are as uncertain as they are
predictable. Proceedings of the National Academy of Sciences 112:8013–8018.
Palace, M., M. Keller, and H. Silva. 2008. Necromass production: Studies in undisturbed and
logged Amazon forests. Ecological Applications 18:873–884.
Pennington, T. D., C. Reynel, and A. Daza. 2004. Illustrated Guide to the Trees of Peru. D. Hunt.
Peterson, C. J. 2007. Consistent influence of tree diameter and species on damage in nine eastern
North America tornado blowdowns. Forest Ecology and Management 250:96–108.
Phillips, O. L., Y. Malhi, N. Higuchi, W. F. Laurance, P. V. Núñez, R. M. Vásquez, S. G.
Laurance, L. V. Ferreira, M. Stern, S. Brown, and J. Grace. 1998. Changes in the Carbon
Balance of Tropical Forests: Evidence from Long-Term Plots. Science 282:439–442.
Plummer, M. 2014. rjags: Bayesian graphical models using MCMC.
Putz, F. E. 1983. Treefall pits and mounds, buried seeds, and the importance of soil disturbance
to pioneer trees on Barro Colorado Island, Panama. Ecology 64:1069–1074.
Quesada, C. A., J. Lloyd, L. O. Anderson, N. M. Fyllas, M. Schwarz, and C. I. Czimczik. 2011.
Soils of Amazonia with particular reference to the RAINFOR sites. Biogeosciences
8:1415–1440.
137
Quesada, C. A., O. L. Phillips, M. Schwarz, C. I. Czimczik, T. R. Baker, S. Patiño, N. M. Fyllas,
M. G. Hodnett, R. Herrera, S. Almeida, E. Alvarez Dávila, A. Arneth, L. Arroyo, K. J.
Chao, N. Dezzeo, T. Erwin, A. di Fiore, N. Higuchi, E. Honorio Coronado, E. M.
Jimenez, T. Killeen, A. T. Lezama, G. Lloyd, G. López-González, F. J. Luizão, Y. Malhi,
A. Monteagudo, D. A. Neill, P. Núñez Vargas, R. Paiva, J. Peacock, M. C. Peñuela, A.
Peña Cruz, N. Pitman, N. Priante Filho, A. Prieto, H. Ramírez, A. Rudas, R. Salomão, A.
J. B. Santos, J. Schmerler, N. Silva, M. Silveira, R. Vásquez, I. Vieira, J. Terborgh, and J.
Lloyd. 2012. Basin-wide variations in Amazon forest structure and function are mediated
by both soils and climate. Biogeosciences 9:2203–2246.
Rich, R. L., L. E. Frelich, and P. B. Reich. 2007. Wind-throw mortality in the southern boreal
forest: effects of species, diameter and stand age. Journal of Ecology 95:1261–1273.
Rifai, S. W., L. O. Anderson, and S. A. Bohlman. In prep. Sea Surface Temperature Anomalies
as the Underlying Cause of Interannual Variation of Catastrophic Wind Disturbance in
the Northwest Amazon.
Rifai, S. W., T. A. P. West, and F. E. Putz. 2015. “Carbon Cowboys” could inflate REDD+
payments through positive measurement bias. Carbon Management 6:151–158.
Roberts D.A., Gardner M., Church R., Ustin S., Scheer G., and Green R.O. 1998. Mapping
chaparral in the Santa Monica Mountains using multiple endmember spectral mixture
models - II. Environmental influences on regional abundance. Remote Sensing of
Environment 65:267–279.
Roberts, D. A., M. O. Smith, and J. B. Adams. 1993. Green vegetation, nonphotosynthetic
vegetation, and soils in AVIRIS data. Remote Sensing of Environment 44:255–269.
Roman, J., R. Knuteson, S. Ackerman, and H. Revercomb. 2015. Predicted changes in the
frequency of extreme precipitable water vapor events. Journal of Climate 28:7057–7070.
Ropelewski, C. F., and M. S. Halpert. 1987. Global and regional scale precipitation patterns
associated with the El Niño/Southern Oscillation. Monthly Weather Review 115:1606–
1626.
Ropelewski, C. F., and P. D. Jones. 1987. An Extension of the Tahiti–Darwin Southern
Oscillation Index. Monthly Weather Review 115:2161–2165.
Rose, S. A. 2000. Seeds, seedlings and gaps - size matters. A study in the tropical rain forest of
Guyana.
Royle, J. A., M. Kéry, R. Gautier, and H. Schmid. 2007. Hierarchical spatial models of
abundance and occurrence from imperfect survey data. Ecological Monographs 77:465–
481.
Rue, H., S. Martino, and N. Chopin. 2009. Approximate Bayesian inference for latent Gaussian
models by using integrated nested Laplace approximations. Journal of the Royal
Statistical Society: Series B (Statistical Methodology) 71:319–392.
138
Rummukainen, M. 2012. Changes in climate and weather extremes in the 21st century. Wiley
Interdisciplinary Reviews: Climate Change 3:115–129.
Running, S. W. 2008. Ecosystem Disturbance, Carbon, and Climate. Science 321:652–653.
Saha, S., S. Moorthi, X. Wu, J. Wang, S. Nadiga, P. Tripp, D. Behringer, Y.-T. Hou, H. Chuang,
M. Iredell, M. Ek, J. Meng, R. Yang, M. P. Mendez, H. van den Dool, Q. Zhang, W.
Wang, M. Chen, and E. Becker. 2014. The NCEP Climate Forecast System Version 2.
Journal of Climate 27:2185–2208.
Schlesinger, M. E., and N. Ramankutty. 1994. An oscillation in the global climate system of
period 65–70 years. Nature 367:723–726.
Schmidt, G., C. Jenkerson, J. Masek, E. Vermote, and F. Gao. 2013. Landsat ecosystem
disturbance adaptive processing system (LEDAPS) algorithm description. Page 27.
USGS Numbered Series, U.S. Geological Survey, Reston, VA.
Shea, K., S. H. Roxburgh, and E. S. J. Rauschert. 2004. Moving from pattern to process:
coexistence mechanisms under intermediate disturbance regimes. Ecology Letters 7:491–
508.
Sheil, D. 2001. Long-term observations of rain forest succession, tree diversity and responses to
disturbance. Plant Ecology 155:183–199.
Sheil, D., and D. F. R. P. Burslem. 2003. Disturbing hypotheses in tropical forests. Trends in
Ecology & Evolution 18:18–26.
Silander, S. R., and A. E. Lugo. 1990. Cecropia peltata L. Silvics of North America.
Silva, C. E., J. R. Kellner, D. B. Clark, and D. A. Clark. 2013. Response of an old-growth
tropical rainforest to transient high temperature and drought. Global Change Biology
19:3423–3434.
Sinton, D. S., J. A. Jones, J. L. Ohmann, and F. J. Swanson. 2000. Windthrow disturbance, forest
composition, and structure in the Bull Run basin, Oregon. Ecology 81:2539–2556.
Sombroek, W. 2001. Spatial and Temporal Patterns of Amazon Rainfall. AMBIO: A Journal of
the Human Environment 30:388–396.
Spiegelhalter, D. J., N. G. Best, B. P. Carlin, and A. van der Linde. 2002. Bayesian measures of
model complexity and fit. Journal of the Royal Statistical Society: Series B (Statistical
Methodology) 64:583–639.
Ter Steege, H., N. C. A. Pitman, O. L. Phillips, J. Chave, D. Sabatier, A. Duque, J.-F. Molino,
M.-F. Prévost, R. Spichiger, H. Castellanos, P. von Hildebrand, and R. Vásquez. 2006.
Continental-scale patterns of canopy tree composition and function across Amazonia.
Nature 443:444–447.
139
Swenson, N. G., and B. J. Enquist. 2008. The relationship between stem and branch wood
specific gravity and the ability of each measure to predict leaf area. American Journal of
Botany 95:516–519.
Tan, J., C. Jakob, W. B. Rossow, and G. Tselioudis. 2015. Increases in tropical rainfall driven by
changes in frequency of organized deep convection. Nature 519:451–454.
Thomas, R. Q., J. R. Kellner, D. B. Clark, and D. R. Peart. 2013. Low mortality in tall tropical
trees. Ecology 94:920–929.
De Toledo, J. J., W. E. Magnusson, C. V. Castilho, and H. E. M. Nascimento. 2011. How much
variation in tree mortality is predicted by soil and topography in Central Amazonia?
Forest Ecology and Management 262:331–338.
De Toledo, J. J., W. E. Magnusson, C. V. Castilho, and H. E. M. Nascimento. 2012. Tree mode
of death in Central Amazonia: Effects of soil and topography on tree mortality associated
with storm disturbances. Forest Ecology and Management 263:253–261.
Trumbore, S., P. Brando, and H. Hartmann. 2015. Forest health and global change. Science
349:814–818.
Turner, I. M. 2001. The Ecology of Trees in the Tropical Rain Forest. Cambridge University
Press.
Urquiza Muñoz, J. D., R. I. Negrón Juárez, F. Ramirez Arevalo, R. Tello, M. Vanderwel, J. W.
Lichstein, J. Q. Chambers, and S. A. Bohlman. in review. Landscape-scale consequences
of differential tree mortality from catastrophic wind disturbance in the Amazon.
Ecological Applications.
Violle, C., Z. Pu, and L. Jiang. 2010. Experimental demonstration of the importance of
competition under disturbance. Proceedings of the National Academy of Sciences
107:12925–12929.
Vitousek, P. M., and J. S. Denslow. 1986. Nitrogen and Phosphorus availability in treefall gaps
of a lowland tropical rainforest. Journal of Ecology 74:1167–1178.
Vittorio, A. V. D., R. I. Negrón-Juárez, N. Higuchi, and J. Q. Chambers. 2014. Tropical forest
carbon balance: effects of field- and satellite-based mortality regimes on the dynamics
and the spatial structure of Central Amazon forest biomass. Environmental Research
Letters 9:034010.
Warmerdam, F. 2008. The Geospatial Data Abstraction Library. Pages 87–104 in G. B. Hall and
M. G. Leahy, editors. Open Source Approaches in Spatial Data Handling. Springer Berlin
Heidelberg.
White, E. P., B. J. Enquist, and J. L. Green. 2008. On estimating the exponent of Power-Law
frequency distributions. Ecology 89:905–912.
140
Williamson, G. B., W. F. Laurance, A. A. Oliveira, P. Delamônica, C. Gascon, T. E. Lovejoy,
and L. Pohl. 2000. Amazonian tree mortality during the 1997 El Niño Drought.
Conservation Biology 14:1538–1542.
Zalamea, P.-C., P. R. Stevenson, S. Madriñán, P.-M. Aubert, and P. Heuret. 2008. Growth
pattern and age determination for Cecropia sciadophylla (Urticaceae). American Journal
of Botany 95:263–271.
Zanne, Amy E., Lopez-Gonzalez, G., Coomes, David A., Ilic, Jugo, Jansen, Steven, Lewis,
Simon L., Miller, Regis B., Swenson, Nathan G., Wiemann, Michael C., and Chave,
Jerome. 2009. Data from: Towards a worldwide wood economics spectrum. Dryad
Digital Repository.
141
BIOGRAPHICAL SKETCH
Sami Rifai was born in Houston, Texas, raised mostly in Rhode Island and spent the last
decade in the southeastern US, but nevertheless he considers California home. He spent much of
his youth in zoos where his mother worked as a zookeeper, where rainforest exhibits instilled
upon him an enduring concern about the rates of tropical deforestation and species extinction. He
learned about forest ecology research when he was at the University of California, Irvine - where
he began the oddly circuitous route of pursuing forest ecology with a B.S. in Earth and
Environmental Science from the Earth System Science Department. At UC Irvine, he began
working in Michael Goulden’s laboratory, where he did not make it to the Amazon as he had
hoped, but rather the cold forests of Manitoba, Canada, and the arid woodland ecosystems of
Southern California. Later, he was rejected for a job to work on stream chemistry data in Brazil,
but was offered the opportunity to do an M.S. at the University of Georgia with Daniel
Markewitz. Sami went to the University of Georgia in 2006 and examined how intensive loblolly
plantation management affected soil microbes and carbon for his M.S. research. He started a
Ph.D. in Ecology and Evolutionary Biology at Tulane University in 2009. With the support of a
NASA Biodiversity grant awarded to Jeff Chambers, he conducted his dissertation research on
the effects of blowdowns in the Peruvian Amazon. There he began to learn firsthand about
tropical forests, the gross income disparities of people in the tropics, and the rapid pace of
deforestation. He transferred to Stephanie Bohlman’s lab group at the University of Florida in
2012, where the cumulative UF academic environment influenced Sami’s thoughts on ecology
and conservation. In 2013, he was awarded a NASA Earth and Space Science Fellowship, and
spent the next three years applying himself to the remote sensing of Amazon forest disturbances.
With interests in ecology, earth system science, and conservation - he intends to continue
researching the landscape ecology of Amazon forests.