joint spatio-temporal analysis of a linear and a ...fw19/waveheights_02_25_14.pdf · ally computed...
TRANSCRIPT
Joint spatio-temporal analysis of a linear and a
directional variable: space-time modeling of wave
heights and wave directions in the Adriatic Sea
Fangpo Wang*, Alan E. Gelfand* and Giovanna Jona-Lasinio†
*Department of Statistical Science, Duke University
*{fw19, alan}@stat.duke.edu
†Department of Statistical Science, University of Rome Sapienza
February 25, 2014
Abstract
It is valuable to have a better understanding of factors that influence sea motion and to provide
more accurate forecasts. In particular, we are motivated by data on wave heights and outgoing wave
directions over a region in the Adriatic sea during the time of a storm, with the overarching goal
of understanding the association between wave directions and wave heights to enable improved
prediction of wave behavior. Our contribution is to develop a fully model-based approach to cap-
ture joint structured spatial and temporal dependence between a linear and an angular variable.
Model fitting is carried out using a suitable data augmented Markov chain Monte Carlo (MCMC)
algorithm. We illustrate with data outputs from a deterministic wave model for a region in the
1
Adriatic Sea. The proposed joint model framework enables both spatial interpolation and temporal
forecasting.
KEYWORDS Angular variable; Bayesian kriging; hierarchical model; latent variables; Markov
chain Monte Carlo; projected Gaussian process; significant wave height
1. INTRODUCTION
Better understanding and more accurate prediction of sea motion might help to reduce risks in
various marine related operations, such as navigation and safety of ships, coastal erosion and oil
spill motion. However, sea motion data are mostly available as the outputs from deterministic
models, especially if we are interested in exploring a relatively broad area of the sea. Recently,
deterministic models based on the dynamics and thermodynamics of the atmosphere have been
used to forecast the weather with increasing reliability. The outputs from such models are usu-
ally computed at several spatial and temporal resolutions. Building upon earlier work (Wang and
Gelfand, 2012) which focused solely upon wave directions, in this paper, we extend our interest
to include wave heights along with outgoing wave directions and provide a framework for joint
modeling these two measurements. More generally, we offer a modeling approach for joint spatial
and spatio-temporal analysis of an angular and a linear variable.
Wang and Gelfand (2012) have dealt with wave directions associated with spatial locations over
time. In fact, our available data include both significant wave heights and outgoing wave directions
(formally defined below) at the same spatial and temporal resolution. Figure 1 displays both the
outgoing wave directions and wave heights over a region of Adriatic sea, for a subset of available
locations, at a particular time point during a storm. The wave heights are displayed in the image
plot, which is an additional layer over the arrow plots of the directions. Perhaps not surprising,
there is strong visual evidence of spatial dependence for both heights and directions, although with
quite different patterns. Rather than separately modeling the wave heights and wave directions, it
is natural to think of building a joint model to accommodate the underlying association between
them. (The analysis in Section 2.4 and Table 2 below, supports this.)
2
The association between wave directions and wave heights might seem to be similar to that say,
between wind directions and wind speeds. However, for wind data, we usually observe the N-S
and W-E components as linear data which induce a direction. We can then add wind speed as a
third linear variable. In the marine application, we observe a height and only a direction. For wind
data, this would be akin to having wind speed and only wind direction. To our knowledge, there
are no such joint models for spatial wave heights and wave directions.
More broadly, the contribution of this paper is to develop a fully model-based approach to cap-
ture joint structured spatial dependence for modeling linear data and directional data at different
spatial locations. We employ a projected Gaussian process for directions and, given direction, a
linear Gaussian process for heights. We show that Bayesian model fitting under such specification
is straightforward using a suitable data augmented Markov chain Monte Carlo (MCMC) algorithm.
This joint model framework allows natural extension to space-time data and can directly incorpo-
rate space-time covariate information, enabling, within its specification, both spatial interpolation
and temporal forecasting.
Wave height, like wind speed, is a linear variable, which can be treated in several ways (Kalnay,
2002; Wilks, 2006; Jona Lasinio et al., 2007). Work on space and space-time models of significant
wave heights can be found in Baxevani et al. (2009) and Ailliot et al. (2011). Again, wave direc-
tions are circular variables, measured in degrees relative to a fixed orientation. There is a relatively
smaller literature on modeling wave directions. Some recent work to build a space and space-time
model of wave directions can be found in Jona Lasinio et al. (2012) and Wang and Gelfand (2012).
Modeling linear and circular variables in a marine context has been considered in a likelihood
framework by Lagona and Picone (2013) and Bulla et al. (2012). These papers focus on classifica-
tion of sea regimes in the Adriatic basin. Classification is based on wind and waves directions and
wind speed and wave heights coming from a coastal meteorological monitoring station along with
the Ancona buoy. A hidden Markov model for the analysis of the time series of bivariate circular
observations (wind and wave directions) is proposed in Lagona and Picone (2013). They assume
the data are sampled from bivariate circular densities, whose parameters are driven by the evolu-
tion of a latent Markov chain. In Bulla et al. (2012), wind and wave data are clustered, again by
3
12 14 16 18 20
4041
4243
4445
April 05 12:00
Longitude
Latit
ude
0
1
2
3
4
5
Figure 1: Plot of wave heights (meters) and wave directions for 200 locations at 12:00 on April5th, 2010
pursuing a hidden Markov approach. Toroidal clusters are defined by a class of bivariate von Mises
densities while skew-elliptical clusters are defined by mixed linear models with positive random
effects. With just a single location, none of this work is spatial.
We often simultaneously observe realizations of a circular variable Θ and a linear variable
X as pairs (θ1, x1), . . . , (θn, xn). In addition to the foregoing examples, Jammalamadaka and
SenGupta (2001, chap. 8.5) suggest pairs such as the direction of departure and the distance to the
destination or wind direction and humidity. In the literature to date, the discussion usually focuses
on conditional modeling. For instance, if the angular variable Θ depends on linear variables, it
is referred to as a Linear-Circular regression model. i.e. a regression specification to use the
4
linear covariates to explain a directional response. Such modeling typically adopts the von Mises
distribution and with a suitable link function which maps Rk to (−π, π), e.g., the arctan function,
where k is the dimension of covariates (Gould, 1969; Laycock, 1975; Johnson and Wehrly, 1978;
Fisher and Lee, 1992). Regression under the general projected normal was proposed in Wang and
Gelfand (2013).
Conditioning in the opposite direction, now using an angular covariate to explain a linear re-
sponse, i.e., if the linear variable X depends on independent angular variables Θ’s, we have a
regression model with a linear response and angular covariates, known as a Circular-Linear re-
gression model. Again, a link function is employed (Mardia, 1976; Mardia and Sutton, 1978;
Johnson and Wehrly, 1978). A flexible approach is to employ trigonometric polynomials (see
Jammalamadaka and SenGupta, 2001).
Measuring association between a circular and a linear variable is not as obvious as that between
two linear variables. Since the angular variable Θ has the support on the circle, the pair (Θ, X)
has the support on a cylinder. If their relationship can be written as X = a + b cos Θ + c sin Θ,
Mardia (1976) proposes a measure using the ordinary multiple correlation in the regression setting
between X and (cos Θ, sin Θ). Similarly, Johnson and Wehrly (1977) discuss the dominant canon-
ical correlation coefficient between X and (cos Θ, sin Θ). The modeling that we offer below is in
the spirit of this representation, of conditioning X on Θ. In particular, we build our joint model
through a conditional times a marginal specification where the marginal specification is a space
or space-time directional data process and then, conditionally, we specify a space or space-time
linear process. The only other spatial work involving a circular-linear regression model we are
aware of is the recent paper of Modlin et al. (2012). They model wind angle and wind speed given
wind angle. They model wind angle as a wrapped normal model introducing conditionally autore-
gressive (CAR) spatial random effects. They model wind speed on the log scale, conditional on
wind angle with a cosine link, again adding CAR spatial random effects. For our setting with wave
directions, we feel a projected Gaussian process is physically more appropriate than a wrapping
model. Furthermore, adding wave height given wave direction we find a natural link function.
The format of this paper is as follows. Section 2 develops a static joint model of spatial wave
5
heights and wave directions, with a data illustration. In Section 3, we consider strategies for spatio-
temporal extension. For example, a dynamic model specification is straightforward but difficult to
fit. A specification that imagines time to be continuous is easier to work with. For the joint space-
time setting, we propose an illustrative model for the process to investigate and compare calm sea
state with stormy sea state. We illustrate using data for the Adriatic sea, off the coast of Italy.
Section 4 offers a summary and future challenges.
2. STATIC JOINT MODELS
In this section, we propose a framework for jointly modeling spatial linear variables and circular
variables. In Section 2.1, we provide the model details. Section 2.2 discusses the model fitting un-
der this specification using data obtained at a collection of spatial locations. Section 2.3 illustrates
the post-model fitting Bayesian kriging within this framework. In Section 2.4, we provide real data
examples for illustration.
2.1 Model specification
We begin with a single linear variable and a single directional variable which, illustratively, we
will refer to as the wave height and wave direction in the sequel. We build a joint parametric model
for the wave height H and the wave direction Θ by introducing a latent variable R (which will be
specified below), in the form
f(H,Θ|Ψh,Ψθ) =
∫f(H,Θ, R|Ψh,Ψθ) dR
=
∫f(H|Θ, R,Ψh)f(Θ, R|Ψθ) dR, (1)
where Ψh and Ψθ are sets of parameters associated with the conditional model for height and the
marginal model for direction, respectively, and are elaborated below.
To briefly review the projected normal model, suppose a random vector Y = (Y1, Y2)T follows
a bivariate normal distribution, with mean µ and covariance matrix Σ. The corresponding random
unit vector U = Y/‖Y‖ is said to follow a circular projected normal distribution (Small, 1996;
6
Mardia and Jupp, 2000) with the same parameters, denoted as PN2(µ,Σ).
Transforming Y (equivalently U) to an angular random variable Θ through Θ = arctan∗(Y2/Y1) =
arctan∗(U2/U1), the density for Θ can be derived (Mardia, 1972, pp. 52). Here arctan∗(S/C) is
formally defined as arctan(S/C) if C > 0, S ≥ 0; π/2 if C = 0, S > 0; arctan(S/C) + π if
C < 0; arctan(S/C) + 2π if C ≥ 0, S < 0; undefined if C = 0, S = 0.
This distribution is almost intractable and suggests that we introduce a latent variable R =
||Y|| and work with the joint distribution of R and Θ, easily obtained through polar coordinate
transformation from the joint distribution of Y1 and Y2. In fact, it takes the form,
f(r, θ|µ,Σ) = (2π)−1|Σ|−12 exp
(−(ru− µ)TΣ−1(ru− µ)
2
)r, (2)
where u = (cos θ, sin θ)T. Wang and Gelfand (2013) note that the general projected normal
distribution is not fully identified; U = Y/‖Y‖ is invariant to scale transformation. In Σ = σ21 ρσ1σ2
ρσ1σ2 σ22
, Wang and Gelfand (2013) set σ1 = τ and σ2 = 1 to ensure identifiability,
resulting in a four-parameter (µ1, µ2, τ , ρ) distribution with Σ = T =
τ 2θ ρτθ
ρτθ 1
.
Next, for location s in the domain of interestD, denote the linear variable (height) by H(s) and
the angular variable (direction) by Θ(s). As the spatial model for Θ(s)|Ψθ, we propose a stationary
projected Gaussian process (Wang and Gelfand, 2012) with a constant mean µ = (µ1, µ2)T and
separable cross-covariance Cθ(s, s′) = %θ(s − s′;φθ) · T , where φθ is the decay parameter asso-
ciated with the correlation function %θ(·) and T is a 2 × 2 matrix defined as T =
τ 2θ ρτθ
ρτθ 1
.
Altogether, we let Ψθ = {µ, T, φθ}.
In particular, as in Wang and Gelfand (2012), the projected Gaussian process is induced from
an inline bivariate Gaussian process Y(s) = (Y1(s), Y2(s))T with a constant mean µ and cross-
covarianceCθ(s, s′). Essentially, we can transform back and forth between the two random variable
spaces (Θ(s), R(s)) and (Y1(s), Y2(s)) through Y1(s) = R(s) cos Θ(s) and Y2(s) = R(s) sin Θ(s),
thus defining the latent process, R(s). As above and as detailed in Wang and Gelfand (2012), this
7
latent R(s) is introduced to facilitate model fitting of the projected Gaussian process. Thus, we
obtain f(Θ, R|Ψθ) under the integral in expression (1) in the same fashion as we obtain expression
(2).
At the top level of the hierarchy in (1), f(H|Θ, R,Ψh) is specified as a univariate Bayesian
spatial regression (a customary geostatistical model) for the wave height H(s)|Θ(s), R(s),Ψh,
with Θ(s) and R(s) included in the mean through a link function g(·). That is,
H(s) = g(Θ(s), R(s)) + w(s) + ε(s).
As usual, the residual is partitioned into the spatial effect term w(s) and the non-spatial error term
ε(s), where w(s) is assumed to follow a zero mean stationary Gaussian process with covariance
function Ch(s − s′) and ε(s)’s are uncorrelated pure errors. Under this conditional specification,
we are actually implementing a regression model with both circular (Θ) and linear (R) covariates.
A natural choice for the link function g(·) will revert to the linear regression on Y1(s) and Y2(s)
from the “unobserved” inline Gaussian process Y(s). Explicitly, we have
H(s) = β0 + β1R(s) cos Θ(s) + β2R(s) sin Θ(s) + w(s) + ε(s) (3)
= β0 + β1Y1(s) + β2Y2(s) + w(s) + ε(s)
= X(s)Tβ + w(s) + ε(s) (4)
where the spatial random effect w(s) follows a zero-centered GP with covariance function Ch =
σ2h%h(s− s′;φh), the error term ε(s)
iid∼ N(0, τ 2h).
In expression (4), we denote the regression coefficient vector β = (β0, β1, β2)T and the lo-
cation specific covariate vector as X(s) = (1, R(s) cos Θ(s), R(s) sin Θ(s))T. Under this model
specification, the parameters associated with the heights are Ψh = {β, φh, σ2h, τ
2h}. Altogether,
(H(s), Y1(s), Y2(s))T specifies a trivariate Gaussian process whose mean structure and cross-covariance
structure, under the above specification, is easily calculated. (See Appendix A.)
The coefficients of the spatial regression, β1 and β2, naturally provide information regarding
8
the association between the circular random variable Θ and the linear variableH under the specific
specification in (3). When β1 and β2 are both 0, fitting the joint model essentially becomes fitting
the wave heights and the wave directions separately as a spatial regression model and a projected
Gaussian process model. Also, we can use the square of the multiple correlation coefficient, R2H|Y,
as a measure of the strength of the conditional dependence (explanation), as is customary in linear
regression. Notably, it is for a regression where the covariates are not observed. However, using
Appendix A, it can be obtained, as a parametric function, from the joint dependence structure and
is free of location. In particular, if ∆ = β21τ
2θ + β2
2 + 2β1β2τθρ,
R2H|Y =
∆
∆ + σ2h + τ 2
h
. (5)
The posterior distribution of R2H|Y can be obtained directly from posterior samples of the parame-
ters after model fitting.
2.2 Fitting
Suppose we have a joint spatial model for (H(s),Θ(s)), as specified in the previous subsection. We
then have the following parameters to estimate: Ψθ = {µ, τ 2θ , ρ, φθ} and Ψh = {β, φh, σ2
h, τ2h}.
We have observations (h,θ), where h = (h(s1), . . . , h(sn))T and θ = (θ(s1), . . . , θ(sn))T. We
need the likelihood, the joint distribution f(h(s1), . . . , h(sn), θ(s1), . . . , θ(sn); Ψθ,Ψh).
Since Θ(s)|Ψθ follows a stationary projected Gaussian process, according to the definition, we
have the corresponding “unobserved” latent linear variable Y1(si) = R(si) cos Θ(si) and Y2(si) =
R(si) sin Θ(si), i = 1, . . . , n. We note that Y1 = (Y (s1), . . . , Y1(sn))T and Y2 = (Y2(s1), . . . , Y2(sn))T,
the realizations of the inline Gaussian process Y = (YT1 ,Y
T2 )T follow a multivariate normal
with mean (µ111×n, µ211×n)T and covariance matrix Σ̃θ = T ⊗ Γθ(φθ), where {Γθ(φθ)}j,k =
%θ(sj − sk;φθ), j, k = 1, . . . , n. For the conditional layer, the model fitting is exactly the same as
that for a customary spatial regression. Therefore, the update of β, τ 2h and σ2
h is rather standard.
We denote Σh = σ2hΓh(φh) + τ 2
hIn, where {Γh(φh)}j,k = %h(sj − sk;φh), j, k = 1, . . . , n.
The full conditional distribution for R(si) is a product of two terms, one from f(Θ, R|Ψθ)
9
and the other from f(H|Θ, R,Ψh). Block updating of the latent vector R is more difficult under
the joint modeling framework than that for just the marginal projected Gaussian process. There-
fore, we resort to conditional updating as R(si)|R(−si),Ψθ,Ψh,θ,h. As described in Wang and
Gelfand (2012), the properties of the inline GP are utilized to obtain the conditional distribution
Y(si)|Y(−si),Ψθ. The details of the posterior computation steps are shown in Appendix B.
To complete the Bayesian model, we specify the priors for the hyperparameters as follows.
For Ψθ, conjugacy for µ arises under a bivariate normal prior, e.g. µ ∼ N2(0, λµI2). For τ 2θ , we
choose the prior as an inverse Gamma IG(aτθ, bτθ) with mean bτθ/(aτθ − 1) = 1 while a uniform
prior on (−1, 1) is used for ρ. For the decay parameter φθ of the exponential covariance function,
we have worked with continuous uniform priors having support which allows small ranges up to
ranges larger than the maximum distance over the region. For the MCMC implementation, the full
conditional distributions are given in Appendix B. The parameters τ 2θ , ρ and φθ are updated using
a Metropolis-Hastings step. For Ψh, conjugacy for β arises under a multivariate Gaussian prior,
e.g., β ∼ N3(0, λβI3). For the decay parameter φh, similarly, we use continuous uniform priors
with support allowing small ranges up to ranges to the maximum distance over the region. As
discussed in Wang and Gelfand (2012), from our limited experience, a projected Gaussian process
seems to require a broader range than a linear process over the same region. As for σ2h and τ 2
h ,
inverse Gamma priors are used, IG(aσh, bσh) and IG(aτh, bτh), respectively.
2.3 Implementing kriging
For prediction at a new location s0, our goal is to obtain a joint distribution of H(s0) and Θ(s0)
given the data (θ,h), expressed as
f(θ(s0), h(s0)|θ,h)) =
∫ ∫ ∫f(θ(s0), h(s0), r(s0)|r,Ψ,θ,h)f(r,Ψ|θ,h) drdr0dΨ,
whereR(s0) is the latent random variable associated with Θ(s0) and the parameters Ψ = {Ψθ,Ψh}.
For kriging of the circular variable at the new location, Θ(s0), we simplify by starting from the
lower level, the joint distribution of Y(s0) = (Y1(s0), Y2(s0))T and Y∗ = (Y1(s1), Y2(s1), . . . , Y1(sn), Y2(sn))T
10
is, Y(s0)
Y∗
∼MVN
µ(s0)
µ∗
,
1 ρT0,Y(φθ)
ρ0,Y(φθ) ΓY(φθ)
⊗ T ,
where µ∗ = (µT(s1), . . . ,µT(sn))T, {ΓY(φθ)}j,k = %θ(sj − sk;φθ) and {ρ0,Y(φθ)}j = %θ(s0 −
sj;φθ), j, k = 1, . . . , n. Thus, the conditional distribution for Y(s0)|Y∗ is a bivariate normal with
mean Es0 and covariance matrix Σs0 , where
Es0 = µ(s0) + ρT
0,Y(φθ)⊗ T · Γ−1Y (φθ)⊗ T−1(Y∗ − µ∗),
Σs0 = (1− ρT
0,Y(φθ)Γ−1Y (φθ)ρ0,Y(φθ))⊗ T.
In fact, the conditional distribution for Θ(s0)|Y∗ is a general projected normal PN2(Es0 ,Σs0).
The latent variable R(si) associated with the i-th location is updated during the model fitting. At
the g-th iteration, we gather the posterior samples of Y∗ through y1(si)(g) = r(si)
(g) cos θ(si) and
y2(si)(g) = r(si)
(g) sin θ(si), i = 1, . . . , n and g = 1, . . . , G. Finally, we are able to draw sam-
ples of Θ(s0) from the predictive distribution f(θ(s0)|θ,h), since there exists an explicit form for
Θ(s0)|Y∗, equivalently, Θ(s0)|Θ,R. As a byproduct of sampling Θ(s0), we also obtain a sample
ofR(s0) at the g-th iteration, r(s0)(g). Let r(g) be the realization ofR = (R(s1), . . . , R(sn))T at the
g-th iteration. We can evaluate f(θ(s0)|r(g),Ψ(g),θ) using a fine grid of points on [0, 2π) and take
the average of the density values on each grid. This resulting average is a usual Rao-Blackwellized
estimate of the predictive density at the kriged location s0. As a mixture of general projected
normal densities, its form is very flexible. (Theorem 1 in Wang and Gelfand (2012) shows that
mixtures of projected normals are dense in the class of all directional data distributions.)
For kriging of the linear variable at the new location, H(s0), again we simplify, now starting
from the joint distribution of H(s0) and H = (H(s1), . . . , H(sn))T,
H(s0)
H
∼MVN
X(s0)Tβ
Xβ
,
τ 2h + σ2
h τ 2hρ
T0,H(φh)
τ 2hρ0,H(φh) τ 2
hΓh(φh) + σ2hIn
,
11
where X(s0) = (1, R(s0) cos Θ(s0), R(s0) sin Θ(s0))T, {Γh(φh)}j,k = %h(sj−sk;φh), {ρ0,H(φh)}j =
%h(s0−sj;φh), j, k = 1, . . . , n and X is a n×3 matrix with the i-th row as (1, R(si) cos Θ(si), R(si) sin Θ(si)),
i = 1, . . . , n. At the g-th iteration, the posterior samples, θ(s0)(g) and r(s0)(g), are included in
x(s0)(g) as the covariates. The remaining kriging steps for the linear variable H(s0) are standard,
and thus we do not provide further details here.
2.4 A data example
We study wave heights and directions in the Adriatic Sea, a semi-enclosed basin (Picone, 2013).
The orography of the Adriatic Sea plays a key role in any study of associated wave behavior as
most of the waves tend to travel from north-northwest in a south-easterly direction along the major
axis of the basin.
The wave height data we employ in this paper are significant wave heights. The significant
wave height (SWH) is essentially the average height of the highest waves in a wave group. The
traditional definition of the significant wave height, denoted as Hs, is the average of the one third
highest wave heights (trough to crest) observed in the region. Recently, a more commonly used
definition is four times the standard deviation of the vertical displacement of the sea surface (eleva-
tion), which is calculated on the basis of the wave spectrum. Corresponding to the latter definition,
a notation Hm0 is often used. With regard to directions, outgoing wave directions are often pre-
ferred to incoming and are recorded in degrees relative to a fixed orientation.
Monitored buoy data, supplying wave height and wave direction measurements would seem to
be attractive for such analysis. However, at present, buoy networks are too sparse to be used as a
data source for spatial analysis. Therefore, we employ an alternative data source for wave heights
and wave directions, outputs from deterministic models, usually climatic forecasts computed at
several spatial and temporal resolutions. These models are increasingly accurate and may be
eventually calibrated using buoy data, a data fusion problem. However, here we illustrate solely
with deterministic model output.
We employ the foregoing general modeling framework, illustrating with deterministic model
output. In particular, we use data outputs from a deterministic wave model implemented by ISPRA
12
(Istituto Superiore per la Protezione e la Ricerca Ambientale) using the Hydro-Meteorological-
Marine System (‘Sistema Idro-Meteo-Mare’, SIMM) (Speranza et al., 2004, 2007). This integrated
forecasting system includes a wind forecast model (BOLAM) predicting the surface wind (wind
at 10 meters above the sea surface) over the entire Mediterranean and wave models (in deep and
shallow waters).
The deep water WAve Model (WAM) describes the evolution of sea wave spectra by solving
energy transport equations, using as input only the wind forecast produced from the BOLAM.
Wave spectra are locally modified using a source function describing the wind energy, the energy
redistribution due to nonlinear wave interactions, and energy dissipation due to wave fracture. The
WAM produces significant wave height, mean wave directions (always computed over one hour
time window), and the peak wave period. The model provides forecasts over a grid with 10 × 10
km cells every hour. The ISPRA dataset has forecasts on a total of 4941 grid points over the Italian
Mediterranean. Over the Adriatic Sea area, we have 1494 points.
According to meteorologists, the definition of “calm” in the Adriatic Sea refers to a time period
with wave height lower than 1 meter. When the wave height is between 1 and 2 meters, a pre-
storm warning is triggered and above 2 meters there is a storm warning. We plot the data, wave
directions and wave heights, for an illustrative hour under each sea state, in Figures 2(a)-(c), with
arrow plots for the wave directions and image plot (color) for the wave heights. We apply the joint
modeling framework proposed in Section 2.1 separately to each of the three datasets, respectively
representing calm, transition and storm in the Adriatic Sea, as shown in Figures 2(a)-(c). In the
absence of covariates, we employ the joint model in Section 2.1.
We model the heights directly. With concern about negative predictions, we also fitted our
model with log wave heights. The results do not change; negative predictions are infrequent, even
in the calm state. We also note that the stationarity assumption is likely unrealistic for coastal data
due to anticipated orographic influences; our model specification is an initial attempt.
As above, there are 1494 locations (grid points) in the Adriatic Sea and we randomly select
300 (approximately 20%) locations for the purpose of validation. We note that the maximum
interpoint distance over set of locations is 850 km. For the fitting, we run the MCMC algorithm for
13
12 14 16 18 20
4041
4243
4445
April 02 06:00
Longitude
Latit
ude
0
1
2
3
4
5
12 14 16 18 20
4041
4243
4445
April 05 07:00
Longitude
Latit
ude
0
1
2
3
4
5
(a) calm (b) transition
12 14 16 18 20
4041
4243
4445
April 05 18:00
Longitude
Latit
ude
0
1
2
3
4
5
(c) storm
Figure 2: Plot of wave heights (meters) and wave directions under three sea states: (a) calm, (b)transition and (c) storm
25000 iterations with a burn-in of 5000 and thinning by collecting every fifth sample. The kriging
estimates are computed using 4000 posterior samples.
The prior setting for Ψθ is as follows: µ ∼ N(0, 4I), τ 2θ ∼ IG(2, 1) and ρ ∼ Unif(−1, 1).
A continuous uniform prior Unif(0.0004, 0.1) is used for φθ, corresponding to a maximum and
minimum range of 7500 km and 30 km respectively. In fact, the strong spatial dependence for the
14
directions under storm conditions suggest that we may require a range beyond the largest pairwise
distance in the dataset when there is little variation in wave directions. For the parameters associ-
ated with the heights Ψh = {β, τ 2h , σ
2h, φh}, a continuous uniform prior Unif(0.005, 0.5) is used for
φh, corresponding to a maximum and minimum range of 600 km and 60 km respectively. Since φh
is the decay parameter in the regular linear spatial regression model, we can adopt customary pri-
ors for ranges with linear variables. We employ a non-informative prior for β as β ∼ N3(0, 10I).
Inverse gamma priors are used for τ 2h and for σ2
h, again IG(2, 1). These inverse gamma priors are
weak in that they have infinite variance with mean 1. We do not find evidence of much sensitivity
to their centering.
We compare our results with those based upon independently modeling the spatial wave height
and the spatial wave direction, a nested case of our joint modeling framework obtained by setting
β1 and β2 to be zero. In the independence case, the marginal model for wave heights is a spatial
regression with constant mean β0, written as
H(s) = β0 + w(s) + ε(s), w(s) ∼ GP (0, σ2h%h(s− s′;φh)), ε(s)
iid∼ N(0, τ 2h),
where the parameters set is Ψhm = {β0, φh, σ2h, τ
2h}. The marginal model for wave directions
becomes a constant mean projected Gaussian process model.
The posterior summaries of parameters are shown in Table 1, for both the joint model and the
independence model above. The 95% credible intervals are obtained as approximate highest pos-
terior density (HPD) intervals following Chen and Shao (1999). Model comparison between the
joint model and the independence model is provided in Table 2. Regarding choice of model com-
parison criterion, for circular variables we utilize the continuous ranked probability score (CRPS)
and the predictive log scoring loss (PLSL), following the discussion in Wang and Gelfand (2013).
For the linear variables, predictive mean square error and average length of 95% credible intervals
are evaluated for each model.
During a calm day, we observe small values of wave heights (lower than 1 meter) and large
variation in wave directions over the region of interest. Conversely, at a time point during the storm,
15
Table 1: Posterior summaries of the parameters: the joint model (H,Θ) and independent modelsH , Θ
parameter mean lower upper mean lower upper
data (Calm) (H , Θ) H , Θ
β0 0.2446 0.1965 0.2950 0.2331 0.1859 0.2809β1 0.0125 -0.0209 0.0471 - - -β2 0.0653 0.0501 0.0796 - - -σ2h 0.0051 0.0046 0.0056 0.0055 0.0051 0.0060τ2h 0.0011 0.0010 0.0013 0.0011 0.0010 0.0013φh 0.0101 0.0084 0.0126 0.0126 0.0105 0.0116φθ 0.0020 0.0019 0.0022 0.0019 0.0017 0.0022µ1 0.2856 -0.4046 0.9755 0.3039 -0.4310 1.0000µ2 -0.5055 -1.7615 0.8260 -0.4141 -1.6671 0.9952τ2θ 0.2717 0.2275 0.3137 0.2747 0.2388 0.3180ρ -0.4874 -0.5713 -0.4268 -0.4790 -0.5451 -0.4078
data (Transition) (H , Θ) H , Θ
β0 0.4660 0.3244 0.6176 0.5394 0.3685 0.6892β1 0.1887 0.1281 0.2453 - - -β2 0.3062 0.2478 0.3706 - - -σ2h 0.0344 0.0306 0.0389 0.0413 0.0376 0.0482τ2h 0.0021 0.0019 0.0024 0.0021 0.0018 0.0023φh 0.0101 0.0084 0.0126 0.0084 0.0072 0.0090φθ 0.0006 0.0005 0.0007 0.0006 0.0005 0.0007µ1 -0.1066 -2.3975 2.1270 0.0617 -2.0922 2.2151µ2 -0.2935 -1.7928 1.3506 -0.2152 -1.8580 1.3850τ2θ 2.4400 2.0958 2.8055 2.2318 1.9602 2.5818ρ -0.2264 -0.3069 -0.1473 -0.2257 -0.2994 -0.1430
data (Storm) (H, Θ) H , Θ
β0 0.3795 0.1897 0.5776 0.4737 0.0734 0.8790β1 -0.3052 -0.4049 -0.2066 - - -β2 0.8794 0.7588 0.9883 - - -σ2h 0.0702 0.0546 0.0826 0.1930 0.1677 0.2283τ2h 0.0033 0.0028 0.0038 0.0034 0.0028 0.0039φh 0.0105 0.0087 0.0122 0.0070 0.0063 0.0076φθ 0.0006 0.0005 0.0007 0.0006 0.0005 0.0007µ1 0.5853 -0.8418 1.8745 0.6990 -0.5824 2.0369µ2 -0.6064 -2.0949 0.9097 -0.5152 -2.0546 0.9437τ2θ 0.7216 0.6515 0.7917 0.7312 0.6652 0.8070ρ -0.6433 -0.6827 -0.6018 -0.6631 -0.7107 -0.6105
we observe that wave directions are fairly homogeneous while there is a large variation in wave
heights over the region (ranging from 1 meter up to 6 meters). During the transition period, there is
considerable variation both in wave heights and wave directions. From Table 1, the posterior means
16
Table 2: Model comparison: the joint model (H,Θ) and independent models H , Θ
(a) calm (b) transition (c) stormFeature (H,Θ) H , Θ (H,Θ) H , Θ (H,Θ) H , Θ
Predictive Mean Square Error (height) 0.0006 0.0006 0.0031 0.0030 0.0109 0.0108Average Length of 95% Credible Interval (height) 0.1821 0.1788 0.3405 0.3586 0.5163 0.6345
mean CRPS for wave direction 0.0407 0.0408 0.0276 0.0279 0.0213 0.0223PLSL for wave direction -977 -974 -1098 -1104 -1321 -1318
of the spatial variance component for wave heights under the joint model (σ2h) are 0.0051, 0.0344
and 0.0702 for the three different datasets. Since the contribution of τ 2h is relatively negligible,
these differences across time points agree with our remark regarding increasing variation of wave
heights from calm to storm.
For the wave directions, there is no single parameter which can capture the variation across the
region. Specifically, it is difficult to glean much from the posterior summaries of the parameters of
the marginal projected normal distribution since the four parameters of a general projected normal
altogether determine the shapes in a complex fashion. However, the decay parameter φθ associated
with the spatial correlation kernel of the projected Gaussian process provide some indication of the
smoothness of the surface of directions. For the calm day, φθ is much larger than in the other two
cases, indicating less smoothness in this surface, as expected.
With prediction as an objective, we can consider the predictive densities of wave height and
wave direction at kriged locations. For illustration, the results for one selected location at the three
different sea states are shown in Figure 3. During the calm period, the joint model does not show
much improvement in predictions in terms of concentration of the predictive densities. During the
transition period, the difference between the joint model and independent models becomes a bit
more clear; both of the predictive densities in Figure 3(b) show that the joint model outperforms
the independent one. Finally, in Figure 3(c), we see considerable benefit to the joint model.
Turning to the model comparison results in Table 2, during the calm days, the model compari-
son criteria find the performance of the two models to be essentially equivalent. For the transition
case, the joint model emerges as slightly better, with shorter average length of 95% credible in-
tervals and slightly smaller average CPRS and PLSL. For the time point chosen during a storm,
17
−1.0 −0.5 0.0 0.5 1.0
0.0
0.5
1.0
1.5
2.0
2.5
3.0
θ
dens
ity
0.5 0.6 0.7 0.8
02
46
8
h
dens
ity
(a) calm
0.2 0.4 0.6 0.8 1.0
02
46
8
θ
dens
ity
1.2 1.4 1.6 1.8 2.0 2.2
01
23
4
h
dens
ity
(b) transition
1.9 2.0 2.1 2.2 2.3
05
1015
2025
3035
θ
dens
ity
4.5 5.0 5.5 6.0 6.5 7.0
0.0
0.5
1.0
1.5
2.0
h
dens
ity
(c) storm
Figure 3: Predictive densities of directions (left) and heights (right) at one kriged location duringdifferent sea states: the solid line is for the joint model (H,Θ), the dashed line is for the indepen-dent model H,Θ, and the vertical line shows the held-out observation
18
the biggest gain is achieved. We have a substantially narrower average length of credible intervals
and a roughly 5% smaller value of CRPS. Such model comparison results are supported by the
predictive density examination in Figure 3.
As mentioned previously, the independence model is nested in the joint model framework. The
posterior summaries of β1 and β2 in Table 1 concur with these findings. For the time point during
a calm day, β1 and β2 are nearly zero, which supports the independence of wave heights and wave
directions. As we move from calm to transition to storm, the β1 and β2 become increasingly differ-
ent from 0. The multiple correlation coefficient is computed, as in (5) and we show the posterior
densities of the multiple correlation coefficient for all three scenarios in Figure 4. The figure con-
cisely supports Table 1, revealing stronger association between height and direction as we move
from calm to storm. Finally, we note that the negative β1 and positive β2 are not unexpected. With
storms tending to travel from the north-northwest to the southeast, and with the sin increasing and
cos decreasing in these directions, this supplies an expectation of an increase in wave height as a
function of θ as the storm develops.
3. JOINT SPACE-TIME MODELS
In Section 2, we focused on joint modeling of spatial wave directions and wave heights at a static
time slice and showed the benefit from jointly modeling these two variables, especially during
a storm. It is natural to envision an underlying process for heights and directions in continuous
space and time; fitting such a model would enable both interpolation and forecasting at future time
points. Since such data are also available across time at hourly resolution using the ISPRA output,
we consider such a process model.
To see the behavior of wave heights over the region, we select 10 illustrative locations, shown
in upper panel of Figure 5. Their corresponding time series of hourly heights across the first ten
days in April, 2010 are plotted in the lower panel of Figure 5. Following the foregoing definition
of sea motion states based on significant wave heights, we observe a range of behavior, including
a transition from calm to storm and back to calm.
19
0.0 0.2 0.4 0.6 0.8 1.0
05
1015
2025
3035
the square of the multiple correlation coefficient
Den
sity
calmtransitionstorm
Figure 4: The posterior densities of the square of the multiple correlation coefficient R2H|Y
So, now, denote the wave height and wave direction measurements at location s and time t
as H(s, t) and Θ(s, t). In Section 3.1, we offer a space-time model for the pair (H(s, t),Θ(s, t))
under the continuous space-time setting. For illustration, we use two portions of the ISPRA output,
a calm window and a storm window (highlighted in Figure 5).
3.1 Model specification
We have the wave height H(s, t) and the wave direction Θ(s, t) both over the space-time domain,
s ∈ D and t ∈ (0, K). We adopt the same joint framework shown in expression (1), where Ψh and
Ψθ are sets of parameters associated with the conditional model for height and the marginal model
for direction, respectively.
Again, this joint space-time model is provided conditionally. For the marginal distribution
of the circular variables Θ(s, t)|Ψθ, we propose a stationary spatio-temporal projected Gaussian
20
12 14 16 18 20
4041
4243
4445
46
Longitude
Latit
ude
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●●
●
●
●
01
23
45
6
Wave heights over time at specific locations
Date
Hei
ghts
in m
eter
s
04/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/1004/01 04/02 04/03 04/04 04/05 04/06 04/07 04/08 04/09 04/10
Figure 5: Time series of wave heights (lower panel) for 10 illustrative locations (upper panel) fromn = 200 in the Adriatic Sea region in April 2010. The two time windows (calm and stormy) usedin Section 3.3 are highlighted.
21
process with a constant mean µ = (µ1, µ2)T and separable cross-covariance form,
Cθ ((s, t), (s′, t′)) = %θ,s(s− s′;φθ,s)%θ,t(t− t′;φθ,t) · T, (6)
where %θ,s is the spatial correlation and %θ,t is the temporal correlation. The parameters associated
with the directions Ψθ = (µ, T, φθ,s, φθ,t), φθ,s and φθ,t are the decay parameters associated with
their corresponding correlation functions %θ,s and %θ,t and T is still defined as in the previous
section, T =
τ 2θ ρτθ
ρτθ 1
. This model has been considered in Wang and Gelfand (2012).
As in the static case, we only need one space-time covariance function. For simplicity, we
assume separability in space and time in expression (6). Future work will have us investigating
space-time dependent covariance functions such as those in Gneiting (2002) and Stein (2005). Un-
der this model specification, the projected Gaussian process is induced from an inline Gaussian
process Y(s, t) with a constant mean µ and cross-covariance Cθ((s, t), (s′, t′)). So, we can trans-
form back and forth between the spaces for the pairs of random variables, (Θ(s, t), R(s, t)) and
(Y1(s, t), Y2(s, t)) through Y1(s, t) = R(s, t) cos Θ(s, t) and Y2(s, t) = R(s, t) sin Θ(s, t). A latent
variable R(s, t) has again been introduced to facilitate the model fitting.
At the conditional level, we introduce
H(s, t) = X(s, t)Tβ + w(s, t) + ε(s, t)
= β0 + β1Y1(s, t) + β2Y2(s, t) + w(s, t) + ε(s, t).(7)
In fact, we simplify w(s, t) to w(s) where, again, the spatial random effect w(s) follows a
zero-centered GP with covariance function Ch = σ2h%h(s − s′;φh) and the random the error term
ε(s, t)iid∼ N(0, τ 2
h). Now the parameters associated with the heights are Ψh = {β, φh, σ2h, τ
2h}.
This model asserts that the temporal dependence between H(s, t) is fully contributed by the latent
components Y1(s, t) and Y2(s, t). Restoring w(s, t) and perhaps allowing time dependent coeffi-
cients introduces identifiability challenges and thus is deferred to future consideration. Note, from
the previous section, that the coefficient vector, β, changes with sea state. Since, for now, we are
22
not considering dynamics in β, it is only sensible to fit the model in (7) to data in either a calm
window or a storm window.
3.2 Fitting, kriging and forecasting
The model fitting for the joint space-time model is straightforward and only requires some modifi-
cation of the fitting for the static case with details in Section 2.2. We now have observations h,θ at
a collection of locations s1, . . . , sn and a collection of time points t1, . . . , tk, where h = {h(si, tj)}
and θ = {θ(si, tj)}, where i = 1, . . . , n and j = 1, . . . , k. The parameters we need to estimate
are Ψh = {β, φh, σ2h, τ
2h} and Ψθ = (µ, T, φθ,s, φθ,t). In addition, we have to update the latent
R(si, tj), where i = 1, . . . , n and j = 1, . . . , k. Again, we update R(si, tj) conditioning on all the
other latent variables R(−(si, tj)) by collecting the relevant terms in the likelihood, from both the
conditional, f(H|Θ, R,Ψh), and marginal, f(Θ, R|Ψθ), specifications.
Kriging of both measurements at a new location s0 at any of the observed time points is stan-
dard; we omit the details here. Instead, we focus on the one-step ahead prediction to the time point
tk+1 for the n locations with observations. We start from the joint distribution,
Ytk+1
Yt1:tk
∼MVN
µtk+1
µt1:tk
,Γk+1(φθ,t)⊗ Γn(φθ,s)⊗ T
,
where Ytk+1= (Y1(s1, tk+1), Y2(s1, tk+1), . . . , Y1(sn, tk+1), Y2(sn, tk+1)T, Yt1:tk = (YT
t1, . . . ,YT
tk)T,
Ytj = (Y1(s1, tj), Y2(s1, tj), . . . , Y1(sn, tj), Y2(sn, tj))T, Γk+1(φθ,t) = {%θ,t(tj − tq;φθ,t)}, j, q =
1, . . . , k + 1 and Γn(φθ,s) = {%θ,s(si − sp;φθ,s)}, i, p = 1, . . . , n. Thus, we obtain the conditional
distribution Ytk+1|Yt1:tk ,Ψθ from which the posterior samples of Ytk+1
can be easily obtained at
each iteration. Then, the posterior sampling of the wave height at a future time point H(si, tk+1)
remains the same as that in the static case.
3.3 Data examples
For illustration, we randomly select a small fraction of the locations (n = 200) from the same data
set in Section 2.4. These locations are those in both Figure 1 and Figure 5. Recalling the definition
23
of calm and storm in Section 2.4, Figure 5 also provides information regarding the time intervals
of calm and storm. Our goal here is to illustrate the prediction at a future time point tk+1 within
each interval, using our joint space-time model.
First, we select data from a calm period, April 2nd to April 3rd, 2010. The time points are
collected every two hours. Thus starting from April 2nd, 00:00 to April 3rd 22:00, we have 24
time points in total. In addition, we select data from a severe storm period, April 5th to April 6th.
Again, the time points are collected at the same temporal resolution, every two hours. Thus starting
from April 5th, 00:00 to April 6th 22:00, with 24 time points in total. In both cases, we hold out
the observation at the last time point for forecasting validation.
The posterior summaries of the regression coefficients for the conditional specification are
reported in Table 3. During the calm period, β1 and β2 are nearly zero, agreeing with the results
for a single time slice during calm period shown in Section 2.4. This suggests that approximate
independence of wave heights and wave directions sustains during a calm period. However, during
a storm period, β1 and β2 clearly depart from zero, indicating strong dependence of wave heights
and wave directions during the storm, again consistent with the results in Section 2.4. We see that,
again, σ2h dominates the variation in heights and we see more than a twenty-fold increase during
the storm period. With regard to the decay parameters, φθ,s is roughly double in calm compared
with storm, i.e., the spatial range is roughly half. On the other hand, φθ,t is roughly half in calm
compared with storm, i.e., the temporal range is roughly double.
Lastly, we are interested in the result of prediction at the future time point tk+1. Table 4 pro-
vides the summaries of prediction for both wave heights and wave directions. We do not use any
information at time tk+1 to fit the model; however, the model provides reasonable prediction. We
predict direction better in a storm (expected since there is less variability in direction during a
storm) and height dramatically better during a calm period (again, expected because heights are
lower and less variable when it is calm), in agreement with the results from Table 3.
As a last comment here, the reader might wonder why, with time observed discretely, we
didn’t consider a dynamic model specification. In fact, we have investigated such forms. The
implicit challenge lies in the transition specification. That is, the marginal distributions of the
24
Table 3: Posterior summaries of the parameters (mean and HPD)(calm) (storm)
parameter mean lower upper mean lower upperβ0 0.1612 0.0563 0.2998 0.5483 -0.0218 1.1750β1 -0.0117 -0.0388 0.0161 -0.1439 -0.2805 -0.0171β2 0.0329 0.0112 0.0563 0.2061 0.0650 0.3529σ2h 0.3561 0.2916 0.4315 8.2797 6.8463 10.0983τ2h 0.0020 0.0017 0.0024 0.0091 0.0064 0.0119φh 0.0051 0.0048 0.0056 0.0047 0.0043 0.0051µ1 0.2999 -0.2808 0.8955 0.1835 -0.6192 0.9448µ2 -0.1431 -0.9733 0.6035 -0.0868 -0.9079 0.6537τ2 0.6129 0.5665 0.6448 0.9245 0.8879 0.9972ρ -0.7095 -0.7359 -0.6878 -0.6251 -0.6656 -0.5905φθ,s 0.0043 0.0037 0.0050 0.0022 0.0020 0.0024φθ,t 0.0775 0.0750 0.0803 0.1667 0.1487 0.1782
Table 4: Summaries for one-step ahead predictioncalm storm
Predictive Mean Square Error (height) 0.0522 0.4497mean CRPS for wave direction 0.0973 0.0647
directions under the projected normal have shapes that depend jointly, in a complex way, on all four
parameters in µ and T . We would need to introduce a four dimensional joint dynamic specification
for them; dynamics in just say, µ are not enough. Then, we would have to address dynamics in
the parameters for the conditional specification for heights given directions. Altogether, we found
such second stage vector autoregressive models very difficult to specify, even more difficult to fit,
especially with shorter time series. Moreover, if these parameters are only indexed by time, then
we will still have space-time separability.
4. SUMMARY AND FUTURE WORK
In this paper, we have proposed a general framework for jointly modeling spatially indexed linear
variables and circular variables. Also, we have discussed the extension to the space-time setting.
We have focused on introducing the joint modeling, without much attention to covariates. In
particular, for wave data, wind information would be a potentially useful covariate. Also, if the
region is large enough, a trend surface might be worthwhile to consider. However, it may be
25
challenging to interpret the trend through the projection.
Much more work can be done for the joint space-time modeling of directional data and linear
data. Working within a calm period or a storm period, we can consider space-time dependence as
suggested in Section 3.1, that is space-time dependence in the bivariate Gaussian process inducing
space-time dependence in the projected Gaussian process. More attractive, and potentially more
useful, would be a model across time that would accommodate calm, storm, and transition. To
do so with regard to heights would require, at the least, introduction of a β0(t) process, i.e., a
time-dependent intercept. In fact, we would also want time-dependent slope parameters, β1(t) and
β2(t). Anticipating dependence between slopes and intercept, this leads to a three dimensional
process over time. Another aspect to consider is the fact that variability in height changes over sea
state. This suggest that we also add σ2h(t), perhaps as a log Gaussian process. (Of course, range
might be time dependent, adding a further model feature.) Then, there is the matter of dynamics
in the space-time projected Gaussian process. Here, we would want to extend the specification
in Section 3.1 to allow continuous time in µ and T . We encounter similar difficulties to those
raised at the end of the previous section. All told, there are clearly lots of challenging modeling
opportunities.
Included in the output of the deterministic models is mean wave period and peak wave period.
The mean wave period, Tm, is the mean of all wave periods in time series representing a certain
sea state, as opposed to peak wave period, Tp, which is the wave period with the highest energy.
Also, a trend surface might also be worthwhile to introduce. It would be useful to investigate these
additional variables with regard to improved interpretation and prediction.
Lastly, the reader will also appreciate that we can extend our joint modeling approach to handle
multiple linear and directional variables in space and in space and time. Following our condition-
ing approach, with say p directional variables we would specify a marginal 2p dimensional linear
Gaussian process to project to a p dimensional directional process. Then, with say r linear vari-
ables, we would need an r dimensional Gaussian process, appropriately conditioned on the 2p
dimensional process. It is easy to imagine such specifications but model fitting will become ex-
tremely demanding.
26
ACKNOWLEDGMENT
The work of the first two authors was supported in part by NSF CMG 0934595 and NSF CDI
0940671. The authors thank ISPRA (Istituto Superiore per la Protezione e la Ricerca Ambientale)
for the use of data output from the wave model of its SIMM hydro-meteo-marine forecasting
system.
APPENDIX
A: Mean and cross-covariance structures of (H(s), Y1(s), Y2(s))T
Recall from Section 2.1, (H(s), Y1(s), Y2(s))T follows a trivariate Gaussian process. The mean
structure of this Gaussian process is (β0 + β1µ1 + β2µ2, µ1, µ2)T and the cross-covariance is
CH,Y(s, s′) = %θ(s− s′;φθ)
v11 v12 v13
v21T
v31
+ %h(s− s′;φh)
σ2h + τ 2
h1s=s′ 01×2
02×1 02×2
,
where v11 = β21τ
2θ + 2β1β2ρτθ + β2
2 , v12 = v21 = β1τ2θ + β2ρτθ, v13 = v31 = β1ρτθ + β2 and
T =
τ 2θ ρτθ
ρτθ 1
.
B: Posterior Computation: The full conditionals of Ψh and Ψθ
Recall from Section 2.2, Ψh = {β, φh, σ2h, τ
2h} and Ψθ = {µ, T, φθ}, where T =
τ 2θ ρτθ
ρτθ 1
.
We define two covariance matrices Σ̃θ = T ⊗Γθ(φθ) and Σ̃∗θ = Γθ(φθ)⊗T . Their correspond-
ing precision matrices are Q̃θ = Σ̃−1θ = T−1 ⊗ Γ−1
θ (φθ) and Q̃∗θ = Σ̃∗−1
θ = Γ−1θ (φθ) ⊗ T−1. Γ−1
θ,ij
denotes the element on the i-th row and j-th column of the inverse matrix of Γθ(φθ).
Following expression (3), the top or conditional hierarchy can be written as,
H(s) = β0 +R(s)[β1 cos Θ(s) + β2 sin Θ(s)] + w(s) + ε(s),
= β0 +R(s)β4(s) + w(s) + ε(s).
27
As defined in Section 2.2, Σh = σ2hΓ(φh) + τ 2
hIn. The inverse matrix of Σh is denoted as Qh, with
Qh,ij representing its element on the i-th row and j-th column. We then denote H̃(s) = H(s)−β0,
thus H̃(s) = R(s)β4(s) +w(s) + ε(s). The conditional distribution of H̃(si) given H̃(−si) can be
easily obtained as a normal distribution, with R(si) involved in the mean.
The full conditionals for the parameters of the joint process model:
• µ ∼ N2(Eµ, Q−1µ ), where the precision matrix of this bivariate normal Qµ = ATΣ̃−1
θ A +
I2/λµ, Eµ = Q−1µ ATΣ̃−1
θ Y and A =
1n×1 0n×1
0n×1 1n×1
.
•
R(si) ∝ riI(0,∞)(ri)
exp
(−
(Γ−1θ,iiu
Ti T−1ui + β2
4(si))r2i − 2ri(Γ
−1θ,iiu
Ti T−1Esi + β4(si)Esi2)
2
),
where ui = (cos θ(si), sin θ(si))T, Esi = µ − Q−1
si
∑j 6=i Q̃
∗θ,ij(Y(sj) − µ),Qsi = Q̃∗θ,ii =
Γ−1θ,ii · T−1, Esi2 = h̃(si) +Q−1
h,ii
∑j 6=iQh,ij[h̃(sj)− rjβ4(sj)] and Q̃∗θ,ij = Γ−1
θ,ij · T−1.
• φθ ∝ |Γθ(φθ)|−1exp(−(Y − Aµ)TT−1 ⊗ Γ−1
θ (φθ)(Y − Aµ)/2).
•
τ 2θ , ρ ∝ |T |−n/2exp
(−(Y − Aµ)TT−1 ⊗ Γ−1
θ (φθ)(Y − Aµ)/2)
(τ 2θ )−aτθ−1exp(−bτθ/τ 2
θ ).
• β ∼ N3(Eβ, Q−1β ), where the precision matrix of the multivariate normal Qβ = XTΣ−1
h X,
Eβ = Q−1β XTΣ−1
h h, h = (h(s1), . . . , h(sn))T, X is a n × 3 matrix with the i-th row as
(1, r(si) cos θ(si), r(si) sin θ(si)), i = 1, . . . , n.
• φh ∝ |Σh|−1/2exp(− (h−Xβ)TΣ−1
h (h−Xβ)
2
), where Σh = τ 2
hΓ−1h (φh) + σ2
hIn.
28
•
τ 2h , σ
2h ∝ |Σh|−1/2exp
(−(h−Xβ)TΣ−1
h (h−Xβ)
2
)(τ 2h)−aτh−1
exp(−bτh/τ 2h)(σ2
h)−aσh−1exp(−bσh/τ 2
h).
REFERENCES
Ailliot, P., Baxevani, A., Cuzol, A., Monbet, V., and Raillard, N. (2011), “Space-time Models for
Moving Fields with an Application to Significant Wave Height Fields,” Environmetrics, 22, 354–
369.
Baxevani, A., Caires, S., and Rychlik, I. (2009), “Spatio-temporal Statistical Modelling of Signifi-
cant Wave Height,” Environmetrics, 20, 14–31.
Bulla, J., Lagona, F., Maruotti, A., and Picone, M. (2012), “A Multivariate Hidden Markov Model
for the Identification of Sea Regimes from Incomplete Skewed and Circular Time Series,” Jour-
nal of Agricultural, Biological, and Environmental Statistics, 17(4), 544–567.
Chen, M.-H., and Shao, Q.-M. (1999), “Monte Carlo Estimation of Bayesian Credible and HPD
Intervals,” Journal of Computational and Graphical Statistics, 8(1), 69–92.
Fisher, N. I., and Lee, A. J. (1992), “Regression Models for an Angular Response,” Biometrics,
48, 665–677.
Gneiting, T. (2002), “Nonseparable, Stationary Covariance Functions for Space-Time Data,” Jour-
nal of the American Statistical Association, 97, 590–600.
Gould, A. L. (1969), “A Regression Technique for Angular Variates,” Biometrics, 25(4), 683–700.
Jammalamadaka, S. R., and SenGupta, A. (2001), Topics in Circular Statistics World Scientific.
Johnson, R. A., and Wehrly, T. E. (1977), “Measures and Models for Angular Correlation and
Angular-Linear Correlation,” Journal of the Royal Statistical Society. Series B, 39(2), 222–229.
29
Johnson, R. A., and Wehrly, T. E. (1978), “Some Angular-Linear Distributions and related regres-
sion models,” Journal of the American Statistical Association, 73(363), 602–606.
Jona Lasinio, G., Gelfand, A. E., and Jona Lasinio, M. (2012), “Spatial Analysis of Wave Direction
Data Using Wrapped Gaussian Processes,” The Annals of Applied Statistics, 6(4), 1349–1997.
Jona Lasinio, G., Orasi, A., Divino, F., and Conti, P. L. (2007), “Statistical Contributions to the
Analysis of Environmental Risks along the Coastline,” in Societá Italiana di Statistica - rischio
e previsione, Venezia: CLEUP, pp. 255–262.
Kalnay, E. (2002), Atmospheric Modeling, Data Assimilation and Predictability Cambridge Uni-
versity Press.
Lagona, F., and Picone, M. (2013), “Maximum Likelihood Estimation of Bivariate Circular Hidden
Markov Models from Incomplete Data,” Journal of Statistical Computation and Simulation,
83, 1223–1237.
Laycock, P. (1975), “Optimal Design: Regression Models for Directions,” Biometrika, 62, 305–
311.
Mardia, K. V. (1972), Statistics of Directional Data Academic Press (London and New York).
Mardia, K. V. (1976), “Linear-Circular Correlation Coefficients and Rhythmometry,” Biometrika,
63(2), 403–405.
Mardia, K. V., and Jupp, P. E. (2000), Directional Statistics John Wiley & Sons.
Mardia, K. V., and Sutton, T. W. (1978), “A model for cylindrical variables with applications,”
Journal of the Royal Statistical Society. Series B (Methodological), 40, 229–233.
Modlin, D., Fuentes, M., and Reich, B. (2012), “Circular Conditional Autoregressive Modeling of
Vector Fields,” Environmetrics, 23, 46–53.
Picone, M. (2013), Model Based Clustering of Mixed Linear and Circular Data, PhD thesis, Roma
Tre University.
30
Small, C. G. (1996), The Statistical Theory of Shape Springer.
Speranza, A., Accadia, C., Casaioli, M., Mariani, S., Monacelli, G., Inghilesi, R., Tartaglione, N.,
Ruti, P. M., Carillo, A., Bargagli, A., Pisacane, G., Valentinotti, F., and Lavagnini, A. (2004),
“POSEIDON: An integrated system for analysis and forecast of hydrological, meteorological
and surface marine fields in the Mediterranean area,” Nuovo Cimento C, 27, 329–345.
Speranza, A., Accadia, C., Mariani, S., Casaioli, M., Tartaglione, N., G. Monacelli, P. M. R., and
Lavagnini, A. (2007), “SIMM: An integrated forecasting system for the Mediterranean area,”
Meteorol. Appl., 14(4), 337–350.
Stein, M. L. (2005), “Space-Time Covariance Functions,” Journal of the American Statistical As-
sociation, 100, 310–321.
Wang, F., and Gelfand, A. E. (2012), Modeling Space and Space-time Directional Data Using
Projected Gaussian Processes„ Technical report, Duke University.
Wang, F., and Gelfand, A. E. (2013), “Directional Data Analysis under the General Projected
Normal Distribution,” Statistical Methodology, 10(1), 113–127.
Wilks, D. S. (2006), “Comparison of Ensemble-MOS Methods in the Lorenz ’96 Setting,” Meteo-
rological Applications, 13, 243–256.
31