Poster Session


The poster session will be held from 2:30-4:30 on Friday, October 4 in SAS Hall.  A prize will be awarded for the best poster presentation.  


Participants

Earvin Balderama, North Carolina State University

Ashley Bell, Colorado School of Mines

Lisa Bramer, Iowa State University

Jenny Brynjarsdottir, Samsi and Duke University

Won Chang, Pennsylvania State University

Rob Erhardt, Wake Forest University

Miranda Fix, University of Washington

Dan Fortin, Iowa State University

Josh Goldstein, Penn State University

Matt Heaton, The National Center for Atmospheric Research

Karen Kazor, Colorado School of Mines

Eric Laflamme, University of New Hampshire

Myoungji Lee, Texas A & M University

Wendy Meiring, UC-Santa Barbara

Ryan Parker, North Carolina State University

Harrison Quick, University of Minnesota

Katarina Sucic, North Carolina State University

Ying Sun, University of Chicago

Maria A. Terres, Duke University

Grant Weller, Colorado State University

Zhen Zhang, Michigan State University
https://mail.google.com/mail/u/1/images/cleardot.gif



Abstracts

Earvin Balderama (North Carolina State University)

Space-time branching models of alien red bananas

Epidemic-type aftershock sequence (ETAS) models are widely used in seismology to simulate, decluster, and analyze earthquake occurrences. Balderama et al. (2012) fitted a modied version of the ETAS model to sightings of an invasive species of red banana trees (Musa velutina) that spread in a Costa Rican rainforest. The formulation of the triggering density function, which describes the way events cause future occurrences of events, was based on plots of inter-event times and distances for the red banana plants. Here, we assess the triggering density more carefully using a non-parametric stochastic declustering method based on Marsan and Lengline (2008). When the algorithm is applied to the red banana data, the results indicate more intense clustering, especially at smaller spatial-temporal scales, compared to previous estimates, as well as triggering of o spring running primarily to the northwest and the southeast from each parent. The results are also used to obtain estimates of the most likely targets where immigration of red banana plants may be occurring.



Ashley Bell (Colorado School of Mines)

Automated Statistical Quality Control Methods for Atmospheric Radiosonde Data

Radiosondes are an instrument attached to a weather balloon that records variables such as temperature, wind, and geoponential height as it rises. Some archives, like the Upper Air Network Archive, contain an expansive historical set of these observations. Though the spatial (over 2,000 stations) and temporal (beginning in 1920) coverage of this archive is large, the observations have no measure that indicates how trustworthy they are. With proper quality control (QC), this data has the potential to broadly impact many ongoing climate studies. A previous method (Durre et al, 2008), has utilized Lanzante's biweight estimator of mean and standard deviation (1996) to QC historical thermal data using z-scores in a two-step method. The first step estimates a global mean and standard deviation and removes values with an exceptionally large z-scores. The second step estimates mean and standard deviation for 45 day windows centered on a particular day and 3 hour bin and removes values exceeding a threshold. We propose to use Huber's robust estimator of mean and standard deviation in place of Lanzante's estimator. This estimator provides a two-sided estimate of standard deviation and should aid in more accurately detecting outliers in skewed distributions.  The distribution of radiosonde temperatures can be skewed, particularly at lower pressure levels. To compare these two estimators, a rigorous simulation study has been developed to mimic real radiosonde temperatures in various climates. Outliers are introduced to the simulated data, and the QC methods are implemented. Logistic regression is used to determine which method maximizes the number of true outliers captured and minimizes the false positives, and preliminary results indicate that the Huber estimator performs better than the Lanzante.


Lisa Bramer (Iowa State University)

A Model for the Bias of Meteorological Wind Speed and Direction Forecasts  

The ability to accurately and precisely forecast wind speeds has become increasingly important as more wind power is introduced into electricity markets. Physical forecast models most often produce medium and long-range forecasts that are necessary for energy trading but are dependent on unknown initial/boundary conditions and planetary boundary layer (PBL) schemes, leading to the presence of systematic biases in forecasts. Statistical and physical forecast methods can be combined to generate more accurate wind speed forecasts. We generate 54-hour ahead predictions for a location in central Iowa using several ensembles and model the bias of these ensembles using a hierarchical statistical structure.  While observed wind speeds at a height of 10 meters are typically used to represent what happens at 80 meters, observed wind speeds at 80 meters are used here to assess the skill of ensembles and bias corrected models. Additionally, wind direction is strongly related to wind speed in the Midwest.  The incorporation of wind direction in a model is not straightforward, as its domain is on the unit circle.  Here we also present a method for using wind direction in our bias correction model.


Jenny Brynjarsdottir (Samsi and Duke University)

Dimension reduction for two space-time processes - Downscaling temperatures over Antarctica

The field of spatial and spatio-temporal statistics is increasingly faced with the challenge of very large datasets. The classical approach to spatial and spatio-temporal modeling is extremely computationally expensive when the datasets are large. Dimension-reduced modeling approach has proved to be effective in such situations. In this talk we focus on the problem of modeling two spatio-temporal processes where the primary goal is to predict one process from the other and where the datasets for both processes are large. We outline a general dimension-reduced Bayesian hierarchical approach where the spatial structures of both processes are modeled in terms of a low number of basis vectors, hence reducing the spatial dimension of the problem. The temporal evolution of the spatio-temporal processes and their dependence is then modeled through the coefficients (also called amplitudes) of the basis vectors. We present a new method of obtaining data-dependent basis vectors that are geared to the goal of predicting one process from the other: (Orthogonal) Maximum Covariance Patterns. We apply these methods to a statistical downscaling example, where surface temperatures on a coarse grid over the Antarctic are downscaled onto a finer grid.


Won Chang (Pennsylvania State University)

A dimension reduction approach to climate model calibration

Climate models are often used to make projections about future climate. These models are complex and involve unknown parameters. We can learn about these parameters using observations of the climate system and output from the computer model at various parameter settings. When the climate model output is high-dimensional and spatial, parameter inference ("calibration") can be computationally challenging. I describe a new method for parameter inference using Gaussian process models and a dimension reduction approach. I show its applicability with real and simulated examples. My methods allow for a study of the effects of data aggregation by permitting, for the first time, calibration based on 1-D, 2-D and 3-D spatial output and observations in the context of calibration for the climate model parameter called vertical diffusivity. This is joint work with Murali Haran, Roman Olson and Klaus Keller.



Rob Erhardt (Wake Forest University)

Statistical Downscaling of Precipitation on a Spatially Dependent Network Using Regional Climate Models 

We present a detailed simulation methodology for generating rainfall events incorporating future climate change scenarios on a network of meteorological stations.  These events can serve as an input for a hydrological model to study the impacts of climate change on water flow.  To simulate rainfall on a future day t, the method uses the K-nearest neighbor algorithm to identify close neighbors from the historical record, and resamples a past rainfall event with resampling probabilities determined from climate model output.  This preserves the spatial dependence of precipitation on the network, and other important distributional features of rainfall.  Large numbers of arbitrary tuning parameters and model assumptions are avoided through the use of a multivariate Gaussian model relating climate model output to historical rainfall.  The downscaling is based on a regional climate model (RCM) embedded within a general circulation model.  The method is demonstrated using rainfall in North Carolina and the Community Climate Systems Model as a GCM with the Canadian Regional Climate Model as an embedded RCM.



Miranda Fix, E. David Ford, Danaan DeNeve Weeks (University of Washington)

Comparing methods to quantify understory light intensity and spatial variation in a Tsuga heterophylla stand on the Olympic Peninsula.

Spatiotemporal variation of understory light may have great importance for seedling regeneration. Quantifying this heterogeneity can provide target light conditions for silvicultural management as well as improving spatially-explicit models to predict forest dynamics. Digital hemispherical photography is commonly used to estimate the proportion of light reaching the understory, however previous findings indicate that it may overestimate diffuse light. Since diffuse light contributes the greater proportion of the light available for photosynthesis in our temperate rainforest system, we are developing and testing a novel measurement system using the BF3 Sunshine Sensor (Delta-T Devices, Cambridge, UK) which logs instantaneous readings of both Total and Diffuse PPFD (photosynthetic photon flux density). We will compare two methods to quantify the spatial variation of light in a naturally regenerated Tsuga heterophylla stand on the Western Olympic Peninsula in Washington. In the first method, we densely sample our forest plot using digital hemispherical photography to create light maps at different scales. In the second method, we have constructed a motorized tram system to move a BF3 sensor along line transects, while other BF3 sensors are stationed above and below the canopy. We will present sampling designs, preliminary results and geostatistical analysis.



Dan Fortin (Iowa State University)

Principal component model for spatially dependent functional data. 

Estimation of principal component functions is a challenging problem, but when the curves belong to a reproducing kernel Hilbert space estimation of the principal component functions via an eigenvalue decomposition of the covariance function are known to have excellent convergence rates.  These methods have only been developed for independent data, but we investigate how weights based on point location intensity can be used to adjust for spatial autocorrelation in the estimated covariance function.  The resulting covariance function estimate allows for closed form estimates of the eigenfunctions, which are used for a finite basis representation of the observed functional data. The coefficients are assumed to be realizations from a multivariate random field and multivariate techniques are leveraged to model spatial dependence.



Josh Goldstein  (Penn State University)

Spatial Point Process Modeling of Viral Infections

Biologists are often interested in investigating the progression of viral infections. I will describe a parametric spatial point process model for modeling the response of epithelial cells to infections with the human respiratory syncytial virus. The likelihood function for the point process model has an intractable normalizing function which makes likelihood-based inference very challenging. I develop a Bayesian inferential approach for this model via Markov chain Monte Carlo (MCMC). MCMC for this problem is non-trivial so I will describe an algorithm based on the double Metropolis-Hastings algorithm following Liang (2010). Our computational approach enables full inference for the parameters and simplifies checking goodness-of-fit. Preliminary results from both real and simulated data suggest that our model captures the complex attractive and repulsive behavior of this process.  This is joint work with Ivan Simeonov, John Fricks, Murali Haran and Francesca Chiaromonte 


Matt Heaton (The National Center for Atmospheric Research)

Efforts in Calibrating the Lyon-Fedder-Mobarry Magnetosphere-Ionosphere Coupled Computer Model

The Lyon-Fedder-Mobarry global magnetosphere-ionosphere coupled (LFM-MIX) model is a computer model used to study Sun-Earth interactions by simulating aspects of geomagnetic storms. Given a set of input values (initial conditions) and solar wind data, the LFM-MIX model numerically solves the magnetohydrodynamic equations and outputs a bivariate high-dimensional spatio-temporal field of ionospheric energy and energy-flux. A unique aspect of the LFM-MIX model is the ability to run it at multiple fidelities where higher fidelities require more computation but, ultimately, more accurately capture the governing physical processes under study. Of particular interest here are the input settings (and their uncertainties) for the LFM-MIX model such that the corresponding high-fidelity output most closely matches field observations. This work provides an overview of efforts to relate (i.e. calibrate) the output from the LFM-MIX model run at different fidelities to field observations of ionospheric energy and energy-flux by estimating ideal initial conditions. 



Karen Kazor and Amanda Hering (Colorado School of Mines)

Statistical Methods to Identify Regional Wind Regimes

A number of efficiencies in harnessing wind energy can be realized by accurately and objectively identifying predominant regional wind patterns. For example, applications extend to forecasting, synthetic data generation, turbine deployment, and turbine siting. Methods used to identify these regimes in the past have either been subjective in nature or founded on assumptions to which wind data does not adhere. In particular, the meteorological variables recorded are typically not independent, and observations are correlated through time. Thus, we seek to test the effectiveness of clustering methods under scenarios that reflect realistic characteristics of wind data. To do so, data is generated based on the annual trends observed at two different meteorological sites in the Pacific Northwest to create two distinct wind regimes in terms of the atmospheric pressure and wind u and v-components observed at these locations. A number of scenarios in which 1) errors follow either a normal or a skew-t distribution 2) variables are either jointly or independently distributed, and 3) observations are either independent or follow a first-order (vector)autoregressive process, are generated for testing. The performance of an algorithm developed by Benaglia et al. (2009) for the non-parametric estimation of multivariate mixtures is evaluated with the simulated data.  Ultimately, the best clustering method as determined by the simulation will be applied to wind data observed at each of 20 meteorological towers in the Pacific Northwest.



Eric Laflamme,  Yibin Pan,  Ernst Linder (University of New Hampshire)

Downscaling High Resolution Precipitation from Regional Climate Model Output

There is a great societal interest in assessing the impacts of projected climate change on  infrastructure design, such as of dams, bridges and coastal roads.  Such impact assessment  typically requires future projections of high-resolution time series of temperature, precipitation,  solar radiation and related variables.  We propose a methodology based on the work of Kallache  et al. (2011) which applies a CDF transfer function to precipitation extremes and NARCCAP  regional climate model output to develop local-scale projections.  This method has been applied  to 58 locations throughout New England to help assess climate change impact.  The procedure  has been successfully validated across all stations, and results are robust against differences in  regional model selection.



Myoungji Lee (Texas A & M University)

Local Properties of Irregularly Observed Gaussian Fields

The behavior of a variogram at the origin is important in spatial interpolation as it determines the smoothness of a random process. The power-law variogram at the origin is capable of capturing the local behaviors of many known parametric covariance functions and self-similar processes. In addition, the fractal dimension, a scale invariant measure of surface roughness, is a function of the exponent in the power-law model. The estimation of the fractal dimension using the empirical variogram and least squares method has been well studied for a stationary Gaussian process, observed at evenly spaced sampling locations.   This talk presents the estimation of the fractal dimension for observations at unevenly spaced sites by approximating restricted likelihood, conditioning on observations at neighboring sites only. Finite sample performances of the estimates from the approximated and full restricted likelihoods are compared through mean squared errors and Godambe's information measure. The numerical and asymptotic study shows that the locally approximated likelihood is computationally inexpensive yet capturing the local behavior of the variogram well, without requiring the specification of the full variogram. This is a joint work with Michael Stein in the University of Chicago.


Wendy Meiring (UC-Santa Barbara), Petrutza C. Caragea (Iowa State), Jeganathan Chockalingam ( University of Southampton), Peter Atkinson ( University of Southampton)

Modeling Space-Time Variation in the satellite-derived chlorophyll index over India

We study variation in plant growth cycles over a region of India, using the Multi-temporal Medium Resolution Imaging Spectrometer (MERIS) Terrestrial Chlorophyll Index (MTCI) data. Space-time variation in vegetation cycles may be summarized by space-time patterns in the frequency and timing of the Onset of Greenness, Peak of Greenness, and End of Senescence measures within each growth cycle. The frequency and length of each growth cycle differs with vegetation type and climatic changes. In this poster we highlight several challenging space-time estimation problems posed by the diverse vegetation and topography in India, combined with the large size of the satellite data set.


Ryan Parker (North Carolina State University)

Computation in Block Composite Likelihoods with Extensions to Non-Gaussian Response

Recently, block composite likelihood methods have been developed to approximate the likelihood of spatial models having a Gaussian response. These methods block observational sites based on location, and they only consider dependence between neighboring blocks of locations. This blocking scheme allows one to reduce necessary computations to estimate model parameters while maintaining statistical efficiency. We present the computational details for fitting these models with maximum likelihood using Fisher scoring. These methods are being implemented in a forthcoming R package, spacious, that will allow the user to quickly estimate spatial model parameters for large data sets using this block composite structure. In addition, we illustrate how to extend the block composite likelihood to a non-Gaussian continuous response using a Gaussian copula framework. This allows one to consider, for example, a generalized extreme value (GEV) response within a block composite likelihood framework.




Harrison Quick, Sudipto Banerjee and Bradley P. Carlin (University of Minnesota)

Assessing temporal rates of change in California asthma hospitalization data using areally referenced Gaussian processes with region-specific variances

According to the CDC, asthma affects nearly 25 million Americans, with rates increasing over the past 30 years. In our previous work, we investigated 18 years of monthly asthma hospitalization rates in the 58 counties of California using a mean square differentiable temporal process embedded within a dynamic Markov random field framework. We used this approach to evaluate temporal gradients where we were concerned with temporal changes in the residual and fitted rate curves after accounting for several key risk factors. Despite the rise in prevalence, we detected a generally decreasing trend in hospitalization rates statewide, and our gradient estimates revealed progressively muted transitions into and out of the summer months. While our model accounted for both spatial and temporal association, a single variance parameter controlled the variability of the underlying spatiotemporal process. However, the assumption of a single variance parameter may not be valid due to a number of factors, such as differences in population and regional variability in weather patterns. Here, we extend our model to allow for region-specific variance components using an approach similar to the spatially adaptive CAR (SACAR) model (Reich and Hodges, 2008), which we believe can simultaneously help avoid over- and undersmoothing in both our spatiotemporal process and our temporal gradient process. After demonstrating the effectiveness of our new model via simulation, we reanalyze the asthma hospitalization data and compare our findings to those from previous work.


Katarina Sucic (North Carolina State University)

Exploring Sea Surface Salinity through a Spatio-Temporal Context

Recently NASA has launched a new large-scale program using (AQUARIUS) satellites to measure sea surface salinity worldwide.  This is an ambitious project and is an exciting source of data that has thus far been lacking in the scientific community.  Understanding salinity is vital to understanding changes in climate, detecting changes caused by hurricanes, and predicting the magnitude of El Niño and La Niña. The aim of our pilot study is to assess extremes and quantiles of salinity under a spatio-temporal context.  We must find an efficient way to block space and time to simultaneously determine how salinity changes over time for one location, as well as for numerous locations (up to the entire spatial domain).


Ying Sun (University of Chicago) and Montserrat Fuentes (North Carolina State University)

Fused Lasso for Spatial and Temporal Quantile Function Estimation

Quantile functions are important in characterizing the entire probability distribution of a random variable, especially when the tail of a skewed distribution is of interest. This article introduces new temporal and spatial quantile function estimators for spatio-temporal data with a fused Lasso penalty to accommodate the dependence in time and space. The fused Lasso penalty enforces sparsity in both the quantiles and the differences among neighboring quantiles. This method borrows neighboring information for temporal and spatial quantile function estimation, hence it is desirable for applications with features ordered in time or space without replicated observations. To better characterize local dependence, the temporal and spatial fused adaptive Lasso estimators are also proposed by introducing adaptive weights to the fused Lasso penalty. Then, the asymptotic properties are investigated theoretically and the performance is justified by simulations. Our methodology is applied to particulate matter (PM 2.5) data from the Community Multiscale Air Quality (CMAQ) model to characterize the upper quantile values which are crucial to studying spatial association between PM 2.5 concentrations and adverse human health effects.


Maria A. Terres and Alan E. Gelfand (Duke University)

First Flowering Events in Cherry Trees using Survival Models with Spatio-Temporal Covariates

First flowering events in cherry trees are believed to be closely related to temperature patterns during the winter and spring months. Earlier works have incorporated the idea of temperature thresholds, defining chill and heat functions based on these thresholds. However, selection of the thresholds is often arbitrary and shared across species and locations. We propose a survival model with spatially and temporally varying covariates having functional forms representing chill and heat accumulation leading up to first flowering events. The models are fit with a variety of possible chill and heat thresholds in order to select the most appropriate threshold pair. We utilize Cumulative Rank Probability Scores, selecting the threshold pair that minimizes the difference between the predicted cumulative probability curve and the observed cumulative probability curves. We first apply the model using temporally varying covariates to analyze 29 years of flowering data for four cherry species (Cerasus sp.) grown at Mt. Takao in Tokyo, Japan. This allows us to investigate the idea that different species’ relationships with temperature may not be identical; it may instead vary between earlier and later flowering species. Next, the model is applied using spatially and temporally varying covariates to analyze 52 years of flowering data for 45 Cerasus spachiana x C. speciosa trees grown across Japan’s Honshu island. By exploring flowering dates across locations we can explore how the relationship between temperature and first flowering events varies through space.  


Grant Weller and Daniel Cooley (Colorado State University)

A New Representation for Hidden Regular Variation in Multivariate Extremes

Multivariate extreme value methodology has been used to assess risk due to air pollution, coastal flooding, and drought, to name a few.  However, classical modeling techniques in multivariate extremes rely on an assumption of asymptotic dependence, a feature that does not hold even for a bivariate normal distribution with any correlation less than one.  In this case, the classical methodology is unable to distinguish asymptotic independence from exact independence.  A refinement on this deficiency in the heavy-tailed case is provided by the concept of hidden regular variation (Resnick, 2007).  In this work, we propose a representation for a random vector which exhibits hidden regular variation as the sum of independent random vectors with disjoint supports.  We provide asymptotic justification for our representation, which is more amenable to finite-sample estimation than previous characterizations.  Simulation methods are demonstrated, and we present a framework for estimation from this characterization via the EM algorithm.


Zhen Zhang (Michigan State University) 

https://mail.google.com/mail/u/1/images/cleardot.gifA Bayesian Hierarchical Model for Spatial and Functional Clustering in Climate Change Study

The infrared spectral signatures of climate change have been conceptually adopted and widely applied to identifying and attributing atmospheric composition change as developed by NASA. In this work, we will establish a Bayesian hierarchical model for spatial and functional clustering of the climate model data, which are used as surrogates for measured spectra to assess climate change. Our model allows spatiotemporal dependence structures and time-dependent covariates with cluster-specific fixed effect functions that are regularized using wavelet basis. Non-informative priors are employed for the clustering prior model and the covariance structure of random effect functions, and multiple shrinkage priors for fixed effects are adopted and compared by simulation studies. Dimension reduction is achieved by assuming conditional independence between clusters for random effect functions. The model is applied to the spectral signatures of climate change that were observed globally, and produces spatial clustering map that is compared with traditional clustering techniques and variations of the proposed model. We also accommodate the model for high-dimensional data by utilizing high-performance parallel computing designs and efficient algorithms for sparse, block-banded matrices.