Spatial Modeling, Applied Stochastics and Paleoclimatology: Spatial Modeling

Speaker(s)

Montserrat Fuentes (North Carolina State University)
Robert Lund (Clemson University)
Dale Zimmerman (University of Iowa)
Yong Wang (Eastern Kentucky University)
Stefano Castruccio (University of Chicago)
Joe Guinness (University of Chicago)

Description

Environmental science, climatology, and other emerging fields of great societal importance, have in common that they require sound mathematical and statistical tools, in order to properly draw inference for the physical process and accordingly quantify the uncertainty. In this session, a variety of new techniques currently being developed in spatial data modeling and stochastic analysis will be presented, and novel case studies in climatology will be illustrated. The presentations will include probabilists,statisticians, and applied mathematicians, who use the tools of their trade to address practical questions in atmospheric and climate science. Our session will foster discussions on how best to integrate these tools for specific topics including paleoclimate reconstruction, spatial downscaling, and sound stochastic modeling and inference in atmospheric science.

Schedule

Fri, June 22 - Location: STEW 214

Time	Speaker	Title

1:30 - 2:55PM	Montserrat Fuentes	Impact of Climate Change on Human Health using Calibrated Atmospheric Models
	Abstract: Studies on the health impacts of climate change routinely use climate model output as future exposure projection. Uncertainty quantification, usually in the form of sensitivity analysis, has focused predominantly on the variability that arises from different emission scenarios or multi-model ensembles. This study presents a Bayesian spatial quantile regression approach to calibrate climate model output and estimate the risk of premature death due to future heat waves. Specifically, we first estimate the spatial quantile process for climate model output using nonlinear monotonic regression during a historical period. The quantile process is then calibrated using quantile functions estimated from the observed monitoring data. Our model also down-scales the gridded climate model output to the point-level for projecting future exposure over a specific geographical region. The quantile regression approach is motivated by the need to better characterize the tails of future temperature distributions where the greatest health impacts are likely to occur. We apply the methodology to calibrate temperature projections from a regional climate model for the period 2041 to 2050. Accounting for calibration uncertainty, we calculate the number of of excess deaths attributed to future temperature in the US. This is joint work with J. Zhu, H. Chang and J. Davis.
	Robert Lund	Categorical Changepoints: Changes in North Atlantic Basin Hurricanes
	Abstract: This talks overviews changes in the North Atlantic Basin tropical cyclone record since 1851. A changepoint test is first developed for categorical data via maximums of statistics that have marginal chi-squared distributions. Statistically, the asymptotic distribution of the test statistic is shown to involve the supremum of sums of squares of scaled Brownian bridge processes. The results are used to identify times of change (lack of homogeneity) in the North Atlantic Basin tropical cyclone record. We find an increase in storm counts in the 1930s that is generally attributed to the onset of aircraft surveillance. We also find a very significant increase in counts circa 1995 that does not seem to be linked to any changes in data collection techniques. The good news is that no recent changes in the individual storm strengths are seen. The end conclusions are essentially the opposite of recent testimonial given in the United States Senate.
	Dale Zimmerman	On Model-based Design of a Sampling Network for Multivariate Geostatistics
	Abstract: Most of the literature on spatial network design considers univariate spatial inference problems, such as kriging and variogram estimation. This presentation considers model-based frequentist design for multivariate spatial inference problems, in particular co-kriging and estimation of correlation/cross-correlation functions. Some examples are used to illustrate the characteristics of good designs for these purposes. How much efficiency is lost by restricting to collocated designs is explored, and some theoretical results relating optimal collocated designs to optimal univariate designs are given.
3:00-3:30PM	Break
3:30 - 4:55PM	Yong Wang	A semi-parametric approach to multivariate spatial modeling
	Abstract: Building a multivariate spatial model is a challenging task since it involves a great amount of work in terms of formulation and computation. Most existing multivariate spatial models take a parametric approach and as a result are limited in their applicability. We propose a semi-parametric approach to multivariate spatial modeling which offers great flexibility and much improved predictive performance. This modeling approach uses parametric covariograms in the marginal models and a semi-parametric cross-covariogram. A simple assumption of the cross-covariance structure guarantees that the validity of the resulting multivariate covariogram. The semi-parametric feature of the cross-covariogram gives this modeling approach great flexibility to accommodate any form of marginal covariogram. Also, this approach has demonstrated superior predictive performance over several popular existing multivariate spatial models through a series of simulation and real data examples.
	Stefano Castruccio	Space time global models for climate ensembles
	Abstract: Climate models are mathematical models aimed at reproducing physical processes on a global scale and at predicting quantities like temperature and precipitation given some forcing inputs such as CO2 concentration. Climate ensembles are collection of such runs with different initial physical conditions and different forcing scenarios. We have recently developed an approach to statistically model the output to reproduce (emulate) the elements in the ensemble based on a nonlinear regression with the past trajectories of CO2 concentrations as the covariate. The emulator was built on a coarse space resolution, without accounting for any spatial dependence and focusing only on the mean structure. The purpose of this work is to build a statistical model that addresses the issue of emulating space/time dependence at grid resolution, thus leading to a more precise assessment of the model uncertainty. Given the large size of the data, fitting these models requires fast algorithms for gridded data on the sphere x time domain and efficient ways of computing without storing very large matrices. The presence of independent repetitions in the climate runs based on different initial conditions leads to situations specific to computer output analysis; for example, in this setting diagnostic tools such as the variogram can be evaluated without any bias even in the presence of a spatial trend. This is joint work with M.I. Stein
	Joe Guinness	Nonstationary spatial-temporal modeling for large datasets
	Abstract: We describe a class of nonstationary models for spatial-temporal data and methods for estimating the models when the data are frequent in time but not in space, as is the case for the Atmospheric Radiation Measurement Program's temperature data from the Southern Great Plains region. As part of the modeling, we provide a simple and intuitive way to use other meteorological observations as covariates in the covariance function. We also model spatial-temporal jumps in the data caused by weather fronts moving across the region. Finally, we use the fitted model and existing observations to produce a suite of realistic conditional simulations of the temperature process at unobserved locations over a one month period as a way to incorporate uncertainty into the spatial interpolations.