Predicting Drought in the United States Through Spatio-Temporal Ordinal Regression

Abstract

The United States Drought Monitor (USDM) is a discrete measure of drought intensity comprised of six ordered levels, and has been used to classify drought on a weekly basis throughout the continental United States. Environmental covariates representing air temperature, precipitation, and soil moisture are used to create predictive models of the USDM from approximately 2004 - 2014 for the continental United States west of Nevada's eastern border. This is done through a latent Gaussian variable augmentation Bayesian network with parameter expansion. Four models are described to measure the efficiency and significance of modeling the environmental covariates, including a spatially distributed auto-correlation effect, and training the data on the full ten years of data versus only five. Four predictive periods are further compared representing different seasons for a total of sixteen models. Comparisons between these models are made using different formulations of ranked probability scores.