Spatio-temporal forecasting for the US Drought Monitor

Abstract

Abstract The US Drought Monitor is the leading drought monitoring tool in the United States. Updated weekly and freely distributed, it records the drought conditions as geo-referenced polygons showing one of six ordered levels. These levels are determined by a mixture of quantitative environmental measurements and local expert opinion across the entire United States. At present, forecasts of the Drought Monitor only convey the expected direction of drought development (i.e. worsen, persist, subside) and do not communicate any uncertainty. This limits the utility of forecasts. In this paper, we describe a Bayesian spatio-temporal ordinal hierarchical model for use in modelling and projecting drought conditions. The model is flexible, scalable, and interpretable. By viewing drought data as areal rather than point-referenced, we reduce the cost of sampling from the posterior by avoiding dense matrix inversion. Draws from the posterior predictive distribution produce future forecasts of actual drought levels—rather than only the direction of drought development—and all sources of uncertainty are propagated into the posterior. Spatial random effects and an autoregressive model structure capture spatial and temporal dependence, and help ensure smoothness in forecasts over space and time. The result is a framework for modelling and forecasting drought levels and capturing forecast uncertainty.