Multiple Markov Chains for Categorial Drought Prediction on the U.S. Drought Monitor at Weekly Scale

Abstract

Abstract Predicting drought severity is essential for drought management and early warning systems. Although numerous physical model-based and data-driven methods have been put forward for drought prediction, their abilities are largely constrained by data requirements and modeling complexity. There remains a challenging task to efficiently predict categorial drought, especially for the U.S. Drought Monitor (USDM). Aiming at this issue, multiple Markov chains for USDM-based categorial drought prediction are successfully proposed and evaluated in this paper. In particular, this study concentrated on how the Markov order, step size, and training set length affected prediction accuracy (PA). According to experiments from 2000 to 2021, it was found that the 1-step and first-order Markov models had the best accuracy in predicting droughts up to 4 weeks ahead. The PA steadily dropped with increasing step size, and the average accuracy at monthly scale was 88%. In terms of seasonal variability, summer (July–August) had the lowest PA while winter had the highest (January–February). In comparison with the western region, the PA in the eastern United States is 25% higher. Moreover, the length of the training set had an obvious impact on the PA of the model. The PA in 1-step prediction was 87% and 78% under 20- and 5-yr training sets, respectively. The results of the study showed that Markov models predicted categorical drought with high accuracy in the short term and showed different performances on time and space scales. Proper use of Markov models would help disaster managers and policy makers to put mitigation policies and measures into practice.