Evaluating Markov chains and Bayesian networks as probabilistic meteorological drought forecasting tools in the seasonally dry tropics of Costa Rica

Abstract

Meteorological drought is a climatic phenomenon that affects all global climates with social, political, and economic impacts. Consequently, it is essential to develop drought forecasting tools to minimize the impacts on communities. Here, probabilistic models based on Markov chains (first and second order) and Bayesian networks (first and second order) were explored to generate forecasts of meteorological drought events. A Ranked Probability Score (RPS) metric selected the best-performing model. Long-term precipitation data from Liberia Airport in Guanacaste, Costa Rica, from 1937 to 2020 were used to estimate the 1-month Standardized Precipitation Index (SPI-1) characterizing four meteorological drought states (no drought, moderate drought, severe drought, and extreme drought). The validation results showed that both models could reflect the climatic seasonality of the dry and rainy seasons without mistaking 4–5 months of the rain-free dry season for a drought. Bayesian networks outperformed Markov chains in terms of the RPS at both reproducing probabilities of drought states in the rainy season and when compared to the months in which a drought state was observed. Considering the forecasting capability of the latter method, we conclude that these models can help predict meteorological drought with a 1-month lead time in an operational early warning system.