Robust Meteorological Drought Prediction Using Antecedent SST Fluctuations and Machine Learning

While reliable drought prediction is fundamental for drought mitigation and water resources management, it is still a challenge to develop robust drought prediction models due to complex local hydro-climatic conditions and various predictors. Sea surface temperature (SST) is considered as the fundamental predictor to develop drought prediction models. However, traditional models usually extract SST signals from one or several specific sea zones within a given time span, which limits full use of SST signals for drought prediction. Here, we introduce a new meteorological drought prediction approach by using the antecedent SST fluctuation pattern (ASFP) and machine learning techniques (e.g., support vector regression (SVR), random forest (RF), and extreme learning machine (ELM)). Three models (i.e., ASFP-SVR, ASFP-ELM, and ASFP-RF) are developed for ensemble, probability, and deterministic drought predictions. The Colorado, Danube, Orange, and Pearl River basins with frequent droughts over different continents are selected, as the cases, where standardized precipitation evapotranspiration index (SPEI) are predicted at the 1° × 1° resolution with 1- and 3-month lead times. Results show that the ASFP-ELM model can effectively predict space-time evolutions of drought events with satisfactory skills, outperforming the ASFP-SVR and ASFP-RF models. Our study has potential to provide a reliable tool for drought prediction, which further supports the development of drought early warning systems.