SPEI-Based Drought Regionalization and Cluster-Specific Machine Learning Forecasting in the U.S. Black Belt

This study evaluates the historical spatial variability of drought in the Black Belt region of the southeastern U.S. by classifying the study area into drought-variation clusters and assesses the performance of three machine learning models for short-term drought forecasting. The Black Belt is a distinct agricultural subregion of the southeastern U.S., characterized by fertile Blackland soils, strong dependence on precipitation-based farming, limited irrigation infrastructure, uneven climatic monitoring coverage, socioeconomic challenges, and a history of repeated severe drought events that have caused substantial agricultural and water-supply impacts…Monthly gridded Standardized Precipitation Evapotranspiration Index (SPEI) [~ 5 km (3.1 miles) spatial resolution] was computed from precipitation and temperature data from the National Oceanic and Atmospheric Administration (NOAA) National Centers for Environmental Information (NCEI) nClimGrid-Daily dataset, based on Global Historical Climatology Network-Daily (GHCN-Daily) observations, and used as the primary drought indicator for the 1990–2025 study period. SPEI includes precipitation and temperature-driven evapotranspiration, which strongly influences drought onset in warm subtropical regions such as the Black Belt. The 6-month accumulation period ( SPEI₆ SPEI 6 ) is suitable for capturing agricultural drought by representing soil moisture conditions related to crop stress. Since the Black Belt has sparse and unevenly distributed weather stations, high-resolution gridded NOAA data provides spatially continuous, quality-controlled coverage (NOAA NCEI, n.d.). The Silhouette method was used to select influential SPEI₆ SPEI 6 statistical parameters and validate three clustering algorithms (i.e., K-means, Hierarchical, and Expectation–Maximization) to create drought clusters from SPEI₆ SPEI 6 data. K-means performed best, identifying three distinct drought-characteristic clusters. Three machine-learning models, Support Vector Regression (SVR), the M5P model tree (a decision-tree algorithm that fits linear regressions at its terminal nodes), and a Hybrid ensemble model combining SVR and M5P outputs, were trained for each cluster for 1-, 2-, and 3-month forecasting horizons using lagged SPEI₆ SPEI 6 time series. Model performance declined for longer horizons and for the extreme-like drought characteristics, while the Hybrid model achieved the best results. These findings highlight the effectiveness of combining drought clustering with region-specific machine learning models for improved short-term drought forecasting in heterogeneous environments such as the U.S. Black Belt.