Multi-Timescale SPEI Drought Forecasting Using Random Forest Regression over Maharashtra, India

Accurate drought forecasting is crucial for effective agricultural risk management in semi-arid regions, particularly in drought-prone regions of Maharashtra, India, where the majority of the population relies on farming. This study develops a one-month-ahead drought forecasting using random forest regression, an ensemble tree-based machine-learning algorithm, using the Standardized Precipitation Evapotranspiration Index (SPEI) at multiple temporal scales. Random Forest regression models were trained to forecast SPEI-3, SPEI-6, and SPEI-12, incorporating rainfall, temperature, and derived hydro-climatic predictors. Model performance exhibits clear timescale-dependent predictability, with skill increasing for longer accumulation periods: SPEI-3 (R² = 0.55, RMSE = 0.81), SPEI-6 (R² = 0.65, RMSE = 0.69), and SPEI-12 (R² = 0.87, RMSE = 0.38). Corresponding generalization ratios of 62.4%, 71.8%, and 90.5% indicate improved robustness and reduced overfitting at short (SPEI-3) to long (SPEI-12) timescales. Feature importance analysis consistently highlights the current SPEI state, contributing approximately 35–40% of the total importance, followed by the precipitation minus potential evapotranspiration (PPET) balance and other hydro-climatic variables, reflecting the dominant role of drought persistence and climatic memory in one-month-ahead forecasting. The models successfully capture spatial drought patterns, though reduced accuracy is observed for extreme drought magnitudes at shorter timescales, likely due to inherent climate non-stationarity and rapidly evolving predictor relationships. Overall, this study demonstrates the effectiveness of machine-learning-driven, one-month-ahead drought forecasting across multiple SPEI time scales, enabling near-real-time monitoring and early warning depending on the selected accumulation period. The proposed framework provides a scalable foundation for operational drought early-warning systems in Maharashtra and other drought-prone hydro-climatic regions worldwide.

Keywords: SPEI, Drought forecasting, Random Forest