Characterization and forecasting of SPEI-based drought in Southern Telangana using statistical machine learning models

Drought, characterized by prolonged periods of low precipitation, poses significant challenges in Southern Telangana, one of India’s most drought-prone regions. In this study, we used the Standardized Precipitation Evapotranspiration Index (SPEI) computed at the 3-month (SPEI-3) and 6-month (SPEI-6) timescales to characterize and forecast drought conditions. Historical monthly rainfall and temperature data for the period 1988–2020 were obtained from the India Meteorological Department (IMD), Pune, to compute the SPEI indices. The long-term SPEI-3 and SPEI-6 trends revealed frequent severe and extreme droughts across Hyderabad, Mahbubnagar, Nalgonda, and Rangareddy districts, underscoring the region’s vulnerability to monsoon variability and El Niño impacts. To forecast drought, three models, viz., auto-regressive integrated moving average (ARIMA), time delay neural network (TDNN), and support vector regression (SVR) were used. SVR consistently outperformed ARIMA and TDNN, achieving over 90% error reduction in Hyderabad and Nalgonda, with notable improvements across testing datasets. In the testing phase, the SVR model achieved low RMSE values across all districts; 0.318 (Hyderabad), 0.408 (Mahbubnagar), 0.326 (Nalgonda), and 0.320 (Rangareddy) for SPEI-3, and 0.173, 0.195, 0.261, and 0.424, for SPEI-6, respectively, demonstrating its superior predictive accuracy compared with ARIMA and TDNN. TDNN provided moderate gains over ARIMA but generally lagged behind SVR. Overall, SVR demonstrated superior generalization and forecasting accuracy for both SPEI-3 and SPEI-6 indices. These findings offer valuable insights for improving drought preparedness and management in Southern Telangana.