Interpretable transformer model for national scale drought forecasting: Attention-driven insights across India

Abstract

The impacts of climate change are increasingly evident through the rise in severe droughts globally. These events result in intensified socio-economic and environmental effects. Proactive drought management requires effective forecasting and an improved understanding of the underlying hydro-climatic variables. The present study focuses on developing a national-scale drought forecasting model tailored to the diverse climatic zones of India. This model leverages the attention-based transformer framework to forecast SPEI-3 values at a lead time of 30, 60, and 90 days respectively while interpreting the complex spatiotemporal dependencies. The model predicted the SPEI-3 values with Root Means Square Error (RMSE) of 0.67 ± 0.08 and Nash-Sutcliffe Efficiency coefficient (NSE) of 0.51 ± 0.14 at a lead time of 30 days. Prediction uncertainty through quantile forecasting enhances the model's utility for effective decision-making and risk management. Model performance varies on the seasonal scale with higher accuracy in post-monsoon (Oct–Nov) and a relative decline in the pre-monsoon (March–May) season. Among large-scale climate drivers, the Indian Ocean Dipole (IOD) was found to have the highest attention representing its significant influence over Indian drought dynamics compared to other global circulation indices. While involving the static variables, the attention to spatial coordinates was found to be higher than elevation. However, in dynamic variables, precipitation, and past SPEI-3 values exhibited the most significant impact. Plots of temporal attention explain the seasonal variability present in the model's predictions. This research presents a comprehensive model, which advances our knowledge of the dynamics of drought forecasting in India.