Improved flash drought forecasting and attribution : a spatial-temporal causality-aware deep learning approach

Abstract

Flash droughts pose significant challenges to water resource management and agricultural sustainability, making it imperative to improve their predictability to mitigate potential risks. This study presents a novel deep learning framework that integrates a spatial–temporal causality-aware (STC) module into a CNN-LSTM hybrid architecture to enhance flash drought prediction in China’s Greater Bay Area (GBA). Ablation experiments demonstrate that the causality module enhances model generalization (GA = 0.90) and performance (NSE = 0.83), and substantially increases the accuracy of flash drought onset prediction (F1 score = 0.33) compared to baseline models. Explainable Artificial Intelligence (AI) analyses further reveal that incorporating causality strengthens the predictive contributions of key flash drought drivers, including soil moisture memory, downward longwave radiation, and precipitation. Especially, it reveals new insights into drought drivers: downward longwave radiation emerges as a critical yet previously underrecognized predictor of soil moisture variability in humid subtropical climates. Additionally, this study distinguishes the mechanisms underlying slow and flash droughts, highlighting the dominant role of initial soil moisture and persistent shortwave radiation in slow droughts, versus rapid energy imbalances and longwave radiation in flash droughts. Further findings suggest that anthropogenic activities in China’s GBA intensify the complexity of drought mechanisms, increasing both prediction difficulty and regional vulnerability to hydrological extremes. The proposed framework and insights provide a foundation for developing more effective flash drought risk management and adaptation strategies in humid subtropical regions.