Prediction of the standardized compound drought and heat index in regional scale based on multiple deep learning approaches.

Compound drought and heat events (CDHEs) have received increasing attention due to their significant impacts on water resource management, agricultural productivity, and ecosystem stability. Despite notable progress in identifying CDHEs and constructing related indices, most forecasting approaches still rely on conventional statistical frameworks. These are often inadequate for capturing the complex nonlinear interactions and spatiotemporal dependencies among meteorological variables. To address this limitation, we employed four deep learning models (i.e., CNN-LSTM, ConvLSTM, SA-ConvLSTM, and SwinLSTM) to predict the spatiotemporal evolution of the standardized compound drought and heat index (SCDHI). Among them, SA-ConvLSTM and SwinLSTM integrate attention mechanisms, enhancing the models' ability to capture spatial variability and identify relevant features. Using meteorological variables derived from ERA5 reanalysis data, we applied these models to forecast the 10-day (climatological dekad) spatiotemporal evolution of the SCDHI across the Yangtze River Delta (YRD) region in China from 2004 to 2023. Experimental results demonstrate strong predictive performance across all models, with SwinLSTM achieving the highest coefficient of determination (R2 = 0.892). Compared to the baseline ConvLSTM, SwinLSTM improved the R2 by 3.24% and reduced the RMSE and MAE to 0.422 and 0.276, respectively, showing superior accuracy in capturing complex spatiotemporal patterns of compound events. This study provides an effective and generalizable framework for the short-term spatiotemporal prediction of CDHEs, supporting early warning and regional climate resilience efforts.