DCSwinLSTM for spatiotemporal meteorological drought forecasting

Meteorological drought prediction is critical for early warning and climate-risk management; yet, modeling regional drought evolution across multi-timescale remains challenging. Here, we propose DCSwinLSTM, a feature-fusion framework that combines deformable convolution for boundary-sensitive spatial encoding, a Swin transformer for hierarchical multi-scale feature learning, and an LSTM recurrent update for temporal dependency modeling. Using a global gridded dataset (1959-2022) constructed from the standardized precipitation index (SPI) and standardized precipitation evapotranspiration index (SPEI) at 3- and 6-month scales, DCSwinLSTM consistently outperforms mainstream spatiotemporal prediction baselines. On the held-out SPI-3 test set, it achieves mean square error (MSE) 0.4269, mean absolute error (MAE) 0.5050, root mean square error (RMSE) 0.6534, R² 0.5875, and PSNR 20.6669; on SPI-6, it attains MSE 0.2800, MAE 0.4003, RMSE 0.5292, R² 0.7265, and PSNR 22.0230. These results support reliable multi-timescale drought forecasting for risk management and water-resource planning, even in data-scarce regions.