Transformer-based Deep Learning for Spatiotemporal Forecasting of Natural Resource Availability

Abstract

Reliable forecasting of natural resource availability including reservoir storage, groundwater level, and streamflow is central to climate-resilient water management, yet conventional statistical and recurrent deep learning models struggle to jointly capture long-range temporal dependencies and heterogeneous spatial coupling across a basin. This study proposes ST-ResFormer, a spatiotemporal transformer architecture that fuses a graph-attention spatial encoder with a patch-based sparse-attention temporal encoder through a cross-attention fusion block, enabling joint multi-horizon forecasting of three coupled resource variables from heterogeneous data sources (gauge records, satellite soil moisture, GRACE terrestrial water storage anomalies, and reanalysis climate forcing). The framework was evaluated on a 22-year (2001–2023) multi-source dataset covering 48 grid cells of the Sutlej Beas River basin in the northwestern Himalaya, benchmarked against eight baselines spanning classical, recurrent, convolutional, graph-based, and transformer-based forecasters (Persistence, SARIMA, LSTM, ConvLSTM, GCN-LSTM, Informer [2], Autoformer [3], and PatchTST [7]). At a one-month forecast horizon, ST-ResFormer achieved a root-mean-square error (RMSE) of 4.63 and mean absolute error (MAE) of 3.41 in normalized units, corresponding to a 27.3% RMSE reduction relative to the strongest transformer baseline (PatchTST) and a 68.8% reduction relative to Persistence, while sustaining a Nash–Sutcliffe efficiency (NSE) above 0.91 out to a six-month horizon. These results demonstrate that jointly modeling spatial and temporal attention over multi-source resource data yields state-of-the-art, interpretable, and operationally deployable forecasts, offering a transferable template for climate-adaptive natural resource management in data-heterogeneous mountain basins.