Spatiotemporal Modeling of Surface Water–Groundwater Interactions via Multi-Task Transformer-Based Learning

Abstract

A spatiotemporal, multi-task learning (MTL) model for simulating surface water–groundwater (SW-GW) dynamics is developed and applied to the Heihe River Basin, Northwest China. The Transformer-based model (MT-TFT) jointly forecasts surface runoff and groundwater levels, outperforming MTL models built on gated recurrent unit (GRU) and long short-term memory (LSTM) architectures. Compared with single-task learning, adding a coupled groundwater-level task markedly improves surface runoff prediction, achieving a Nash–Sutcliffe efficiency (NSE) of 0.73 and a coefficient of determination (R2) of 0.75. Attention-based interpretability shows that the model assigns the highest weights to time steps with elevated precipitation; as lead time shortens, attention further concentrates on these periods, improving the accuracy of near-term, multi-step forecasts. These results highlight the value of inductive transfer across hydrologic targets and demonstrate that MT-TFT provides an effective, interpretable framework for SW–GW coupling.