A multi-source adaptive transfer learning model integrating spatiotemporal environmental perception for runoff prediction in data-scarce basins of the Loess Plateau

Study region The study area is located on the Loess Plateau of China. The catchments controlled by the Ganguyi, Hejin, Huaxian, Sanmenxia, and Wenjiachuan hydrological stations are selected as source basins, whereas the catchments controlled by the Tongguan and Yanchuan stations are designated as data-scarce target basins. Study focus To overcome data scarcity constraints and enhance physical plausibility, a multi-source adaptive transfer learning model integrating spatiotemporal environmental perception (termed Transformer-LSTM-MATL) is proposed. Based on the Transformer-LSTM model, this model first creates a heterogeneous prior knowledge base with a variety of source basins. A dynamic attention mechanism is incorporated to characterize the evolving environmental conditions in the target basin. By tracking multidimensional environmental features, the model dynamically adjusts the contribution of different source basins, which allows more flexible knowledge transfer than conventional single-source or static transfer approaches. Ultimately, through effective aggregation of multi-source knowledge, the runoff prediction accuracy in data-scarce basins is significantly improved. New hydrological insights for the region Evaluations in two typical data-scarce basins of the Loess Plateau demonstrate that the multi-source transfer model achieves Nash-Sutcliffe Efficiency (NSE) values exceeding 0.8. Compared with single-source transfer models and direct prediction models, NSE increases by 9.77%-168.23%, while the Root Mean Square Error (RMSE) decreases by 21.25%-56.56%. Meanwhile, the Pearson correlation coefficient (R) improves by 1.23%-25.75%. Further analysis of the attention weights indicates that their spatiotemporal variation is closely related to basin landscape characteristics and seasonal hydrological regimes. This pattern suggests that the fusion preferences of the model are generally consistent with hydrological understanding rather than being driven solely by statistical fitting. Overall, by accomplishing a comprehensive validation from simulation metrics to hydrological plausibility, a robust runoff prediction model balancing high predictive accuracy and physical interpretability for data-scarce basins is presented in this study.