ReDF-Net: task-oriented fusion of gridded and in-situ information for daily runoff forecasting

The growing availability of reanalysis gridded products and in-situ observations offers new opportunities for multi-source data fusion in catchment-scale runoff forecasting. However, many existing multi-modal approaches decouple gridded feature extraction from sequence prediction, rely on simple concatenation that struggles with high-dimensional spatiotemporal information, and provide limited quantitative interpretation of the relative contribution of heterogeneous inputs. Here we present ReDF-Net, a task-oriented residual–attention framework for daily (t+1) runoff forecasting that explicitly fuses ERA5-Land surface soil moisture with in-situ observations. ReDF-Net employs a modified residual network to extract compact representations from sequences of daily gridded soil moisture fields, with the historical time window encoded as input channels. Learnable softmax attention weights are embedded within the residual blocks to adaptively reweight features during end-to-end training, ensuring that gridded feature learning is directly optimized by the forecasting loss. The attention-aggregated gridded representation is then fused with site-based predictors and fed into a time-series forecasting backbone for runoff prediction. Beyond predictive accuracy, we quantify global multi-source feature contributions using the converged attention weights and interpretability analysis, and contrast model behavior during flood and non-flood periods to facilitate process-consistent interpretation. The framework is evaluated at Yichang station in the Yangtze River basin and Lanzhou station in the Yellow River basin, and benchmarked against representative recurrent forecasting models. Results demonstrate that task-oriented fusion of ERA5-Land surface soil moisture and in-situ observations improves daily runoff forecasting skill while maintaining training stability, and provides transparent attribution of how different data sources support runoff prediction. The proposed approach offers an interpretable pathway for advanced data–model fusion of hydrological variables at the catchment scale.