<p>Although data-driven models have demonstrated high performance in runoff forecasting, their predictions are inherently sensitive to the quality and representation of precipitation forcing, and the internal mechanisms through which precipitation signals are translated into discharge responses are not always transparent. To address the challenges, this study integrates deep learning (DL) models with explainable artificial intelligence(XAI) techniques to (i) investigate the influence of precipitation-product discrepancies on data-driven runoff modeling, and (ii) reveal the hydrological significance of DL-based predictions across temporal and spatial scales. A long short-term memory (LSTM) model and a hybrid convolutional neural network (CNN)-LSTM model are driven by five sets of reanalysis precipitation datasets. The integrated gradients (IG) method and CNN visualization were employed to interpret the internal mechanisms of the models. The framework was applied in the upper, middle and lower reach of the Yangtze River basin. The results demonstrate that input-induced varabilities are especially pronounced during heavy rainfall and rainstorm events, emphasizing the critical impact of precipitation quality on DL-based runoff prediction. These findings indicate that precipitation discrepancies in hydrologically effective regions are more likely to amplify prediction errors, suggesting that targeted improvements in spatial precipitation representation may enhance model robustness, particularly in basins exhibiting higher sensitivity. Given the pronounced hydro-climatic heterogeneity and frequent flood hazards along the Yangtze River Basin, particularly the transition from snow-influenced upstream regions to rainfall-dominated mid- and lower reaches, these insights are especially relevant for improving region-specific flood forecasting and precipitation data prioritization strategies within this basin.</p>