Abstract: To improve the spatial resolution of satellite precipitation in the Upper Yellow River Basin, a CNN-LSTM model was constructed by integrating a Convolutional Neural Network (CNN) to extract spatial features from environmental factors and a Long Short-Term Memory (LSTM) network to capture the temporal dependencies of precipitation processes. Transfer learning (TL) was then introduced to the CNN-LSTM base model to establish a CNN-LSTM-TL satellite precipitation downscaling model. Additionally, the SHAP (SHapley Additive exPlanations) method was applied to interpret the contributions of environmental factors. The results showed that the proposed satellite precipitation downscaling model achieved precipitation estimation at a spatial resolution of 500 m, with a coefficient of determination (R²) of 0.82, representing improvements of 10.8%, 22.4%, and 43.8% compared with the CNN-LSTM, CNN-TL, and LSTM-TL models, respectively, thereby significantly enhancing precipitation estimation accuracy in the Upper Yellow River Basin. Land surface temperature (LST) contributed most to the model predictions, followed by evapotranspiration (ET). The negative contribution of ET was more pronounced in low-altitude areas, whereas the positive contribution of LST was more evident in high-altitude regions. The proposed approach provides technical support for water resources management and drought monitoring in the Upper Yellow River Basin.