摘要 : Accurate representation of the spatiotemporal distribution of precipitation is critical to climate and hydrology fields. Data-driven methods based on machine learning have exhibited the potential to fuse satellite-derived precipit... 展开 Accurate representation of the spatiotemporal distribution of precipitation is critical to climate and hydrology fields. Data-driven methods based on machine learning have exhibited the potential to fuse satellite-derived precipitation data (SPD) and ground-based precipitation observations (GPOs). However, machine learning is not specifically designed to address geospatial problems and cannot effectively leverage the prior knowledge of precipitation, which limits the performance of fusion models to a certain extent. The present work addresses this problem by proposing a novel data- and knowledge-driven fusion method. Specifically, two schemes are designed to organically integrate the temporal heterogeneity and local spatial autocorrelation knowledge of precipitation into a deep learning model. Then, an integrated loss function is constructed to collaboratively constrain the bias of the fusion results in both temporal and spatial dimensions during the model training process. The proposed method is validated by fusing ten years of the Integrated Multi-satellitE Retrievals (IMERG) final run (IMERGF) daily precipitation data and GPO. The results show that the fused precipitation data generated by the proposed method can reliably capture the temporal and spatial variations of precipitation, increasing the median Kling–Gupta efficiency (KGE) values of the IMERGF daily data in both temporal and spatial dimensions from 0.69 and 0.13 to 0.81 and 0.79, respectively. Moreover, this method demonstrates a substantial performance superiority over the fusion methods based on random forest (RF) and extreme gradient boosting (XGBoost). These promising results imply that the proposed method has great potential in generating accurate precipitation data by the deep fusion of SPD and GPO. 收起