Abstract: The development of a real-time satellite-gauge precipitation fusion approach is essential for advancing the “three lines of defense” system for rainfall monitoring and forecasting. However, existing methods are often constrained by the latency of near-real-time satellite information and the imbalance of precipitation samples, limiting their practical applicability in operational monitoring. To address these issues, this paper proposes a real-time satellite-gauge precipitation fusion method with two key features. First, a spatio-temporal deep learning model is applied to fill gaps in the satellite precipitation data during the latency period. Second, oversampling, undersampling, and hybrid sampling techniques are incorporated into a two-stage fusion framework consisting of classification and quantitative estimation to mitigate the impact of precipitation sample imbalance characteristics. The performance of the proposed method was assessed through case studies of four typical heavy precipitation events over the Chao Lake and Taihu Lake basins in 2020. The results demonstrate that the U-Net model outperforms ConvLSTM in filling the latency gaps, achieving the Root Mean Square Error (E_RMS) below 1.75 mm/h. The hybrid sampling method significantly surpassed undersampling, effectively enhancing the accuracy of the machine learning model in discriminating rain/no-rain and heavy/light precipitation events, thus securing an F1-score exceeding 0.88. Compared to the traditional two-stage fusion method, the proposed approach not only enhances the discrimination ability for both light (0—0.1 mm/h) and heavy (≥25 mm/h) rainfall intensity intervals, but also increases the Kling-Gupta Efficiency (E_KG) by an average of 0.36. This study advances satellite-gauge precipitation fusion from a near-real-time to a truly real-time paradigm and provides a useful supplement to the areal rainfall monitoring technology system within the “three lines of defense” framework.