Abstract: The multi-source precipitation fusion technique serves as an important method for accurately estimating the spatiotemporal distribution of precipitation. However, conventional techniques fail to adequately incorporate the spatial local correlation and temporal dependence of precipitation, thus limiting their ability to reproduce its spatial distribution patterns. To overcome this limitation, this study first selected three satellite precipitation products (IMERG, CMORPH, and GSMaP) and ground-based observation data. Subsequently, it constructed a deep learning-based spatiotemporal precipitation fusion model (3DCNN-ConvLSTM) that integrates a three-dimensional convolutional neural network (3DCNN) and a convolutional long short-term memory neural network (ConvLSTM). The integrated model achieves accurate precipitation estimation by deeply extracting its spatiotemporal variations. The results indicate that at the daily scale, the proposed 3DCNN-ConvLSTM significantly outperformed the original satellite precipitation products in precipitation fusion. The superiority was evinced by the following metrics: the correlation coefficient and Kling-Gupta efficiency coefficient increased to 0.679 and 0.64, while the root mean square error (RMSE) and the mean absolute error (MAE) decreased by 11.7% to 24.4% and 9.3% to 20.7%, respectively. Besides, the model exhibited enhanced accuracy for detecting precipitation events of different intensities. At the monthly scale, the model improved estimation accuracy of monthly precipitation to different degrees, with the more significant improvements occurring in high-precipitation months. At the spatial scale, the integrated model corrected the spatial overestimation of the original satellite precipitation products, yielding the highest correlations and the smallest errors over varying terrains. Compared with other fusion models, the 3DCNN-ConvLSTM presented superior performance in improving precipitation data accuracy. In conclusion, the multi-source precipitation fusion model considering the spatiotemporal correlations can effectively improve the quality of precipitation data in the Fujian-Zhejiang-Jiangxi region, offering practical value in the field of multi-source precipitation fusion.