Remote sensing offers obvious advantages for large-scale soil moisture (SM) mapping. However, the strong variability across different remote sensing-based SM datasets-caused by differences in retrieval algorithms and observation systems-poses substantial challenges for consistent applications. In situ observations, while reliable at the point scale, are spatially sparse and often temporally discontinuous, especially with gravimetric methods. Furthermore, most existing fusion studies are limited to small, environmentally homogeneous regions, resulting in poor model generalization when applied over large and complex areas. To overcome these limitations, we propose a novel multisource SM fusion that integrates diverse geographic and climatic variables into typical machine learning models, including random forest (RF), support vector machines, and back propagation neural networks, to derive high-accuracy, spatiotemporally continuous SM data. This framework identifies intrinsic connections between multisource SM datasets and various geographic and climatic factors, improving the quality and reducing uncertainties of these datasets. A stratified k-fold cross-validation method is used for training and validating the machine learning models. Evaluation using the training subset of in situ SM measurements, sampled from a total of 594 stations across the Huang-Huai-Hai River Basin, showed that the RF-based fusion framework outperformed other models, with a coefficient of determination (R-2) between 0.43 and 0.6. Incorporating geographic and climatic information into the fusion framework proved beneficial, as shown by the following performance hierarchy: Fusion of multisource SM datasets and geographic-climatic environmental factors (median R-2 is an element of[0.38, 0.52]) > Fusion considering only multisource SM datasets (median R-2 is an element of[0.23, 0.36]) > Fusion considering only geographic-climatic environmental factors (median R-2 is an element of[0.14, 0.19]). In addition, the independent testing subset of in situ SM measurements demonstrated that the fusion datasets outperformed both individual satellite/reanalysis SM products and the fusion results from the comprehensive averaging and TC methods. Specifically, the fusion dataset constructed from separate watersheds outperformed the dataset constructed from the whole area.