Snow is a major surface feature of the Songhua River Basin in winter and a critical source of spring runoff. Accurately capturing its spatiotemporal distribution is crucial for hydrological modeling, water resource management, and disaster control. This study focuses on the Songhua River Basin, integrating AMSR-2 brightness temperature, ERA5-Land reanalysis, SRTM DEM, CLCD land cover data, and snow depth observations from 61 meteorological stations over ten snow seasons (2013–2023). Two ensemble learning models, XGBoost and Random Forest (RF), were developed for snow depth estimation. Four feature combination scenarios (S1–S4) were designed to investigate the impact of different input variables on model performance. The SHAP method was employed to analyze feature importance and spatiotemporal heterogeneity. Results showed that Scenario S4, integrating multi-source features, yielded optimal performance. The XGBoost model performed better on the test set (RMSE = 4.44 cm, R² = 0.6738), demonstrating superior generalization compared to the RF model, which exhibited noticeable overfitting. Brightness temperature difference (18.7H–36.5H), air temperature, and longitude were core features. Estimation accuracy showed significant spatiotemporal variations: highest during the stable snow cover period (December–February), peaking in January; good for snow depth ≤ 15 cm, but errors increased significantly for ≥25 cm. Accuracy was better in areas with gentle terrain and homogeneous land surface, with the best performance over cropland and worst over forests. This multi-source data fusion and XGBoost framework effectively enhances snow depth estimation accuracy in the Songhua River Basin.