Abstract: Accurate monitoring of soil moisture content (SMC) in farmland is critical for improving irrigation efficiency, optimizing water management, and supporting sustainable agricultural development. To address the issues of insufficient spatiotemporal coverage and limited information dimensions associated with using a single remote sensing data source for SMC inversion in saline-alkali farmland, this study focuses on saline-alkali farmland in the Hetao Plain. It constructs a deep learning-based multi-source image fusion framework that integrates Landsat 9 OLI and Sentinel-2 MSI data to achieve complementary multi-sensor information and feature enhancement. Building on this, a comprehensive feature system was established by coupling spectral indices with environmental variables. SMC inversion models were developed using Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Convolutional Neural Networks (CNN) to establish SMC inversion models. Shapley Additive Explanations (SHAP) were then employed to analyze variable contributions, and the optimal model was used to generate SMC spatial maps and calculate area statistics. The results indicate that: 1) Deep fusion of multi-source imagery effectively overcomes the limitations of single-sensor data in terms of spatial resolution and spectral consistency, balancing spatiotemporal continuity with information richness. Specifically, while Landsat 9 imagery possesses good radiometric stability and tonal continuity and can reflect regional-scale surface distribution patterns, it has certain shortcomings in terms of detail representation; conversely, Sentinel-2 imagery has high spatial resolution and can clearly depict field boundaries and surface texture features, but may exhibit spectral fluctuations and brightness inconsistencies in localized areas. Through a deep fusion method, the strengths of both data types were effectively integrated, enabling the fused imagery to significantly enhance spatial detail representation while maintaining good spectral continuity, thereby improving the reliability of SMC inversion during the bare soil period. 2) A comparison of the accuracy of SMC inversion models revealed that the CNN model demonstrated the best overall performance, with a coefficient of determination R²=0.798, root mean square error (RMSE) of 2.79, and relative percentage difference (RPD) of 2.10, and the model showed stable performance. The RF model ranked second, with R², RMSE, and RPD values of 0.775, 3.08, and 2.03, respectively, indicating that the model possesses inversion capability. XGBoost performed the worst, with R², RMSE, and RPD of 0.641, 3.21, and 1.95, respectively, demonstrating only moderate inversion capability. The inversion results based on SMC-CNN show that SMC in the study area is primarily concentrated in the 10%–30% range, which closely corresponds to the observed analysis. 3) SHAP analysis indicates that precipitation is the most critical driving factor, with spectral indices contributing the most (46.87%), followed by climatic factors (27.22%), while single-band, soil texture, and topographic factors made relatively limited contributions, at 10.92%, 8.29%, and 6.70%, respectively. This indicates that the spectral indices used in this study effectively capture key information regarding SMC. Furthermore, climate change plays a significant role in SMC variations, whereas topographic variations have little impact on SMC distribution in plain areas. The results indicate that combining deep learning-based multi-source image fusion with coupled modeling of environmental variables provides a generalizable technical approach and theoretical foundation for large-scale, efficient, and interpretable monitoring of SMC during the bare soil period in saline-alkali farmland.