Soil moisture content (SMC) in farmland is one of the most important indicators to optimize the water sources for high irrigation efficiency in sustainable agriculture. However, only a single source from remote sensing data cannot fully meet the requirements of the SMC inversion. Spatiotemporal coverage and information dimensions are often required to accurately monitor the SMC in saline-alkali farmland. In this study, a multi-source image fusion framework was constructed to realize complementary multi-sensor information and feature enhancement using Landsat 9 OLI and Sentinel-2 MSI data. Saline-alkali farmland was taken from the Hetao Plain. A feature system was established to couple the spectral indices with environmental variables. SMC inversion models were developed using extreme gradient boosting (XGBoost), random forest (RF), and convolutional neural networks (CNN). Shapley additive explanations (SHAP) were then employed to determine the variable contributions. The optimal model was used to generate SMC spatial maps for the area statistics. The results indicate that: 1) Deep fusion of multi-source imagery effectively balanced spatiotemporal continuity with information richness, in terms of spatial resolution and spectral consistency, compared with the single-sensor data. Specifically, while Landsat 9 imagery represented the surface distribution at a regional scale, due to its radiometric stability and tonal continuity; Conversely, Sentinel-2 imagery with high spatial resolution was suitable for the field boundaries and surface textures, but there were spectral fluctuations and brightness inconsistencies in localized areas. Both data types were effectively integrated after deep fusion. The fused imagery significantly enhanced the spatial representation with the spectral continuity. Thereby, the performance of SMC inversion was improved during the bare soil period. 2) A comparison was made on the accuracy of SMC inversion models. Among them, the CNN model shared the best overall performance, with a coefficient of determination R²=0.798, root mean square error (RMSE) of 2.79 g/kg, and relative percentage difference (RPD) of 2.10. The RF model ranked second, with R², RMSE, and RPD values of 0.775, 3.08%, and 2.03, respectively. While XGBoost performed the worst, with R², RMSE, and RPD of 0.641, 3.21%, and 1.95, respectively. The SMC-CNN inversion shows that the SMC in the study area was primarily concentrated in the 10%–30% range. 3) SHAP analysis indicated that the precipitation was the most critical driving factor, where the spectral indices contributed the most (46.87%), followed by climatic factors (27.22%). While single-band, soil texture, and topographic factors made relatively limited contributions, at 10.99%, 8.29%, and 6.70%, respectively. The spectral indices were determined to effectively capture key information with the SMC. Furthermore, climate change played a significant role in SMC variations, whereas topographies had little impact on SMC distribution in plain areas. Multi-source image fusion was combined with coupled modeling of environmental variables. The finding can provide a technical approach and theoretical foundation for the large-scale, efficient, and interpretable SMC monitoring during the bare soil period in saline-alkali farmland.

Hyperspectral remote sensing has been one of the most important technologies to monitor soil salinization. However, soil moisture can also interfere with the acquisition of field spectral images and then the prediction of soil salinity. The aim of this study was to effectively eliminate the effect of soil moisture, thereby improving the inversion accuracy of soil salinity. Field spectral reflectance (Ref) was collected from the Yinchuan Plain of China. The soil salt content was then measured to analyze the spectral characteristics of soils with different moisture content. The first derivative of reflectance (FDR), orthogonal signal correction (OSC), and their combination (FDR-OSC) were also employed to transform the spectral data. A systematic evaluation was performed on the correlations of original and transformed spectral data with the soil salt and moisture content. A soil salt content inversion model with the support vector machine (SVM) was finally established after removing the moisture effect. The results showed that: 1) The overall change in soil salt content was striking (0.30-14.98 g/kg), indicating a strong variability across the study area. Soil pH ranged from 7.82 to 9.58, with a small standard deviation and weak variability. Soil moisture content was ranged between 2.49% and 29.29%, representing a moderate variability. 2) An inverse relationship emerged between soil moisture content and the original spectral reflectance, with the three absorption bands near 1430, 1950, and 2200 nm, respectively. The principal absorption wavelength was near 1950 nm, indicating a long-wave drift pattern. The reflectance values were reduced for the better drift of the absorption valley using the OSC, compared with Ref. The bandwidth and depth of the three moisture absorption bands all converged, despite the lower differential reflectance values from the FDR transformation curve. 3) The correlation between soil spectral data and moisture content was followed the descending order: Ref, OSC, FDR, FDR-OSC, indicating that FDR-OSC performed better than OSC and FDR to remove the moisture effect. The correlation between soil spectral data and soil salt content was followed the descending order: FDR-OSC, FDR, OSC, Ref, indicating that the highest sensitivity of spectral data was transformed by FDR-OSC to soil salinity. Only five bands in the 400-2400 nm wavelength range of Ref were selected as the sensitive bands (R²≥0.50, P<0.05), much fewer than 11 and 18 sensitive bands of OSC and FDR, respectively. The highest number of sensitive bands was obtained by FDR-OSC, reaching 26. The increased sensitive bands of OSC, FDR, and FDR-OSC mainly comprised near-infrared long waves. 4) Compared with the Ref-based model, the soil salt content inversion model with FDR-OSC performed better fitting and inversion, as the modeling coefficient of determination(R_c²), verification of coefficient of determination (R_p²), and relative prediction deviation (RPD) were 0.952, 0.960 and 5.04, respectively. 5) Spatial interpolation by inverse distance weighting revealed that the slightly and moderately salinized soils covered 76.0% of the Yinchuan Plain, whereas the strongly salinity soil and salinity soils only accounted for 2.5%. The areas with the high soil salinity were concentrated mainly in the central and northwestern parts of the study area, with the low soil salinity in the northeastern and southern parts. In conclusion, the FDR-OSC can provide an effective way to eliminate the effect of soil moisture on spectral reflectance, thereby improving the inversion accuracy of soil salinity. The findings can also provide scientific evidence for the high-precision monitoring of soil salinization using hyperspectral remote sensing.