Comparison of soil moisture mapping techniques: Evaluating dataset variability and spatial transferability across regions

Abstract Soil moisture (SM) is a key parameter for irrigation monitoring, scheduling, and supporting precision agriculture. In this study, we used Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 optical data to estimate SM at high spatial resolution (20 m) in the Lower Chenab Canal Command (LCC) area of Punjab, Pakistan. To achieve this, we applied a semiempirical water cloud model (WCM) using both the VV and VH polarizations from SAR data. Additionally, two widely used machine learning (ML) models, random forest (RF) and support vector machines (SVM), were employed to estimate SM at the field scale. To assess model reliability and transferability, reference data from two field sites were split using two approaches: (1) stratified random sampling, with 50% of data from each site used for training and 50% for validation, and (2) spatial splitting, where one site was used solely for training and the other for validation. Results showed that with random splitting, ML models, particularly SVM, outperformed the WCM, with a coefficient of determination (R²) of 0.58, root mean square error (RMSE) of 6.78 Vol.%, and mean absolute error (MAE) of 5.71 Vol.%. Moreover, VV outperformed VH with an RMSE (R 2 ) of 7.37 Vol.% (0.51), compared with 7.69 Vol.% (0.46) when a stratified random split was used. In contrast, the spatial split yielded lower accuracy, as WCM with VH polarization performed best, achieving an R 2 of 0.45 and RMSE (MAE) of 7.82 (6.23) Vol.%. The high-resolution SM estimates offer practical value for optimizing irrigation management, particularly in agricultural areas challenged by limited water availability.