Data-Efficient Field-Scale Mapping of Soil Salinity and Its Spatial Controls Using UAV Multispectral Imagery, Proximal Sensing, and Interpretable Machine Learning

Accurate characterization of soil salinity is essential for sustainable agricultural management in arid and semi-arid regions, but conventional field surveys are limited by intensive sampling requirements and poor spatial coverage. This study integrated UAV multispectral sensing, near-surface electromagnetic induction (EMI) measurements, PlanetScope imagery, machine learning, and geospatial statistics to map field-scale soil salinity from limited ground observations in Sehb El Masjoune, central Morocco. Ground data included 26 soil samples analyzed for electrical conductivity (EC) and 500 measurements of apparent electrical conductivity (ECa, EM38-MK2) calibrated against laboratory EC.

Spectral and terrain predictors derived from UAV and PlanetScope imagery were used with Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), Random Forest (RF), and a stacked Ensemble Learning Model (ELM). Regression Kriging (RK) was applied to account for spatially structured residuals. ELM achieved the highest predictive performance (R² = 0.87; RMSE ≈ 4.15), while RF effectively captured nonlinear spatial variability. RK improvements were model-dependent, with PLSR showing the largest gains (11.1% for PlanetScope and 13.8% for UAV predictors), whereas improvements for RF, SVR, and ELM were limited. SHAP analysis identified terrain variables as the strongest controls on predicted salinity, emphasizing the importance of topography in field-scale salinity patterns. Despite their contrasting spatial resolutions, centimetric UAV and 3 m PlanetScope imagery produced consistent representations of saline–sodic areas. Overall, integrating remote sensing with sparse proximal observations and interpretable machine learning enabled data-efficient salinity mapping, while model formulation and representation of spatial controls were more influential than imagery resolution alone.