Mapping Soil Salinity Using Machine Learning and Remote Sensing Data in Semi-Arid Croplands

Abdelwahed Chaaou^1*Hamza Ait-Ichou²Said El Hachemy²Mohamed Chikhaoui³Mustapha Naimi⁴Mohammed Hssaisoune⁵Mohammed El Hafyani⁵Yassine Ait Brahim¹Lhoussaine Bouchaou⁵

• ¹Mohammed VI Polytechnic University (UM6P), Ben Guerir, 43150, Morocco
• ²Faculty of Sciences, Ibn Zohr University, Agadir, 80035, Morocco
• ³Hassan II Institute of Agronomy and Veterinary Medicine,, Rabat, Morocco
• ⁴Hassan II Institute of Agronomy and Veterinary Medicine, Rabat, Morocco
• ⁵Ibn Zohr University, Agadir, 80035, Morocco

Soil salinity significantly constrains agricultural productivity and land sustainability, particularly in irrigated areas. While, remote sensing offers large-scale monitoring capacity, but its accuracy depends on how effectively spectral information is integrated with advanced modeling approaches. This study evaluates the performance of a combined approach based on machine learning (ML) algorithms and satellite-derived predictors for soil salinity mapping in the Béni Amir Sub-perimeter of Tadla plain, Morocco. A total of 43 topsoil samples (0-10 cm) were collected and analyzed for electrical conductivity (ECe) and resampled to 144 samples for model training and testing. Predictor Variables were derived from Landsat-8 OLI data, including salinity indices (OLI-SI, SI, SI1), intensity indices (Int1, Int2), brightness index (BI), land degradation index (LDI), and reflectance values of selected spectral bands (B2-B7) were standardized and transformed with PCA to address multicollinearity. Four ML algorithms, Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Regressor (SVR), and Multi-Layer Perceptron (MLP) were tested. The results show that the Ece ranges from 0.84 to 10.28 dS/m with a standard deviation of 2.29 dS/m, indicating substantial salinity variability across the Béni Amir sub-perimeter. Individual predictors exhibited moderate correlation with Ece (R=0.34-0.72). Among the applied models, KNN achieved the highest accuracy (mean coefficient of determination (R²) = 0.75 [0.73-0.77]; Root Mean Square Error (RMSE) = 0.61 dS/m). The resulting maps revealed a consistent southwestward increase in salinity, following the regional hydraulic flow. KNN classified 49% of the area as moderately saline, 22% as slightly saline, and 20% as non-saline, while the strongly and extremely saline classes covered 8.4% and 0.6%, respectively. RF, SVR, and MLP showed comparable trends, with moderately saline areas ranging between 30-41% and strongly to extremely saline soils below 10%. These findings demonstrated that combining satellite-derived data with ML enables a reliable assessment of soil salinity, supporting management of irrigated agroecosystems.