Abstract

Soil salinity poses a significant threat to agricultural productivity and ecosystem sustainability in arid landscapes, affecting approximately 833 million hectares globally. Traditional soil salinity assessment methods are time-consuming, labor-intensive, and spatially limited. This study presents a comprehensive AI-based approach for mapping soil salinity in arid regions using machine learning algorithms integrated with remote sensing data and ground truth measurements. The research was conducted across 15,000 km² of arid landscape in the Thar Desert region, combining multispectral satellite imagery from Landsat-8 and Sentinel-2 with field-collected electrical conductivity measurements from 2,847 sampling points. Four machine learning algorithms were evaluated: Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Gradient Boosting Machine (GBM). The Random Forest model demonstrated superior performance with an overall accuracy of 89.3%, R² of 0.87, and RMSE of 2.14 dS m⁻¹. Spectral indices including Normalized Difference Salinity Index (NDSI), Salinity Index (SI), and Brightness Index (BI) emerged as the most influential predictor variables. The developed model successfully identified five salinity classes ranging from non-saline (<2 dS m⁻¹) to extremely saline (>16 dS m⁻¹) areas. Results revealed that 34.7% of the study area exhibited moderate to severe salinity levels, with hotspots concentrated around salt lakes and low-lying areas. The AI-based mapping approach provides a cost-effective, scalable solution for monitoring soil salinity dynamics in arid regions, supporting precision agriculture and land management decisions.