Abstract
Soil salinity is a key driver of land degradation in arid regions and remains a severe constraint to sustainable agriculture. Field-based observations are limited in space and time, making it challenging to track seasonal and internal changes. This study presents an innovative hybrid probabilistic framework for assessing soil salinity using multi-source satellite and climate data. A quasi-temporal dataset for 2020 to 2025 was created from Sentinel-2 surface reflectance, ERA5-Land hydroclimatic variables, terrain features, and distance to water. A physically interpretable feature set was developed using spectral indices, soil moisture gradients, aridity indicators, water balance metrics, and combined spectral and climatic interactions. Soil conditions were classified into three regimes (NORMA, SALINE, SAND) using a unified decision scheme based on NDVI, NDMI, and BSI. The hybrid model estimates the probability of transitions between soil regimes across years and regions. Its performance was tested using stratified and group cross-validation, temporal hold-out (2020 to 2023 for training and 2024 to 2025 for testing), and strict external validation based on independent AOI sites. The model achieved stable results with an Accuracy of about 0.97, Macro-F1 of about 0.97, and ROC AUC above 0.97. The proposed approach provides an AI-driven, interpretable, and scalable solution for monitoring soil salinity and supporting land and water management.