Physics-Guided Soil Salinity Modeling Using Multi-Source Remote Sensing and Machine Learning

Soil salinity poses a critical threat to agricultural productivity, particularly in arid and semi-arid regions where improper irrigation practices, climate variability, and groundwater dynamics accelerate land degradation. Traditional salinity assessment methods rely on field-based measurements such as electrical conductivity (EC), which, while accurate, are spatially limited and resource-intensive. With the increasing availability of Earth observation data and cloud-based geospatial platforms, there is a growing opportunity to develop scalable, data-driven approaches for salinity monitoring. This study presents a physics-guided spatio-temporal soil salinity modeling framework implemented in Google Earth Engine (GEE), integrating multi-source satellite data and machine learning concepts within a cloud-native environment. The proposed framework combines optical and Synthetic Aperture Radar (SAR) datasets to capture both surface salt signatures and soil moisture dynamics. Specifically, Sentinel-2 surface reflectance data are used to derive a Soil Salinity Index (SI) based on shortwave infrared (SWIR) bands, while Sentinel-1 SAR backscatter (VV polarization) is employed to account for moisture influence, which often masks or modifies spectral salinity signals. Additional environmental variables including Land Surface Temperature (LST), rainfall, and vegetation indices are incorporated to enhance predictive performance. A hybrid feature engineering approach is adopted by combining physically meaningful indices and temporal encoding to represent seasonal dynamics. This approach enables improved discrimination of saline soils under varying hydrological states, addressing a key limitation in conventional optical-only methods. The framework is implemented entirely within GEE, ensuring scalability, computational efficiency, and reproducibility for large-area analysis. Two ensemble machine learning models, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), are implemented and evaluated using field-based electrical conductivity (EC) measurements. XGBoost showed better predictive power compared to RF based on RMSE and R2 statistical evaluators. The methodology is demonstrated through a case study of Jhelum district in the salt range region of Pakistan, a hypersaline rainfed agricultural landscape characterized by rainfall variability and emerging salinity challenges. The proposed framework demonstrates improved predictive capability through feature interaction and temporal representation, providing spatially explicit salinity maps at high resolution. The proposed approach contributes to the growing field of geospatial big data analytics by demonstrating how cloud-based platforms like GEE can be leveraged to develop interpretable, scalable, and application-oriented environmental monitoring solutions. It provides a practical pathway for integrating physical understanding with data-driven methods, offering significant potential for agricultural management, climate resilience planning, and sustainable land use strategies.