Data fusion of proximal sensing and machine learning for soil salinity mapping in irrigated systems

Soil salinity is a major form of land degradation in irrigated agroecosystems, directly affecting crop productivity, soil health, and long-term sustainability. While proximal sensing techniques such as electromagnetic induction (EMI) are widely used for field-scale salinity mapping, accurately predicting low to moderate salinity levels remains challenging due to strong spatial heterogeneity, depth-dependent processes and limited transferability across fields. This study proposes a data fusion framework integrating proximal sensing (EMI and gamma-ray spectrometry), soil profile-based salinity metrics, and machine learning models to improve soil salinity estimations across four irrigated regions in Arizona and California, United States: Imperial Valley, Yuma, Salinas Valley and Colorado River Indian Tribes (CRIT). A total of 1,015 samples were collected. Electrical conductivity of saturated paste extract (ECe) from samples at 10, 40 and 120 cm, was aggregated into a profile weighted indicator ECew (0-120 cm) to better match the effective sensing depth of EMI measurements. Multiple modelling scenarios were evaluated using apparent electrical conductivity obtained from EMI at nominal depth ranges of 1.6 m and 2.3 m, gamma-ray data, soil taxonomy, and depth information. Six regression algorithms (Ridge, Random Forest, Gradient Boosting, Support Vector Regression, XBoost and CatBoost) were tested using random K-Fold and Leave-One-Field-Out (LOFO) cross-validation to assess spatial transferability.

Results show that profile integration and multi-sensor fusion improve predictive robustness. Across regions, jack-knifed mean square prediction error (MSPE) values ranged from 0.10 to 2.16 relative error, with the highest accuracy in Yuma (MSPE=0.10-0.13) and Imperial Valley (MSPE=0.12-0.27) (good to excellent accuracy), intermediate performance in CRIT (MSPE=0.15-0.36) and clear limitations in Salinas (MSPE=0.16-2.16). The use of ECew consistently reduced prediction error and improved spatial transferability under LOFO validation. Tree-based and boosting models outperformed linear and kernel-based approaches for depth-specific ECe, while Ridge regression proved most robust for ECew. Spiking analysis further demonstrated that incorporating small fraction of local ground-truth samples markedly improved LOFO performance, highlighting a practical balance between sampling effort and predicting accuracy. Overall, the fusion of indirect sensor observations and soil profile-based salinity metrics, enable scalable and transferable mapping of low to moderate soil salinity for operational applications.