Machine Learning Mapping of In Situ Soil Electrical Conductivity Using UAV and Sentinel-2 Data

Differences in resolution, spectral response, and spatial support complicate the joint use of UAV and Sentinel-2 data for mapping soil electrical properties. We developed a scale-matched machine learning framework for in situ soil electrical conductivity (EC) across four irrigated sites in southern Kazakhstan. Random Forest (RF) and Elastic Net (EN) were evaluated using nested leave-one-site-out cross-validation. For UAV data, RF with median aggregation within a 5 m buffer performed best (RMSE = 0.211 dS m−1, Q2 = 0.367) and retained positive Q2 at all withheld sites. Spatial-support matching increased cross-sensor associations, whereas low Lin’s concordance coefficients (0.105–0.135) indicated limited numerical agreement. For Sentinel-2, EN showed the best independent predictive performance and was retained as the primary mapping model, while RF was retained for secondary comparison. Model-specific Areas of Applicability covered 66.08% of the mapped area for EN and 71.92% for RF. All EN predictions exceeding the observed field maximum occurred outside its AOA. Field EC was strongly associated with laboratory-determined total water-soluble salts (Spearman’s ρ = 0.911). Because ECe was not measured, the maps represent spatial variability of in situ EC rather than standard agronomic salinity classes. Sentinel-2 supports regional screening, while UAV imagery resolves finer spatial variability. These findings contribute to sustainable irrigated agriculture by supporting soil salinity monitoring and informed land management decisions aimed at reducing land degradation risks.