Spatial modeling of soil salinity in hyperarid agroecosystems of northern Peru using a hybrid machine learning ensemble

Soil salinization poses a critical threat to agricultural sustainability in arid agroecosystems, limiting food security and economic stability. On the northern coast of Peru, the synergy between inefficient water management and extreme climate variability has accelerated soil degradation, demanding precise spatial monitoring tools. This study proposes a Hybrid Smart Ensemble model to map soil electrical conductivity (EC), integrating optical data (Sentinel-2), dual-polarization SAR texture (Sentinel-1), and thermal imagery (Landsat-9 LST). Three modeling scenarios were evaluated: (1) Random Forest (RF), (2) Deep Neural Network (DNN), and (3) a Hybrid Ensemble. To overcome the typical scarcity of field data in developing regions, a Spatial Data Augmentation strategy based on Tobler's First Law was implemented. The results demonstrated that the hybrid architecture (Scenario 3) achieved superior performance (R² = 0.86; RMSE = 1.70 dS/m), exceeding the baselines of individual classifiers. While the Random Forest provided stability in moderate ranges, the inclusion of Deep Learning proved essential for capturing the nonlinear spectral response of extreme salt crusts (>12 dS/m). Variable importance analysis identified Ground Surface Temperature (LST) and Elevation as dominant predictors, validating the hypothesis that thermal inertia and topographic depressions are the main physical drivers of salinity in hyperarid environments. The resulting map, with a 10 m resolution, reveals "hot spots" in the Chulucanas and Lower Piura sectors, providing valuable information for implementing targeted detection and analysis zones to optimize drainage rehabilitation and soil recovery strategies.While themodel effectively delineates non-saline baselines and severe salt crusts, boundary zones (moderate salinity)remain challenging due to historical sampling limitations. Therefore, the applicability of the model is currentlylimited to hyperarid agroecosystems with edaphic and topographic characteristics similar to those of the LowerPiura Valley; results should be validated with additional field campaigns before implementation in other contexts.