Abstract/Summary

Reliable soil property maps are essential for environmental modeling, yet conventional mapping methods remain costly and time-consuming. We developed a machine learning framework that integrates the Soil-Landscape Estimation and Evaluation Program (SLEEP) with gradient boosting to predict soil properties at regional scales and multiple depths. Our approach addresses multicollinearity through a recursive feature selection algorithm. We applied this framework to a tropical region characterized by a ~700-km longitudinal gradient of contrasting topography, climate, and vegetation (~98,000 km²; NE Brazil), where scarce soil physicochemical data limit environmental modeling. We used six topographical, ten climate, and two vegetation covariates, along with data from 223 soil profiles (~1 profile per 440 km²). Training and testing of our framework demonstrated strong spatial performance (r² = 0.79–0.98 and percent bias = -1.39 to 1.14%). Topographic and climatic factors held greater weight than other variables in predicting soil layers, texture, and sum of bases. Moreover, we used our soil parameters combined with multiple pedotransfer functions (PTFs) to derive soil hydraulic properties. Our PTFs-derived estimates of hydraulic conductivity were considerably lower than high-resolution global predictions available for our study areadue to differences in clay fraction and mineralogy. Therefore, we recommend the use of region-specific PTFs for hydraulic properties based on multi-covariate soil property maps. This cost-effective framework accurately integrates diverse environmental covariates, adapts to varying soil data availability, and scales across spatial resolutions, making it highly transferable to other data-scarce regions.