Abstract

Digital Soil Mapping (DSM) has revolutionized traditional soil survey methods by integrating environmental covariates with advanced statistical and machine learning techniques. This study investigates ensemble modeling approaches for improving DSM accuracy and uncertainty quantification across heterogeneous landscapes. We implemented and compared five ensemble strategies: bagging, boosting, stacking, voting, and Bayesian model averaging, using a comprehensive dataset of 1,250 soil samples across 750 km² in Central Europe. Environmental covariates included terrain attributes, climate variables, remote sensing indices, and legacy soil data. The stacking ensemble approach, combining Random Forest, Gradient Boosting, Support Vector Machines, and Cubist models, achieved the highest prediction accuracy for soil organic carbon (R² = 0.92, RMSE = 0.38%), clay content (R² = 0.88, RMSE = 3.2%), and pH (R² = 0.86, RMSE = 0.41). Ensemble methods reduced prediction uncertainty by 28-35% compared to individual models while providing robust uncertainty estimates through prediction intervals. Spatial cross-validation revealed consistent performance across different landscape units, demonstrating model transferability. This research establishes a framework for operational DSM implementation, offering improved accuracy and reliability for soil resource assessment and management.