Application of Machine Learning Random Forest Algorithm in Digital Soil Mapping of Key Soil Properties for Sustainable Tea Productivity

Abstract Digital soil mapping (DSM) plays a crucial role in optimizing agricultural resource use including sustainable tea productivity. This study predicted key soil properties using a Random Forest (RF) machine-learning algorithm within a DSM framework, integrating soil, climate, organisms, relief, parent material, age, and spatial position (SCORPAN)-based environmental covariates derived from digital elevation model (DEM)-based topographic variables, Sentinel-2A spectral indices, and spatial soil information. A total of 64 topsoil (0–30 cm) samples collected during a soil survey in Ganyange Ward, Tarime District, Tanzania, were analysed in the laboratory. Predicted soil properties included soil reaction (pH), soil organic carbon (SOC), total nitrogen (N), available phosphorus (P), exchangeable base cations (Ca, Mg, and K), cation exchange capacity (CEC), and soil texture fractions (sand, silt, and clay). Model performance was evaluated using spatial cross-validation, where approximately 75% of samples were used for training and 25% for validation in each fold. Accuracy metrics included root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R²). The RF model demonstrated moderate to strong predictive performance for some soil properties, with positive R² values ranging from 36.4% to 78.7%. The highest predictive accuracy was obtained for electrical conductivity (EC) (R² = 78.7%) followed by silt content (R² = 67.1%), and soil pH (R² = 57.4%). Moderate performance was observed for P, SOC, exchangeable K, and clay content. The generated spatial soil property maps indicate moderately favourable soil conditions for tea cultivation, characterized by acidic environments (pH 4.2–4.8) and non-saline conditions. These maps provide a valuable decision-support tool for land-use planning and sustainable tea management.