R-Modelling-Based Quantile Regression Approach for Spatial Prediction and Mapping Soil Properties in Sri Lanka

Reliable soil information is crucial for agriculture planning, decision-making, food insecurity, and environmental impact mitigation. Digital Soil Mapping (DSM) solves data gaps and offers an efficient approach to predict soil variability. This study evaluated the potential of DSM, based on the quantile regression forest (QRF-DSM) approach in the R-modelling language for populating gridded soil properties with sparsely distributed observations (soil profile data from 100 locations) and 67 environmental covariates related to soil forming factors to predict and map bulk density (BD), texture fractions, cation exchange capacity (CEC), soil organic carbon (SOC), and pH of topsoil covering entire Sri Lanka. The optimal covariates for each soil property were selected with the recursive feature elimination (RFE) algorithm, and crossvalidation was employed to assess model performance. The selected model showed a positive correlation with both Pearson’s product-moment correlation coefficient (r) and coefficient of determination (r2). This study ensured precise predictions for the topsoil layer, achieving low mean absolute error (MAE) and low root mean squared error (RMSE) for soil properties: BD (0.11, 0.15), sand (9.33, 13.23), silt (4.29, 5.65), clay (7.12, 9.84), CEC (3.67, 5.32), SOC (0.48, 0.56), pH (0.63, 0.75), and generating mean and uncertainty maps at a 250 m resolution for Sri Lanka. The predicted gridded soil property maps provide information on spatial variability for diverse applications.