Machine Learning approaches to predict Soil Organic Carbon in the North-western Himalayan Forests

Soil organic carbon (SOC) is a fundamental indicator of soil quality and ecosystem functioning, particularly in forest ecosystems. Accurate spatial assessment of SOC is essential for understanding its distribution and supporting sustainable forest management. This study aimed to predict and map SOC distribution across the moist temperate Quercus leucotrichophora forests of Himachal Pradesh, India, using machine learning (ML) techniques and to identify the major environmental drivers of its spatial variability. A total of 100 composite soil samples (0-15 cm depth) were collected from seven districts, and SOC concentration was determined using the Walkley-Black method. A digital soil mapping framework based on the SCORPAN model incorporated 49 environmental covariates derived from Landsat 8 imagery, SRTM elevation data and WorldClim bioclimatic variables. Following multicollinearity assessment using the variance inflation factor, 28 covariates were retained for model development. Four ML algorithms, namely Random Forest (RF), Support Vector Machine (SVM), Generalized Linear Model (GLM) and Cubist, were evaluated using R², RMSE, MAE and regression slope. Measured SOC concentrations ranged from 1.21% to 9.69%, with comparatively higher values observed at selected sites in Chamba district. RF demonstrated superior predictive performance (R² = 0.85; RMSE = 0.83%), outperforming GLM, SVM and Cubist. Variable importance analysis highlighted elevation, multi-resolution index of valley bottom flatness (MRVBF), slope and topographic ruggedness index (TRI) as dominant predictors, emphasising the influence of terrain on SOC variability. The RF-based SOC predictions ranged from 2.82% to 7.71%, while uncertainty was comparatively higher in steep and topographically complex landscapes than in valley floors. Although the outputs were generated at 30 m grid resolution, their effective spatial detail remains constrained by the native resolution of the input covariates. Overall, the findings demonstrate the potential of RF-based digital soil mapping for characterising SOC patterns in heterogeneous Himalayan Forest landscapes and provide a spatial baseline for prioritising forest conservation and soil management interventions.