Fine Resolution Mapping of Forest Soil Organic Carbon Based on Feature Selection and Machine Learning Algorithm

An accurate forest soil organic carbon (SOC) assessment aids in the ecological restoration of forest mining areas, enabling dynamic monitoring of carbon sink accounting and informed land reclamation decisions. Digital soil mapping (DSM) has enhanced soil monitoring, with machine learning and environmental covariates becoming the keys to improving accuracy. This study utilized 32 environmental variables from multispectral, topographic, and soil data, along with 142 soil samples and six machine learning methods to construct a forest SOC model for the Huodong mining district. The performance of Boruta and SHAP (SHapley Additive exPlanations) in optimizing feature selection was evaluated. Ultimately, the optimal machine learning model and feature selection method were applied to map the SOC distribution, with variable contributions quantified using SHAP. The results showed that CatBoost performed best among the six algorithms in predicting the SOC content (R2 = 0.70). Both Boruta and SHAP improved the prediction accuracy, with Boruta achieving the highest precision. Introducing the Boruta model increased R2 by 8.57% (from 0.70 to 0.76) compared to models without feature selection. The spatial distribution mapping revealed higher SOC concentrations in the southern and northern regions and lower levels in the central area, indicating strong spatial heterogeneity. Key factors influencing the SOC distribution included pH, the nitrogen content, sand content, DEM, and B3 band.