A zoning-based machine learning framework for accurate soil organic matter prediction across Mollisol and non-Mollisol regions

Soil organic matter (SOM) is a core indicator of soil fertility and ecosystem function. However, in regions where Mollisol and non-Mollisol coexist, high-precision spatial mapping faces significant challenges due to pronounced terrain heterogeneity and redundancy in high-dimensional covariates. This study proposes a “remote sensing zoning-feature selection optimization-random forest (RSZ-FSO-RF)” framework. By integrating Landsat-8 multi-temporal imagery from 2014–2023 with topographic and climatic factors, and leveraging the Google Earth Engine (GEE) platform, it achieves high-precision remote sensing zoning of Mollisol and non-Mollisol areas (overall accuracy: 92.13%, Kappa coefficient: 0.70). Subsequently, local Random Forest (RF) regression models were established within each zone for SOM prediction, with predictive variables optimized using recursive feature elimination (RFE). Results demonstrate that compared to FAO-zone-based modeling, the RSZ-FSO-RF framework significantly enhances prediction accuracy (R²=0.619, RMSE=6.849 g kg^–1). And further feature optimization continued to enhance model performance (R²=0.627, RMSE=6.781 g kg^–1). Notably, optimal predictor combinations varied significantly across zones, with SOM spatial variability generally higher in non-Mollisol areas than in Mollisol regions. By organically integrating remote sensing zoning with feature selection, this framework effectively mitigates covariate redundancy while accounting for local heterogeneity, significantly enhancing the accuracy and stability of high-resolution SOM mapping. Furthermore, this study provides scientific basis and decision support for soil resource management and sustainable agricultural development under complex topographic conditions.