Surface Soil Organic Carbon Estimation Based on Habitat Patches in Southwest China

Abstract

High-precision digital soil mapping in complex terrain is challenging. This study proposed a new method using the partitioning around medoids clustering algorithm to partition the study area into distinct habitat patch types. Utilizing multisource data and three machine learning models, we estimated soil organic carbon (SOC) content in southwest China. Results showed higher SOC content (0–15 cm) in the southwestern mountains and the northwestern plateaus of Sichuan, while lower in the Sichuan Basin. The prediction uncertainty exhibited a similar pattern. Topographic and climatic variables played crucial roles in SOC estimation. Among the three machine learning models, RF and XGBoost demonstrated higher simulation accuracy than SVM (R2 increased by 2.86%–82.35%). Using the RF feature selection (FS) method to select optimal factors as model input variables improved simulation accuracy compared with using all factors or selecting based on Pearson correlation analysis (R2 increased by 1.75%–64.71%). The study found that a hybrid model based on different habitat patches achieved higher accuracy than the single model for the whole study area (for example, with RF FS method and modeling, R2 increased by 2.17%–34.78%, and RMSE decreased by 2.19%–28.80%). These findings enhanced the accuracy and refinement of existing mapping in southwest China (compared to SoilGrids 1 km and SoilGrids 250 m products, R2 increased by 20.45% and 317.73%, and RMSE decreased by 35.51% and 60.49%). Such improvements better characterized the spatial variability of SOC and provided important implications for future soil carbon stock accounts in complex terrain areas.

Keywords

• Digital soil mapping (DSM)
• feature selection (FS)
• habitat patches
• soil organic carbon (SOC)