Estimation Accuracy Obviously Increased of Soil Organic Carbon Content Based on Habitat Patches Scale Using Interpretable Machine Learning in Southwest China

Although machine learning models have been successfully applied to estimate soil organic carbon (SOC), the lack of interpretability research has resulted in the absence of mechanistic explanations for these models. Therefore, in this study, we adopted the following workflow: first, we divided the whole study area into several subregions using land cover types and habitat patch types based on cluster analysis, and then the model driving features of SOC estimation were selected using the recursive feature elimination (RFE) technique. Second, we compared the estimation accuracy and stability of three machine learning models, and explored the optimal feature combinations and model architecture. Finally, the Shapley additive interpretation (SHAP) model was introduced to interpret the selected SOC content estimation model to verify its reliability. The results showed that the divided modeling framework based on habitat patch type achieved the best performance in SOC content estimation: the maximum increase in R 2 reached 27%, and RMSE decreased by 20% (Habitat Patch 3 vs whole study area). The optimal model combination based on RFE‐RF demonstrated good estimation accuracy across the four habitat patches, with R 2 values of 0.36, 0.49, 0.62, and 0.59, and RMSE values of 22.04, 9.59, 11.67, and 14.57, respectively. Meanwhile, we analyzed the impacts of each input feature on the model prediction from both global and local perspectives. The results showed that DEM, precipitation, and air temperature were the key factors influencing the estimation of SOC content. The SHAP model successfully revealed the relationship between feature variables and SOC, effectively compensating for the limitations of the machine learning model in terms of interpretability.