Optimizing Soil Organic Matter Estimation Through Multi‐Factor Zoning and Tree‐Based Automated Learning

Abstract

Soil organic matter (SOM) is a key indicator of land degradation and soil functioning. It plays an essential role in nutrient cycling, soil structure, and long‐term agroecosystem resilience. The variations of surface cover, soil moisture and soil texture heterogeneity will affect the accuracy of SOM remote sensing mapping. This study developed a zoning variable based on crop type, soil moisture, and soil texture, and incorporated this zonal variable into an automated learning model to improve SOM estimation across heterogeneous cropland. Soil samples collected from the Mollisol region of Nenjiang County in 2014 were used for model construction, while an independent set collected from the same region in 2022 was employed to examine temporal robustness. Four tree based algorithms, including Random Forest, Gradient Boosting Decision Tree, AdaBoost, and XGBoost, were optimized with the tree based pipeline optimization tool (TPOT). The GBDT model that incorporated the zoning variable achieved the best performance, with an R 2 of 0.64, an RMSE of 8.24 g/kg, and a MAE of 6.06 g/kg, outperforming the model without zoning. Spatial mapping showed a consistent east‐to‐west decreasing gradient of SOM, and a mean increase of 2.04 g/kg from 2014 to 2022. By incorporating the information of crop type, soil moisture, and soil texture, the zoning variable leads to more reliable SOM estimation across heterogeneous cropland at the regional scale.