Modeling study demonstrates improved soil organic matter estimation using multi-factor environmental zoning in heterogeneous cropland, suggesting enhanced remote sensing monitoring.
Key Points
Develop a multi-factor zoning framework integrating crop type, soil moisture, and soil texture into automated machine learning pipelines to enhance regional soil organic matter (SOM) mapping accuracy across heterogeneous cropland.
Trained tree-based algorithms (Random Forest, Gradient Boosting Decision Tree [GBDT], AdaBoost, and XGBoost) using the Tree-based Pipeline Optimization Tool (TPOT) on soil samples collected from Nenjiang County in 2014.
Evaluated temporal robustness using an independent sample set collected from the same region in 2022.
Integrated a multi-factor zonal variable accounting for surface cover, soil moisture, and soil texture heterogeneity into the automated learning pipeline.
The GBDT model incorporating the multi-factor zoning variable achieved the highest accuracy, with an R² of 0.64, an RMSE of 8.24 g/kg, and an MAE of 6.06 g/kg, outperforming models without zoning.
Spatial mapping demonstrated a consistent east-to-west decreasing gradient of SOM and a mean increase of 2.04 g/kg from 2014 to 2022 across the region.