Mapping soil organic matter contents under the impact of soil erosion intensity

Long-term soil erosion exacerbates the progressive depletion of soil organic matter (SOM). Accurate assessment of long-term SOM dynamics therefore requires explicit consideration of erosion as a key factor. This study investigated the spatio-temporal dynamics of SOM in northeastern China by incorporating the effects of varying erosion intensities. Specifically, Landsat time-series data combined with environmental variables were used to map SOM changes over multiple decades. A novel time-series prediction framework was developed, integrating Stacking and local regression strategies. By combining the nonlinear flexibility of machine learning, the spatial adaptivity of geographically weighted regression (GWR), and the feature-learning capacity of convolutional neural networks (CNN), the framework improved SOM mapping accuracy and robustness, raising R² from 0.582 to 0.639 and reducing RMSE from 13.39 g kg⁻¹ to 12.45 g kg⁻¹. Incorporating local regression via zonal modeling of erosion zones further enhanced performance, achieving an R² of 0.715 and an RMSE of 10.83 g kg⁻¹. Moreover, uncertainty was reduced from 1.45 ± 0.05 g kg⁻¹ to 1.26 ± 0.06 g kg⁻¹, significantly improving mapping stability and prediction reliability. These results demonstrate that local regression, by explicitly accounting for erosion heterogeneity, effectively mitigates spatial variability in SOM, particularly under conditions of limited and unevenly distributed training data. This study provides a valuable reference for long-term SOM monitoring and sustainable cropland management.