Improving the accuracy of soil organic matter mapping in typical Planosol areas based on prior knowledge and probability hybrid model

Abstract

The use of remote sensing techniques for mapping soil organic matter (SOM) in black soil regions is well established. However, in areas where Planosols are interspersed with non-Planosols, tilling impacts the soil spectra of tilled soils at varying times and to different extents. As a result, errors may arise when modeling Planosols and non-Planosols collectively using conventional methods. This study developed a probability hybrid model specifically designed for the interlayered zones of Planosol and non-Planosol soils to accurately reflect the content and spatial distribution of SOM. A total of 712 topsoil samples were collected from the 852 Farm, a typical area with the interlayered zones of Planosol and non-Planosol soils in northeastern China. Cloud-free Sentinel-2 images were obtained during the bare soil period from April to May between 2021 and 2023. The spatial distribution of Planosol was detected, and the probability of soil classification was calculated using a random forest model. Based on soil classification probabilities, global models, multi-temporal ordinary hybrid models, and multi-temporal probability hybrid models were developed respectively. The results of SOM mapping using these different strategies were compared. Under seasonal reductive leaching, Planosol exhibits a distinct eluvial horizon beneath the topsoil. Long-term tilling leads to the mixing of this eluvial horizon with the topsoil in Planosol, resulting in spectral characteristics that differ significantly from those of other soil types. Accordingly, we propose a new remote sensing index—the Normalized Difference Planosol Index (NDPI), to reflect the upturning degree of the eluvial horizon and get “whiteness degree” information. We evaluated the effect of adding this index as an input on the detection of Planosol and the accuracy of SOM mapping. The results of the study show that (1) May is the optimal time window for SOM mapping and Planosol detection in the typical interlayered area of Planosol and non-Planosol soils. (2) Based on the random forest model combined with multi-period May bare soil images can accurately detect the spatial distribution of Planosol with the highest accuracy, the overall accuracy is 97.66 %; (3) The hybrid models outperform the global model, with the probability hybrid model achieving the highest accuracy (R²=0.8056, RMSE=4.2869 g/kg) and the mapping is more continuous and smoother. (4) The inclusion of NDPI improves the accuracy of Planosol spatial distribution detection and SOM mapping in Planosol areas, resulting in an increase in the Kappa coefficient by 0.0168 and an improvement in R² by 0.0122. The present study innovatively utilizes remote sensing imagery to monitor Planosol, thus expanding the application of remote sensing technology in digital soil mapping.