Digital Soil Mapping of Topsoil Organic Matter in Southern Hilly Areas Based on GNNWR Modeling

Objective The aims were to screen suitable methods for digital mapping of soil organic matter (SOM) in the southern hilly region, and to provide a scientific basis for clarifying its spatial distribution characteristics and estimating soil carbon storage.

Method Taking the SOM contents of 1220 surface soil samples (0 - 20 cm) from the soil testing and Formulated Fertilization Project in Ningxiang as the research object, the influence of environmental factors was explored, such as climate, topography, vegetation and parent material on the spatial heterogeneity of SOM distribution. Spatial prediction models for SOM were established using Geographically Neural Network Weighted Regression (GNNWR), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Geographically Weighted Regression (GWR), Multiple Linear Regression (MLR) and Ordinary Kriging (OK), and the spatial distribution map of SOM in the study area was drawn.

Result The GNNWR model achieved the highest R² (0.60) and the lowest RMSE (7.44 g kg⁻¹) on the validation set, with a significantly higher fitting accuracy than the other six models, and could better capture the spatial distribution heterogeneity of SOM under complex topographic conditions. The high-value areas of SOM in the study area were distributed in strips in the east and central parts, while the SOM content in the high-terrain areas in the west and south was relatively low.

Conclusion Climatic, topographic, biological and parent material factors are important variables affecting the spatial prediction accuracy of SOM content. SOM spatial mapping based on GNNWR performs excellently under complex geographical environment conditions, and can provide new solutions and ideas for spatial mapping of soil properties.