High-resolution soil organic carbon mapping in the Yellow River Delta utilizing Sentinel-2 composite and geographically weighted machine learning algorithms

Abstract. Soil organic carbon is vital for climate change mitigation and soil fertility enhancement. We aim to achieve high-resolution mapping of soil organic carbon density (SOCD) in the Yellow River Delta using Sentinel-2 satellite imagery and machine learning algorithms. We evaluated the potential of exposed soil composite reflectance (ESCR) for SOCD mapping, incorporating four machine learning algorithms: random forest, artificial neural network, geographically weighted random forest, and geographically weighted artificial neural network (GWANN). The study compared the predictive performance of these algorithms, spotlighting the capability of GWANN methods in addressing spatial nonstationary. The GWANN model incorporating ESCR achieved the highest accuracy (R2=0.50). The spatial distribution trends and uncertainties of SOCD mapped using random forest and GWANN were examined, revealing the robustness of GWANN in generating accurate SOCD maps and the superior spatial scalability and transferability of random forest.