关键词: 土壤有机碳, 数字土壤制图, 机器学习, 极端梯度提升树（XGBoost）, 人工神经网络（ANN）, 地理加权回归（GWR）, 海南岛

Abstract: In view of the key role of soil organic carbon (SOC) in maintaining the stability and the carbon cycle balance of tropical ecosystems, Hainan Island was selected as the research area. Based on the data of 884 soil organic carbon sampling points in Hainan Island, combined with the soil survey data, climate data, biological data and multi-source remote sensing data (including Landsat 5 optical images, ALOS-1 SAR data, etc.) of Hainan Island, the random forest algorithm (RF) was used for feature selection. The prediction accuracy of extreme gradient boosting tree (XGBoost), artificial neural network (ANN) and geographically weighted regression (GWR) models were systematically compared and the corresponding SOC spatial distribution was generated. On this basis, the spatial distribution characteristics of SOC in Hainan Island were analyzed. The results showed that the spatial distribution of SOC was affected by many environmental factors such as climate, topography and parent material. The importance of the simple ratio index (SR) and the structure insensitive pigment index (SIPI) in biological data was more than 25%, the importance of elevation was the third, the importance was 18%, and the importance of each climate data was about 10%. The SOC content in the central mountainous area, the eastern coastal area and the southern mangrove area of Hainan Island was high, while the SOC content in the western coastal area was low. The carbon storage of different land use types showed forest soil (10.10 g/kg)>wetland soil (8.91 g/kg)> farmland soil (8.90 g/kg) > grassland soil (8.19 g/kg) > urban and construction land soil (8.04 g/kg).

Key words: soil organic carbon, digital soil mapping, machine learning, extreme gradient boosting（XGBoost）, artificial neural network（ANN）, geographically weighted regression（GWR）, Hainan Island