Abstract: [Purposes]To address the issues of scarce ground samples and insufficient accuracy in carbon stock estimation in tropical cloudy forest regions,this study conducted a remote sensing inversion of carbon stock at the regional scale.[Methods]Based on Sentinel-2 multispectral imagery and WCMC carbon density data,a multi-dimensional feature set integrating spectral reflectance,vegetation indices,tasseled cap transformation components,and topographic factors was constructed.A random forest regression model was then employed to estimate carbon stock.[Findings]On the validation set,the model achieved a correlation coefficient of 0.72 and an RMSE of 16.10 tC/hm².Feature analysis indicated that the shortwave infrared bands (B11,B12)and the tasseled cap wetness component (TC_W)were the dominant predictors,with a cumulative importance exceeding 0.28,while topographic factors played a supplementary role.[Conclusions]The estimated total forest carbon stock in Xishuangbanna is approximately 158.89 TgC,exhibiting a spatial pattern of higher carbon density in peripheral areas and lower density in the central region.Carbon density increases with elevation,while low-elevation valley areas show significant reductions due to human activities．

Key words: carbon storage estimation, Sentinel-2, random forest algorithm, Xishuangbanna, tropical forest, GEE platform