Mapping forest aboveground carbon stock of combined stratified sampling and RFRK model with mean annual temperature and precipitation

Accurately estimating forest aboveground carbon stock (ACS) is essential for achieving carbon neutrality. At present, most non-parametric models still have errors in estimating carbon stock in regions. Given the autocorrelation inherent in spatial interpolation, combining non-parametric models with spatial interpolation offers significant potential. In this study, we combined the random forest (RF) with the ordinary kriging and co-kriging of the mean annual temperature, precipitation, slope, and elevation to establish the random forest residual kriging (RFRK) model. Meanwhile, we also developed the multiple linear regression residual kriging (MLRRK) model and the random forest residual kriging (RFRK) model. Finally, we selected the optimal model for the estimation and mapping of the ACS. The results indicate that: (1) the model achieves an R² of 0.871, P of 90.4%, and RMSE of 3.948 t/hm²; (2) the RFCK model with mean annual precipitation (RFCKpre) outperforms the one with mean annual temperature (RFCKtem), while the RFOK model exhibits the lowest accuracy; (3) the RFCKpre exponential model has the highest accuracy, with the highest R² of 0.63 and RI (0.23), the lowest RMSE of 9.3 and SSR (41,612). These findings suggest that the RFRKpre model has improved the accuracy of estimating the ACS of regional forests.