Abstract: Prediction of forest soil respiration is of great significance to study carbon cycling and climate change in terrestrial ecosystems. The prediction models of different scales (global, national and regional) have different functions. Due to the limited number of sample points, researches on soil respiration models of regional scale are less. In this paper, a new idea of regional modeling was proposed, that was a linear model based on the finite sample points established to modify the predicted values of large-scale forest soil respiration model. The results showed that the accuracies of the global- and national- scale models applied in Guangxi Province and Guangdong Province were decreased by 40% and 53%, respectively compared with the original models. To modify the predicted data of the two large scale models, the parameters (such as the forest type, temperature and precipitation) with different categories were selected respectively to build linear models based on the 32 sample points in this two provinces. Among the parameters, the best linear model was built by the precipitation, which improved the modification accuracy by 11%-38%. Therefore, the modeling method in this study could solve the problem that the number of sample points in the region was not enough to model. Moreover, it also better realized the regional application and optimization of large-scale mode, which was applied simply and easily for other regions base on the linear model parameters.