As the largest carbon sink in terrestrial ecosystems, forest ecosystems play an important role in balancing global carbon budget and mitigating the effect of global warming. Recently, the widely used approach for quantifying aboveground forest carbon storage is to use forest resources inventory data (FID) and the relationship between dominant tree species and its biomass. Based on this method, combining forest resources inventory data (FID) and remotely sensed images can lead to spatial distribution of predicted aboveground forest carbon storage at both regional and global scales. This method can well provide the results required by the relevant departments where decisions are made. However, how to combine the data of sample plots and image pixels to get the accurate spatial distribution of aboveground forest carbon is still a great challenge that scientists are experiencing. In this study, artificial neural networks were applied to develop models to predict aboveground forest carbon storage according to sample data and Landsat Thematic Mapper^TM data. A total of 930 sample plots obtained from Lin'an county forest resources inventory in 2004 were used. For each plot, aboveground forest biomass was calculated based on allometric equations of four dominant tree species from literatures. The biomass was then converted to forest carbon using standard coefficients. These values of forest carbon storage in sample plots were then used as the target values for an error back-propagation neural network (BPNN) model. At the same time, three spectral band ratios (TM₄/(TM₅+ TM₇), 1/TM₂, and (TM₄- TM₃)/(TM₄+TM₃)) computed from the Landsat TM imagery were selected based on the correlation analysis and input into the BPNN model together with the target values. In addition, suitable parameters, and structure were chosen to simulate aboveground forest carbon storage and its spatial distribution. The results showed the BPNN algorithm could accurately generate the spatial distributions of forest carbon density and changes. The obtained estimates were quite similar to the observed values at the sample locations. The root mean square error (RMSE) of the aboveground forest carbon storage for the sample plots was 5.45Mg/hm². The correlation coefficient between predicted and observed values was 0.61 (significant at the 99% level of confidence). The mean estimate of carbon density for the whole study area was 0.98Mg (10.89 Mg/hm²) which was smaller than the average from the sample plots with a relative error of only 13%. Although the RMSE remained relatively large, the predictions were more accurate compared to those from previous studies. The finding implies that artificial neural networks are a promising tool that can be used to estimate and simulate forest carbon storage and analyze forest carbon budget for large areas. However, an accurate simulation of the regional aboveground forest carbon is not only dependent on the high quality of the used images and the data of variables that are highly correlated to the aboveground forest biomass, but also determined by the algorithm that leads to the models to estimate and simulate the aboveground forest carbon. Regarding the BPNN algorithm in this paper, some inherent problems, such as being easily trapped in local minima and too slowly converged, leading to failure in searching for a global optimal solution, still exist. Consequently, further research is needed to improve and optimize the BPNN algorithm in order to obtain more accurate estimates of aboveground forest carbon at large scales.