[Remote sensing estimation of natural forest biomass based on an artificial neural network]

Based on Landsat TM images and with the natural forest area of Wangqing in Jilin Province as a case, a nonlinear RS (remote sensing) modeling system of forest biomass was built by using a back-propogation artificial neural network (B-P ANN). In addition to RS data, the factors representing terrain conditions, such as elevation, slope, aspect and site type, were also included as independent variables in the modeling system. The standard B-P ANN was made more robust by reducing the size of input data and by improving the training algorithms, thereby leading to faster convergence speed and stronger capabilities of self-study and self-adaptation. The simulation results showed that the robust B-P ANN was able to utilize previous knowledge of data sets, and to automatically determine reasonable models. Model predictions of forest biomass were successful, with the mean relative errors and the mean absolute of relative errors for needle-leaved, broad-leaved, and mixed forests being -1.47%, 2.38% and 3.56%, and 6.33%, 8.46% and 8.91% , respectively. A forest biomass distribution map was derived, and the overall accuracy of the map was 88.04%.