Estimating Impervious Surfaces Using Multi-Endmembers SOM Neural Network

Because of the population explosion and industrialization, the regional ecological environment and sustainable development of society will be affected by the aimless accelaration of the pace of urbanization. The impervious surfaces which characterized the urbanization and the ecological environment was just a important parameter. Due to strong spatial heterogeneity of urban areas, RS technology was used to extract surface parameters quantitatively become a hotspot of current researches one.Taking Changsha, Hunan as an example, this study discussed V-I-S model for the problem of mixed pixels in medium resolution remote sensing images. The SOM neural network was used to classify the urban land cover of Changsha. We combined with the fuzzy set theory and SOM to decomposed the mixed pixel,and then,we combined with the spectral correlation matching and decomposed the mixed pixel with multi-endmembers. The main works are as follows:(1) In the support of MATLAB R2009a,we improved the unsupervised SOM,making it become a supervised one, The experimental results show that the image classification of urban land cover based on SOM neural network which we used, the overall accuracy and Kappa coefficient of its results are higher than the results of traditional methods.(2) We combined with the fuzzy set theory and SOM to decomposed the mixed pixel.This method was realized in the environment of MATLAB R2009a. The results of this method are four fraction maps of land use type,and then,we would obtain the fraction map of impervious surfaces.The RMSE of its result was 0.2215.(3) Based on the SOM for the decomposition of mixed pixel, we combined with the spectral correlation matching, and decomposed the mixed pixel with multi-endmembers. This method was realized in the environment of MATLAB R2009a. The results of this method are four fraction maps of land use type,and then,we would obtain the fraction map of impervious surfaces.The RMSE of its result was 0.1916.The experimental results show that the multi-endmembers SOM neural network provides a promising superiority in the decomposition of mixed pixel.