Study on Urban Remote Sensing Classification Based on Improved RBF Network and Normalized Difference Indexes

Abstract

Aiming at the complexity of ground objects in urban area, and the difficulty in distinguishing ground objects using spectral characteristics, we extracted normalized different indexes, namely Modified Normalized Difference Water Index (MNDWI), Soil Adjusted Vegetation Index (SAVI ) and Normalized Difference Building Index (NDBI) , as the key auxiliary information for land use classification of urban area. To solve problems of RBF neural network, such as local minimum values and discrete output value in output layer, we used max-min distance means to initialize RBF center, and introduced equilibrium factor into Gauss function to improve RBF neural network learning algorithm. On this basis, a new urban area classification model was proposed based on improved RBF network and normalized difference indexes. At last, NanChong city in SiChuan province of China was taken as the study area, and TM images was used as experiment data to test the model proposed in this paper. The results showed that, based on the improved RBF network, with the help of spectral band information, the classification overall accuracy was 89.97%, Kappa coefficient was 0.88; using both spectral band information and normalized difference indexes, the classification overall accuracy was 95.02%, Kappa coefficient is 0.94, the classification overall accuracy was improved by 5.05%. Also, the experiment results showed that, with the help of spectral band information and normalized difference indexes, the classification overall accuracy of MLC, BP and improved RBF network was 90.12%, 93.63%, 95.02%, respectively, which means RBF has an advantage of fusing geological parameters in classification.