Bp Neural Network-based High-resolution Remote Sensing Image Classification

With the development of science and technology progress, satellite sensor continued progress. Present remote sensing satellite has been able to provide, the meters level high-resolution remote sensing image data. Extracting the spectrum and texture features information of High resolution remote sensing images into scientific research achievement, in order to give service to the construction of the national defense and economic construction. It has become a key subject of the remote sensing image classification areas.This paper discusses the multilayer back-propagation (BP) neural network classification algorithm which applied to high-resolution remote sensing image classification areas. Compare with the other statistical classification algorithm, the BP neural network classification algorithm has strong learning ability, combined with remote sensing image texture, spectral, slope, slopes and other information, extracting features information more easily. It is widely used in the pattern recognition domain. But the BP neural network has some shortcomings, such as it is hard to determine the structure of the BP neural network, learning speed is slow, training is easy to fall into the local minimum. So this.paper gives several solutions, such as we use genetic algorithm to solve the network structure setting, and use change learning rate method and momentum factors method in order to accelerate learning speed and improve the classification accuracy of high-resolution remote sensing image.This paper take the Quickbird high resolution image of Beijing city as based experimental data, and use the MATLAB software to develop the improved BP neural network program. We use this program to classify the remote sensing images, then we use the error matrix to evaluate the accuracy of classification result. Compare with other classification methods, we found that the overall classification accuracy rate of improved BP neural network classification is 93.4%, which is more than traditional BP neural network 2.4%, maximum likelihood method 5.5%, minimum distance method 10.2%, ISODATA method 20.2%.To sum up, the improved BP neural network classification method get classification result quickly and a higher classification accuracy. Use the method for the classification of urban high resolution remote sensing images, seek a rapid and suitable method to classify the similar urban area. Auxiliary urban planning department understand city social economy activity development trends of urban resource planning and rational utilization guidance.