SAR Image Terrain Classification Based On The Deep RBF Network

Currently, most of Synthetic Aperture Radar(SAR) image terrain classification algorithms which can be regarded to classify the characteristics through shadow learning only learn the low-level characteristics and usually gain low recognition rate. Therefore, how to design complex classifier becomes a hot research. This paper proposes a Radial Base Function(RBF) data classification model which has three hidden layers based on the idea of deep learning. This model transforms the input data from low-dimensional space to high-dimensional space through the first hidden layer and the input data become linear and separable characteristics, and then high-level characteristics can be extracted by self-learning through the second hidden layer. Then these high-level characteristics will be mapped to higher dimensional linear separable space to get higher-level characteristics to classify. The deep RBF data classification model、improved deep RBF classifier and optimized deep RBF classifier can be used on classification problems of UCI database,SAR image terrain data and texture image with multi-feature and multi-class to gain higher robustness and better classification performance than these surface learning such as Support Vector Machine(SVM) and RBF. The paper does following works:1. A SAR image terrain classification method with RBF network and Sparse Auto Encoder(SAE) network has been proposed to solve the problem that the feature learning of present shadow machine learning method is not good and the classification accuracy is low. The algorithm idea based on deep learning can expand monolayer RBF network and SAE network to three-layer deep neural network model to train the SAR image texton feature gained, and the classification accuracy is higher than ones by shadow learning methods(SVM、RBF) to prove that the method is feasibility and effectiveness.2. A deep RBF classifier with RBF network and Restricted Boltzmann Machines(RBM) network has been proposed. Because of the limitation of SAE network in feature extraction, especially the drawbacks of lower classification accuracy than shadow learning methods on the classification of multi-feature and multi-sample database, the SAE network can be replaced by RBM network. The classifier can get higher classification accuracy than SAR image terrain classification method based on RBF network and SAE network and shadow learning methods(SVM 、 RBF) on theclassification problems of SAR image、multi-feature and multi-class UCI database and texture image.3. The learning and optimization algorithms of evolutionary neural network have been finished. To solve the problem of the long training time and complex parameter adjustment of deep RBF classifier based on RBF network and RBM network, an optimized method based on Genetic Algorithm、an optimized method based on Particle Swarm Optimization and an optimized method based on immune have been proposed. Through these optimized methods to optimize clustering center and the center of the range of RBF network, the deep RBF classifier obtains better data classification performance, shorter training time and lower parameter adjustment complexity.