Polarmetric SAR Image Classification Based On Deep SVM

The Pol SAR has been a hot area of research in recent years for the polarimetric scattering it provided.Polarimetric SAR image classification,as one of the maintasks for Polarimetric SAR image understanding,has been playing an important role in both civil and military applications.Support vector machine(SVM)is a statistical classification of an effective supervision methods, has been widely used in many fields.This method has been in the field of machine learning attention and considerable development.Based on supervised SVM algorithm polarimetric SAR terrain classification study the issue, combined with practical problems, improvements to existing algorithms, and put forward improved algorithms.1. LSSVM algorithm in the SVM have been studied,taking into account the drawbacks for using of LSSVM model to solve classification problem:high computation complexity,the solution is not sparse and the LSSVM model is highly affected by the noise in the training samples,fuzzy LSSVM has been combined with fast sparse LSSVM algorithm and an fuzzy sparse LSSVM algorithm has been proposed in the paper.due to the importance of fuzzy membership to the fuzzy LSSVM,two calculation methods were taken based on the distance from samples to category class center,ie Euclidean distance and Kernel distance.the proposed algorithm is used for polarmetric SAR data classification,view the result and compare the algorithm through comparative experiments,the proposed algorithm performance stronger.2. LSSVM kernel function have been studied. As solving nonlinear classification problem, the first task is to select qualified kernel function, the samples are mapped into a high dimensional space, in order to achieve high dimensional space linearly separable, so choose the kernel function is the key. Commonly used kernel functions for radial basis function, but at the time of fitting more complex functions, the resulting effect is not satisfactory, this paper analyzes the support vector machine kernel function condition, then the kernel function LSSVM classifier Morlet wavelet kernel function instead of the kernel function in line with the conditions, we propose a wavelet kernel LSSVM algorithm for polarimetric SAR data classification through comparativeexperiments show that the use of higher Morlet wavelet kernel function classification accuracy.3. The wavelet LSSVM algorithm proposed have been expanded.The algorithm have been combined with the network architecture of deep svm, a deep wavelet LSSVM algorithm has been proposed, the model is mainly designed two hidden units, the support vector corresponding to the lower value as the activation level of training samples, and so training the depth of wavelet kernel sparse LSSVM classifier, taking into account the time complexity of the algorithm will be solved using the Lasso algorithm second layer, and so build depth Lasso model, the depth and the depth of the wavelet kernel Lasso model sparse LSSVM algorithm UCI classification comparative data and polarimetric SAR data.