SAR Image Classification Based On Statistical Distribution And Deep Belief Network

Synthetic Aperture Radar breaks the limitation of weather?light and other conditions and can obtain a large amount of information.The obtained information is applied to urban planning,cover classification,disaster prevention,environmental risk assessment,urban detection and many other aspects.In the field of SAR image processing,image classification is a hotpot of SAR image understanding and interpretation system.This paper proposed a new SAR image classification method which is based on DBN and superpixel segmentation.Firstly,according to the statistical distribution of SAR image,we proposed a new type of Restricted Boltzmann Machine,then we introduce space information,combine the preliminary classification result with the superpixel segmentation result in order to refine the classification result.The main work of this paper is as follows:(1)We utilize the original pixel values of SAR image instead of designing complex low-level features,the method of this paper make full use of the characteristic that DBN can obtain features by means of unsupervised manners,as a result,we can obtain a kind of feature that is discriminative and suitable for SAR image classification.(2)A new type of RBM,is proposed and is applied to SAR image classification.Firstly,the deduction of the new type of RBM is described in detail,and this new type of RBM is utilized to stack a DBN,then the stacked DBN is trained and fine-tuned using the input matrix,the new type of RBM combines the SAR image statistical properties with the characteristics of classical RBM,compared to traditional RBM,this new type of RBM can fit the data of SAR image better.The experiment results prove that our method can get a higher classification accuracy.(3)We proposed a SAR image classification method that is based on superpixel optimization algorithm.We make a further optimization of the classification result by combining the preliminary classification result with the superpixel segmentation to reassign labels in one superpixel.The method of this paper can reduce the wrongly classified pixels and divide the homogeneous area more effectively by introducing spacial information.The experiment results shows the advantages of this method and theimprovement of homogeneity of region.