Polarization SAR Image Classification Based On Deep Bayesian Learning

A pixel in a polarimetric synthetic aperture radar(Pol SAR)image corresponds to a feature in a small area on the ground,and the pixels of the same feature form a continuous area with certain structural and texture characteristics.In the polarization SAR image classification task,the image classification method usually adopted is semi-automatic,that is,the tag of the whole image is inferred automatically based on some artificial supervision information.Due to the large amount of speckle noise and mixed pixels in polarified SAR images,the monitoring information given manually in these semi-automatic classification methods is limited and not completely reliable.Therefore,how to learn effective information from these limited and not completely reliable monitoring tags becomes an important task.However,how to combine the effective monitoring information learned with the statistical characteristics of polarimetric SAR data and the structural characteristics of images to more effectively infer the labels of the whole image has become a difficult point in the classification of Pol SAR images.This project aims at the theory and application of Pol SAR image classification,and relies on the general project of national natural science foundation of China: ”deep collaborative representation learning and classification of Pol SAR image based on brain inspiration(No.61671350)”,and makes the following exploratory and original work.Mentioned earlier,in fact,texture and structure characteristics of the image as well as artificial tags limited not accuracy can be regarded as a priori knowledge of the Pol SAR image data classification questions,the prior knowledge will be from the prior bayesian model,likelihood,and posterior three angles,through a combination of deep learning excellent learning ability and bayesian learning excellent knowledge reasoning ability,realize each depth of the bayesian learning model.More specific innovative contributions are described below:·In order to solve the problem of limited and not completely reliable monitoring information given manually,two different generative/discriminant mixed deep bayesian learning models are proposed to model the data and classification likelihood function,and the maximum posterior distribution is achieved by combining the prior information so as to realize the effective classification of polarized SAR images.First,generation term is used to fit the likelihood distribution of Pol SAR data,and discriminant term is used to classify the likelihood distribution.By learning the statistical distribution characteristics and image characteristics of the data at the same time,it can make up for the limited supervision information,so as to get a high classification accu-racy with less supervision information.Secondly,a robust classifier learning strategy is proposed.Under the framework of learning based on bayesian dictionary,the influence of unreliable information on classification performance is reduced by using the prior distribution assumption of the incomplete reliable supervised information as the classification likelihood function,so that the model can obtain a high classification accuracy when the marking information is not completely accurate.·According to structural features of terrains in Pol SAR image,three different deep bayesian learning model are proposed to solve the Pol SAR image classification problems in some of the theory and application: first,in view of the image number of pixels of different categories in the imbalance problem,put forward a kind of price sensitive hidden depth of the bayesian learning model.By using different hidden spaces to describe the likelihood distribution of data in the feature domain,and by weighting different coefficients to the likelihood functions in different spaces,different classification bias between classes can be induced,so as to realize the effective classification of unbalanced classes.Due to the adaptive computing characteristics of the cost coefficient,the proposed algorithm has good performance for the balanced and unbalanced Pol SAR data classification.Secondly,aiming at the problem of inaccurate parameter point estimation in the method based on statistical distribution in polarimetric SAR classification model,the distribution parameters in Wishart classifier of Pol SAR were learned by variational learning from more pixels.The mixture Wishart likelihood distribution is established for the data,and the lower bound of the variational distribution of the real data is improved by the variational bayesian network.The categories that cannot be correctly distinguished by Wishart classifier can be correctly classified in this model.Meanwhile,the tag priori modeling of regional consistency is carried out to realize the transformation from maximum likelihood classification to maximum posteriori classification.Finally,a generative deep bayesian learning model is proposed to complete the structured tag matrix.By learning a likelihood function for the label distribution of Pol SAR images,we can deduce the label of the whole image from the limited supervision information,instead of classifying the pixels in the whole image.Due to the processing of image sampling,the algorithm not only reduces the computational complexity,but also significantly improves the classification accuracy under the same supervised information.