POL-SAR Image Classification Based On Deep Learning

Polarimetric synthetic aperture radar(Pol SAR)is an advanced means to obtain remote sensing information,more and more researchers pay their attention to this filed.Pol SAR is used to measure the scattering properties of each resolution unit in four different polarizations to obtain the corresponding polarization information of the target.Therefore,the Pol SAR can record more target electromagnetic scattering information.At the same time,due to the characteristics of the Pol SAR image,it can provide more basis for interpretation and analysis of the image,thus it has more research value than ordinary remote sensing image.Pol SAR image classification is an important research content of Pol SAR image interpretation,and it is also one of the research hotspots.Pol SAR has a wide range of applications in the civil and military fields,so it is of great significance to further study the Pol SAR images.So how to classify Pol SAR images exactly is particularly important.In this paper,we use deep learning to classify the Pol SAR images,and good results are obtained.The research contents of this paper include the following aspects:First,because of many unlabeled data available for Pol SAR data,we use deep belief network to Pol SAR image classification,which can make full use of all unlabeled data to model Pol SAR data.Firstly,the basic structure and basic principle of DBN are studied deeply.In view of the particularity of the characteristics of the Pol SAR data extracted in this paper,we extend the general binary RBM to Gaussian RBM(GRBM).Second,the traditional DBN-based Pol SAR classification method is to directly train the polarization coherence matrix T or the polarization covariance matrix C as the input of the network,this can only extract part of the characteristics of Pol SAR,many feature information can not be used,which will lead to poor classification results.In this paper,we present a multi-feature data combined with Pol SAR original coherent matrix T,polarized scattering feature and polarized SAR texture feature.Combining the scattering mechanism with spatial features of Pol SAR data,thus the classification effect has been improved.Third,in Pol SAR image processing,the data are generally vectorized,and polarized SAR data itself has a tensor structure.Because vectorization can lead to the destruction and loss of spatial information,and there will be redundant information between Pol SAR image pixels,thus,in this paper,we proposes a dimensionality reduction algorithm based on tensor PCA and a DBN algorithm,which can effectively improve the classification accuracy.