Research On PolSAR Image Denoising And Classification Based On Deep Learning

Polarimetric SAR images have multiple polarization channels,which can obtain rich ground object information.Denoising and classification,as the main means of polarimetric SAR image interpretation,have certain scientific significance and practical value.Aiming at the problems of unable to obtain clean SAR images and low efficiency in identifying ground features in SAR images,this paper takes GF-3 image data from three areas in China as the research object,and the polarimetric SAR image denoising model was established by deep learning with the help of N2 N network model,and the polarimetric SAR image classification model was established by deep learning with Deep Lab V3+ algorithm,which realized the rapid denoising and accurate classification of polarimetric SAR image.The main research contents and results are as follows:(1)Polarimetric SAR image denoising based on N2 N network modelThe speckle noise of SAR image is a multiplicative noise,which is essentially different from the additive noise of optical image.The noise removal method of optical remote sensing image is often not applicable to SAR image.Therefore,this paper proposes a polarimetric SAR image denoising method based on N2 N network model.Firstly,Gaussian noise and Poisson noise are added to the original SAR image to form a paired training set with the original SAR image.The open source N2 N pre-training model is used to train the denoising network model,and the denoising model of SAR image is obtained.Then,this model is used to denoise SAR images for several times,and the peak signal-to-noise ratio is used as the accuracy evaluation index.When its value tends to be stable,the denoising is stopped to obtain relatively clean SAR images.Finally,the original SAR images and relatively clean SAR images after denoising are used to train the U-Net network model,and the denoising model obtained can realize fast denoising of SAR images.The denoising effect of the proposed method was compared with Lee,Gauss and BM3 D filtering methods,and the equivalent number of looks,edge preservation index and peak signal-to-noise ratio were taken as the evaluation indexes of the denoising effect.The results show that the method proposed in this paper has the best denoising effect,which proves the feasibility and effectiveness of this method for SAR image denoising.(2)Polarimetric SAR image classification based on Deep Lab V3+ algorithmThe deep learning method can not only abstract higher dimensional features from the original data but also make full use of the advantages of three-dimensional spatial information.In this paper,a polarimetric SAR image classification method based on Deep Lab V3+ algorithm is proposed.In this method,the training set of SAR images was firstly mapped using Arcgis10.8,and then use Python to convert it to cityscapes format.Finally,the classification model is established by Deep Lab V3+ algorithm.Its advantage is that the resolution of the feature response can be effectively controlled,the receptive field can be expanded,and the semantic segmentation results can be improved without increasing the amount of computation.Meanwhile,the Decode module is introduced to gradually restore the image spatial information and capture clear object boundary,so as to improve the accuracy of semantic segmentation.This classification method was compared with Wishart unsupervised classification,SVM classification and object-oriented decision tree classification,taking the overall classification accuracy and Kappa coefficient as the accuracy evaluation indexes,the results show that the proposed classification method has the highest accuracy,which proves the feasibility and effectiveness of the proposed method for SAR image classification.This paper includes 35 figures,17 tables and 74 references.