Research On SAR Image Classification Based On Deep Learning And Noise Analysis

Synthetic aperture Radar(SAR)can work under all-weather condition and provide abundant information of ground surface.Thus,SAR images have prominent advantages in target detection and classification.SAR images are widely used in agricultural assessment,military reconnaissance,environmental monitoring,disaster warning and geological exploration.Therefore,SAR image interpretation has become an important research topic in the field of remote sensing.SAR image classification is one of the key researches of SAR image interpretation.In recent years,SAR image classification has received extensive attention of academics.However,due to the unique imaging characteristics of SAR,it is difficult to interpret SAR images with traditional research methods.At the same time,the existence of speckle noise severly affects the accuracy of SAR image classification.Therefore,it is of great theoretical and practical value to develop SAR image classification algorithms.In this paper,the SAR image classification algorithm based on deep learning is studied,and the influence of speckle noise is discussed.By studying the multi-level features of SAR images,the deep learning model can gradually learn high-level and features from SAR iamges.However,due to the existence of coherent speckle noise,the image content and features are severly affected.With speckle noise reduced,deep learning model can extract more effective features.In view of the above problems,this paper summarizes the effects of speckle moise on SAR image classification with deep learning method applied for SAR images classification.The main work of the article is as follows:1)The influence of speckle noise on the deep learning model for SAR image classficatiuon is studied.Firstly,a typical SAR image classification algorithm based on deep belief network is established.As to the labled regions,festures of different types are studied through the deep belief network.Thus,the classification of unknown types of SAR images is completed by the learned network.In view of the influence of speckle noise in SAR image,three denoise methods are applied to suppress the speckle noise,which tend to compare the influence of noise on the deep belief network.The article also conduct the comparative study on the simulated SAR image classification with different noise levels.2)A novel SAR image classification algorithm combining deep belief network and region filter is proposed in this article.The region filter is proposed to reduce speckle noise while preserve boundary information.Then,through the unsupervised feature learning and supervised fine-tuning of deep belief network,the classification is well completed.The approach can effectively resist noise interference and learn discriminative features.Experimental results demonstrate that the proposed method provides fine improvements in noise immunity and classification accuracy.Moreover,the proposed algorithm shows better classification ability in boundary areas.In order to verify the performance advantage of SAR image classification algorithm proposed in this paper,the comparison experiment is carried out on both simulated and real SAR images.The experimental results show that,compared with other SAR image classification algorithms,the proposed algorithm has higher recognition accuracy and better classification effect for different regions of SAR.