Research On SAR Image Terrain Classification Based On Deep Networks

Synthetic aperture radar(SAR)utilizes microwave imaging technology and has the advantages of working all day and all time,which has been widely used in military and civilian fields.SAR image terrain classification is one of the important applications of SAR image interpretation and analysis.Its purpose is to distinguish the category which each pixel belongs to.With the rapid development of SAR imaging technology,SAR images have evolved from low and medium resolution to high resolution.Compared with traditional lowand medium-resolution SAR images,high-resolution SAR images have many remarkable features such as rich structural information,complex scene information,and large datasets.Therefore,traditional low and medium-resolution SAR image classification methods are difficult to meet the application requirements of high-resolution SAR image classification.And designing effective features is the key to solve high resolution classification problems in SAR image.To address the above problems,this paper studies the SAR image classification based on traditional texture features,SAR image classification based on convolutional neural networks,SAR image classification based on superpixel and convolution-deconvolution networks,SAR image classification based on generative adversarial networks.The contents of this paper are arranged as follows.1.The first part introduces the research background and significance of SAR image terrain classification,the research status at home and abroad,the development of deep learning and the main research contents and structure of this paper.2.The second part studies the SAR image classification based on the texture features.First,this thesis briefly introduces some common texture features.Then,the extraction method of multilevel local pattern histogram texture features is described in details,which can be divided into three steps: image quantization,pattern matrix splitting and histogram computation and combination.Finally,the measured SAR data is used to verify the effectiveness of the algorithm.3.The third part studies the SAR image classification using deep convolutional neural networks.First,the paper introduces the basic theoretical knowledge of deep learning,including the M-P neuron model,back propagation algorithm and convolutional neural networks.Then,the SAR image classification based on convolutional neural networks and the SAR image classification based on superpixel and convolution-deconvolution networks are introduced respectively.Finally,the above two schemes are verified by utilizing the measured SAR images.Compared with the traditional texture features,the SAR image classification based on deep networks can effectively learn feature representation and obtain higher recognition rate.4.The fourth part studies the SAR image classification based on generative adversarial networks.First,it introduces the relevant theoretical knowledge of generative adversarial networks.The basic model and application of the deep convolutional generative adversarial networks are then described in details.Finally,the generative adversarial networks is applied to the SAR image classification problem,and the SAR image classification based on the generative adversarial networks is used in the experiment with the real SAR image.The experimental results show that the algorithm can still obtain satisfactory classification results when the number of training samples is limited.