High Resolution SAR Image Classification Based On Deep Learning And Wavelet Transform

Synthetic Aperture Radar(SAR)image classification is one of the hot topics in the field of remote sensing.In recent years,along with the development of remote sensing technology,SAR image resolution is rapidly improving.In the later stage of interpretation,high resolution SAR images are also more difficult than low-resolution SAR images.High resolution SAR images have more complex scenes than the previous low-resolution images,and have more rich texture,contain more strong scattering points,and have more sampling points when the image is interpreted.So the classification of high resolution SAR has become very challenging.In recent years,with the development of deep learning technology,the deep learning model has shown a very good performance in the field of image processing.Deep learning has become one of the most popular areas in the field of image processing.In this paper,we use the deep learning model to extract the image depth feature information,and combined with the traditional wavelet transform,so as to obtain complete and comprehensive feature representation of the original SAR image,and finally combine the classifier to complete the SAR image classification.The main results in this paper are as follows:(1)Introduce the contractive auto-encoder(CAE)model into the field of SAR image classification,and the CAE model has excellent reconstruction and good robustness for the original image.We use these two characteristics to build a stack contractive autoencoder(SCAE)network to obtain advanced features of the image.Through the construction of the stack network,it can effectively improve the effect of feature extraction,so as to learn deeper level of features from the original input image,and provide more effective and excellent features for the later image classification.At the same time,two parallel SCAE networks are constructed for different initial image features,and the overall network is finetuned to further improve the classification effect.(2)Based on the traditional convolution neural network,a new type of unsupervised feature learning model is constructed.By constructing a multi-level network,we can extract the effective feature information from the original input image data.The hierarchical structure is robust to the small perturbations in the input image data,and the multi-level network is constructed so that each filter in the network can combine low-level features into higher-level feature representations to extract deeper level of advanced features.At the same time,we use the convolutional auto-encoder to pre-train the parameters of convolution layer of the network model to obtain better initialization parameters.(3)We combine 4the deep learning model with the traditional wavelet transform method to give full play to their respective advantages.The stationary wavelet transform has good translation invariance,and it has no restriction on the size of the input image.Through the stationary wavelet transform decomposition of the original SAR image,the low frequency and high frequency image information of the image can be extracted,and these feature information can express the scene information in the image completely.We construct two parallel stack contractive auto-encoder networks by using the high-frequency image information and the low-frequency image information obtained by the stationary wavelet transform as the input,and through the fine-tuning of the whole network system,the image classification is finally completed.The experiments show that the final classification effect is very satisfactory.Texture features such as wavelet energy feature is a very classic image feature representation.We use the two-dimensional discrete wavelet transform to decompose the original SAR image to obtain the texture features of the image.We combine the extracted texture feature information with the features extracted by our proposed new type of model based on the traditional convolutional neural network,so as to complete and comprehensively represent the feature information of the original SAR image,which provides a very favorable basis for the later classification.