PolSAR Image Classification Based On Spatial Information

In recent years,the high-profile remote sensing system has achieved unprecedented breakthroughs,which enables people to obtain land use information through more methods.As one of the most important ways to obtain land use information,the Polarimetric Synthetic Aperture Radar(PolSAR)shows more advantages than other traditional methods.The classification of PolSAR images is an important way of processing and interpreting PolSAR images.Therefore,the classification of PolSAR images has always been the focus of the global researchers.Thanks to the heat of artificial intelligence,more and more scholars have introduced the idea of deep learning to the classification of PolSAR images.However,PolSAR images are different from optical images,and there must be several problems.First,the data format of PolSAR image is different from that of optical image.The optical image is formed by the visible light sensor,and each pixel is represented by a gray value.So the gray value can be the original feature,and it can be the input of the deep learning model.But the PolSAR image is formed by the microwave sensor and each pixel is represented by a polarization scattering matrix which cannot be the input of deep learning model.Second,There are serious speckle noises in PolSAR images,and the speckle noises affect the classification.Finally,Labelling pixels is a time-consuming process and only a small number of labelled pixels are available.However,deep learning models usually need a large number of labelled data.The spatial information in PolSAR image is a very important information.In this paper,we use the spatial information to try solving the above three problems,and several classification algorithms are proposed based on spatial information.(1)In the second chapter,a new PolSAR image classification method based on superpixels and stacked sparse autoencoder is proposed.This method not only makes deep learning model has a reasonable input but also suppresses the speckle noises.First,the PolSAR image is oversegmented into superpixels.Second,a real-valued feature vector is constructed for each pixel in the image.Third,the stacked sparse autoencoder is trained.Finally,the stacked sparse autoencoder is used for prediction.In this method,we take the superpixel spatial information into account and the classification accuracy is improved.(2)For the fact that the number of labelled pixels is very small,a new PolSAR image classification method based on semi-supervised learning and ensemble learning is proposed in the third chapter.There are three classification models in this method.First,a real-valued feature vector is constructed for each pixel in the image.Second,the training samples are increased by semi-supervised learning.When the three models are trained,the three models are used for prediction,and the final results are obtained through the integration of the three results.In this method,the superpixel spatial information is taken into account,which ensures the reliability and diversity of the added training samples.Ensemble learning is used in this method,so the final accuracy is improved.The problem that the number of labelled pixels is very small is effectively solved in this method.(3)A new PolSAR image classification method based on superpixels and convolutional neural network is proposed in the fourth chapter.In this method,the speckle noises are further suppressed.First,a real-valued feature vector is constructed for each pixel in the image.Second,feature maps are constructed by using the neighborhood information of a pixel.Third,the convolutional neural network is trained.Finally,the convolutional neural network is used for prediction.In this method,the superpixel spatial information and neighborhood information are both taken into account.So the speckle noises are further suppressed and the classification accuracy is improved.