PolSAR Image Classification And Change Detection Based On Deep Learning

Polarimetric Synthetic Aperture Radar(PolSAR)image classification and change detection are two hot topics in Remote Sensing(RS),and they are widely used in monitoring the earth covering and the evolution of the earth surface.In recent years,the RS technology develops rapidly,which produces more and more highly complicated PolSAR images,and traditional methods can not meet the requirement of users to accomplish the tasks precisely and efficiently.On the other hand,various deep learning models are booming in many field-s,including nature image processing and speech processing.These successful applications show that deep learning is powerful in feature extraction and data-processing,however its application to RS is at the beginning.Different from the nature image,there are many spe-cific polarimetric properties in PolSAR image and it has the shortages of plentiful contents,serious noises,small labeled samples and so on.To solve these problems,in this dissertation we make fully use of the properties of PolSAR data and design several specific deep learning models for PolSAR data to accomplish the tasks of PolSAR image classification and change detection.What’s more,they can be used in other related tasks independently.A series of algorithms are shown as follows:1.For the particularity of PolSAR data,a new type of Restricted Boltzmann Machine(RBM)is specially defined,which we name the Wishart-Bernoulli RBM(WBRBM),and is used to form a deep network named as Wishart DBN(W-DBN).Numerous unlabeled PolSAR pixels are fully used to model PolSAR pixels by W-DBN.In addition,the coherency matrix is used directly to represent a PolSAR pixel without any manual feature extraction,which is simple and time saving.Local spatial information,together with the confusion matrix,is used in this work to clean the preliminary classification result obtained by the method based on W-DBN.With help of the prior knowledge of PolSAR data and local spatial information,the proposed method overcomes shortcomings of traditional methods,which are sensitiveto extracted features and slow to execute.The experiment results show that the proposedmethod produces better results and is much faster than traditional methods.2.To improve the efficiency of the classification algorithm,a specific deep model for Pol-SAR image classification is proposed,which is named Wishart DSN(W-DSN).First of all,a fast implementation of Wishart distance is achieved by a special linear transformation,which speeds up the classification of PolSAR image and makes it possible to use this polarimetric information in the following neural network(NN).Then,a single-hidden-layer NN based on the fast Wishart distance is defined for PolSAR image classification,which is named Wishart network(WN)and improves the classification accuracy.Finally,a multi-layer NN is formed by stacking WNs,which is in fact the proposed deep learning architecture W-DSN for PolSAR image classification and improves the classification accuracy further.In addi-tion,the structure of WN can be expanded in a straightforward way by adding hidden units if necessary,as well as the structure of the W-DSN.As a preliminary exploration on for-mulating specific deep learning architecture for PolSAR image classification,the proposed methods may establish a simple but clever connection between PolSAR image interpretation and deep learning.The experiment results tested on real PolSAR image show that the fast implementation of Wishart distance is very efficient(a PolSAR image with 768 000 pix-els can be classified in 0.53 s),and both the single hidden-layer architecture WN and the deep learning architecture W-DSN for PolSAR image classification perform well and work efficiently.3.The previous two works solve two main difficulties in PolSAR image processing,i.e.,data modeling and highly efficient classification.However since it is hard to obtain manual labeled samples,there is another difficulty named small sample problem.To tackle this problem,a novel version of Generative Adversarial Network(GAN)named Task-Oriented GAN is proposed in this work.It mainly consists of three parts,i.e.,Generator(G-Net),Discriminator(D-Net)and TaskNet(T-Net).Both G-Net and T-Net are designed specifically for PolSAR data,where the former generates fake-data and the latter aims to accomplish different tasks,i.e.,classification or clustering.D-Net is similar to that in the general GAN,which distinguishes fake-data from the real one and helps G-Net generate more reliable fake-data.Combined with D-Net,G-Net learns to analyze the property of PolSAR data without a distribution assumption.For classification task,the T-Net is set to be a Classifier.The generated fake-data from G-Net can be used as real-data after training,so the proposed method performs well even if the manual-labeled data is small.The experiment results show that the proposed method performs well in dealing with PolSAR image classification,where small manual-labeled data is needed.4.To detect areas of change in multitemporal PolSAR images,this work presents a novel version of Convolutional Neural Network(CNN),which is named Local Restricted Con-volutional Neural Network(LRCNN).Unlike the general CNN,LRCNN comprises only convolutional layers and a spatial-constraint called Local Restriction(LR)is imposed on the output layer.In the training of LRCNN,the polarimetric property of SAR image is fully used instead of manual-labeled pixels.At first,a similarity measure of polarimetric SAR data is proposed and the Layered Difference Images(LDIs)of polarimetric SAR images are produced accordingly.Next,the LDIs are transformed into Discriminative Enhanced LDIs(DELDIs)to benefit the LRCNN training.With these new difference images,LRCNN is trained by a regression pre-training and a classification fine-tuning.Finally,the change de-tection result showing changed areas is directly generated from the output of the trained LR-CNN.The relation of LRCNN and the traditional way for change detection is also discussed to illustrate the proposed method from an overall point of view.Tested on one simulated dataset and two real datasets,the effectiveness of LRCNN is certified and it outperforms var-ious traditional algorithms.In fact,the experimental results demonstrate that the proposed LRCNN for change detection not only recognizes different types of changed/unchanged da-ta,but also ensures noise insensitivity without losing details in changed areas.5.For the task of change detection in large-scale PolSAR image,a new method based on Looking-Around-and-Into mode is proposed,which is inspired by the way of visual search-ing in an electronic map.In this mode,an Attention Proposal Convolutional Auto-Encode(APCAE)and a Recurrent Convolutional Neural Network(Recurrent CNN)are proposed,to accomplish the Look Around process and the Look Into process respectively.First,the large-scale PolSAR image is zoomed into a small size by down-sampling and the corre-sponding Difference Image(DI)is obtained,then APCAE is used to locate the candidate domains which contains changed area.Second,all candidate domains are sorted in an im-portance descending order so that important domains can be detected in prior.Third,zoom the candidate domains into different scales and the Recurrent CNN is employed to produce multi-scale change detection results.Repeat these steps until all candidate domains are detected.This algorithm works efficiently like this.It looks around firstly to locate the can-didate domain(Look Around)and then analyze them in different scales(Look Into),which is called a Looking-Around-and-Into mode.Experiment results show that it performs very well in the task of change detection in large-scale PolSAR image.