PolSAR Image Classification With Sparse Representation And Deep Learning

Polarimetric synthetic Aperture Radar(PolSAR)is an active sensor which can obtain high quality remote sensing images from airborne or spaceborne systems without the restriction of illumination and other external conditions.This system is widely used in military and civil fields with the ability to detect targets by radar echo.Therefore,PolSAR image processing technology has received more and more attention,among which PolSAR image classification is the key technology of PolSAR image processing.After decades of development,researchers have proposed a variety of methods for PolSAR image classification.Especially,with the rapid development of sparse representation and deep learning,the related methods are also applied to PolSAR image classification.However,sparse representation and deep learning still have a lot of room for improvement in the application of PolSAR image classification.Based on the above background,in order to further improve the performance of sparse representation and deep learning in PolSAR image classification,this paper proposes a series of related methods,which are summarized as follows:1.A multilayer dictionary learning architecture is proposed in this paper.The recently proposed projective dictionary pair learning(DPL)model can learn a projective dictionary pair composed of analysis dictionary and synthesis dictionary at the same time.When the DPL model is optimized,its discriminative performance and representation performance will be balanced.Moreover,its model structure is simple,so it can achieve good classification results while taking into account time consumption.Moreover,in recent years,deep learning has been widely used in various fields,one of the key reasons for its excellent performance is its deep architecture.Inspired by this reason,this paper modifies the DPL model appropriately and extends it to multilayer architecture.A multilayer projective dictionary pair learning(MDPL)model is proposed to extract PolSAR data features.Because of the multilayer structure of MDPL and the excellent feature extraction performance of its basic model DPL,the feature extracted from MDPL has high abstractness,which is helpful to obtain good classification results.Then,the feature extracted from MDPL is input into Softmax classifier,which is simple in structure and good in classification performance.2.A new method for fitting different features of PolSAR images is proposed.In most of the PolSAR image classification methods,more than one feature is extracted from PolSAR image,such as polarimetric coherence matrix,polarimetric covariance matrix and polarimetric target decomposition.However,in the current methods,the common processing method is to directly pull all features into a feature vector,which to some extent will separate the relationship between the features.The semi-coupled dictionary learning(SCDL)model proposed in recent years can be used to solve this problem.The SCDL model proposed to solve the super-resolution problem.The feature relationship between high-resolution image and low-resolution image can be found,but its optimization method is more complicated,and DPL model can be used to solve this problem.Based on the SCDL and DPL models,a semicoupled projective dictionary pair learning(SDPL)model is proposed.The SDPL model can reveal the relationship between different features.Its optimization method is simple and the classification performance corresponding to SDPL is good.In addition,sparse auto-encoder(SAE)is used to extract the features,and SAE can adaptively extract the nonlinear relations among the elements in the input data.Therefore,a good classification result of PolSAR image can be obtained.3.The nonlinear transformation is introduced to sparse representation.In recent years,both sparse representation and deep learning have been successfully applied to PolSAR image classification,but the results obtained by sparse representation are often not as good as those obtained by deep learning.In this paper,it is considered that one of the main reasons for the excellent performance of deep learning is its nonlinear transformation.However,sparse representation is often a linear operation in the original data space.Therefore,on the basis of DPL model,a non-linear projective dictionary pair learning(NDPL)model is proposed.Similar to the DPL model,the NDPL model also needs to learn the dictionary pair composed of analysis dictionary and synthesis dictionary at the same time.The analysis dictionary plays the role of coding and synthesis dictionary plays the role of decoding,but the encoding process and decoding process introduce nonlinear transformation.Therefore,its ability to extract features is significantly enhanced.Moreover,the NDPL model also inherits the simple framework of the DPL model,and its optimization method is relatively simple and it does not require too many labeled training samples.Therefore,the NDPL model can use a small number of labeled training samples to obtain excellent classification results,and its time consumption is also less.4.The fully convolutional network(FCN)is successfully applied to PolSAR image classification.Deep learning is widely used in PolSAR image classification.However,the common deep learning models have some problems in PolSAR image classification.For example,the deep belief network(DBN)and SAE are limited by their model architectures and cannot utilize the spatial information of the image effectively,while the convolutional neural network(CNN)must use the neighborhood of the pixel to get the classification result of a single pixel,which leads to repeated calculation and repeated memory consumption.The recently proposed FCN method can be used to deal with the dense prediction problem,and the PolSAR image classification is exactly the dense prediction problem.Therefore,FCN is the most suitable method for PolSAR image classification.However,due to the different size of each PolSAR image,the FCN model cannot be directly applied to PolSAR image classification.In order to solve this problem,a sliding window fully convolutional network(SFCN)is proposed in this paper.However,with the increase of image size,the computation burden and memory consumption of SFCN will also be greatly increased.This problem can be solved by sparse coding.By sparse coding,the image can be down-sampled to a smaller size with little information loss.The down-sampled image can be classified by SFCN,and the classification result can be up-sampled to the original image size.In this way,competitive results can be achieved.5.Deep reconstruction-classification network(DRCN)and adversarial training are introduced to PolSAR image classification.As we all know,image marking is a very human and material consumption,so it is very meaningful if very few labeled training samples can be used to get good classification results.The recently proposed DRCN method can be used to solve this problem,which can deal with two tasks: supervised image classification and unsupervised image reconstruction,and the two tasks share feature coding.The shared feature coding can reconstruct all the pixels of the image and classify some labeled pixels.The shared feature coding can also correctly classify the remaining pixels in terms of the distribution of the data.Therefore,unsupervised image reconstruction is a useful auxiliary task for supervised image classification.The higher-order inconsistencies between the true image and reconstructed image can be detected and revised by adversarial training,so that can help DRCN get good classification results.Based on SFCN,adversarial reconstructionclassification network(ARCN)is proposed by introducing DRCN and adversarial training.ARCN can be trained from pixel to pixel,end to end,and it can obtain very competitive image classification results with very few labeled training samples.