Remote Sensing Imagery Terrain Classification With Discriminative Sparse Coding And Structural Dictionary Learing

Synthetic Aperture Radar(SAR)is a senor that can be equipped on a space-craft or a air-craft which is able to work on all-day and all-whether condition.SAR images are widely employed in the fields of target detection and recognition,terrain map,resource exploration.SAR image land-cover classification is an important and primary task for these applications.So far,there are many algorithms which have been proposed for the SAR image land-cover classification.But most of the traditional methods are based on the low-level feature and the improvement is very limited.This is mainly due to two factors.The one is the SAR image is subjected to the speckle and for another,there does not exist the typical visual elements which represents the unique semantic.Therefore,in the SAR image land-cover classification task,it is very important how to extract a robust and discriminative feature without affection from the speckle.Recently,there have been many algorithms based on the mid-level or high-level information or feature and they are successfully applied in the natural scene classification and annotation tasks.Sparse representation(coding)model has been proven to be one of the effective methods on image analysis and classification.As an effective and robust feature,it is successfully employed in the scene classification and the object recognition for all type of images and the good performance is obtained.Moreover,the contextual information and priors are considered in the classification algorithms,which play a key role in the improvement of the classification accuracy and the reduction of the algorithm complexity.Therefore the classification based on the contextual information and priors becomes a hot spot in the field of the image classification.In this dissertation,a deep and systematic research is done based on sparse representation model,which focuses on the key issues on the remote sensing image,especially SAR image,land-cover classification.The main contributions can be listed as follows.(1)It has been demonstrated that the collaborative representation-based classification(CRC)is less complex and more effective than the sparse representation-based classification(SRC).However in the CRC model,all of atoms are employed to encode a sample and it is very low efficient.To solve this issue,an adaptive dictionary is proposed.Compared with the CRC,only a small part of training samples are employed to construct a dictionary,instead of all the samples.This locality implies the sparsity and reduces the computational complexity.On the other hand,there exist large numbers of structures in hyperspectral image and it is not independent among pixels in a local region.It is assumed that given a dictionary,their coef-ficients have the similar non-zero entries.Based on this assumption,two joint collaborative representation models based on the local spatial context and the non-local spatial context are proposed for hyperspectral image classification.It is shown in the experiment results that the proposed algorithms are faster than the CRC and the SRC with higher classification accuracy.(2)The traditional sparse representation model can approximate the data with a few of atoms and the good performances are obtained in the image restoration.In the SAR image land-cover classification,a robust,sparse and discriminative feature is expected.For this,a dis-criminative sparse coding and a dictionary learning algorithm are proposed.In here,the classifier is considered as a prior which the coefficients satisfy the sparsity prior and the max-margin criterion.The experiment results show that the proposed algorithm performs better than the traditional one in the SAR image land-cover classification task.(3)With the increase of the resolution,there exist much complex texture structures in the SAR image.The single layer sparse coding model only encodes the each image pixel or image patch independently,ignoring the relationship among the patches and the ability to representation is very limited.In order to improve the performance of SAR image land-cover classification,not only the priors on the classifier are fully employed to make the coefficients satisfy the max-margin criterion,but also a new hierarchical sparse coding model is proposed to encode the spatial contextual information in the SAR image.The experiment results show that the proposed algorithm is effective.(4)It is a key issue how to a discriminative dictionary,which provides the sparse discrim-inative features for image classification.For this,a novel dictionary learning method is proposed based on the inter-class difference,where the features are trained by a good and discriminative dictionary.In addition,it is necessary to impose several priors and constraints to reduce the number of required training samples.To achieve this,the inter-class difference and structural priors are considered as constraints and imposed on the dictionary learning.Furthermore,a linear classifier is employed to guide sparse feature extraction and the dic-tionary learning and it is trained to classify the test sample efficiently.It is demonstrated by the experiment that the proposed algorithm performs better than other algorithms.(5)Features are of great importance for SAR imagery terrain classification,but low-level features usually readily suffer from the speckle noise and they are incapable or inaccurate to capture some complex and irregular texture structure.Therefore,a novel feature learning framework is proposed to address this problem,in which some mid-level and high-level features are simultaneously learned by exploiting the spatial context constraints and sparse priors.More specifically,the mid-level features served as the intermediates are extracted from several initialized low-level features by considering the spatial constraints to reduce the influence of the speckle noise.Then,more abstract and discriminative high-level features are learned with an effective dictionary learning algorithm so as to represent the complex structures in SAR imagery.The proposed framework is evaluated on both artificial synthesis and real SAR imagery.It is shown from quantitative evaluations and visual results that the proposed algorithm performs better than other compared algorithms.It also can be demonstrated that both the feature selection and the dictionary learning jointly contribute to the improvement of performance.(6)Remote sensing imagery contains complex texture structures and the dependencies among patches in a local region.To exploit this,structured sparsity priors are imposed on the sparse model.In addition,to learn a completed dictionary for remote sensing,a large number of training set is necessary,which will reduce the efficiency of dictionary learning.Further-more,there exists the inherent structures in the dictionary.Inspired by the idea of online learning,an online structured dictionary learning algorithm is proposed.When the the train-ing set is change,the dictionary is updated in real time,instead of re-training.In addition,the inherent structure of dictionary is considered and the dictionary is updated group-by-group.The experiment results demonstrates that the proposed algorithm performs better than other algorithms and is more efficient than the off-line dictionary learning.