High-resolution SAR Image Classification Based On Discriminant Feature Learning

The synthetic aperture radar(SAR)is an active detection system with long distance,allweather and timeless imaging capacity.It has been widely utilized in a variety of civil and military areas.With the development of SAR imaging techniques and the increasing of spatial resolution,how to obtain an effective representation of SAR images plays a pivotal role in many understanding and interpretation applications.After analyzed the recent development on feature extraction techniques,a deeper study on feature learning of high-resolution SAR images was made in this thesis by combing with the recent progress on machine learning and computer vision.Meanwhile,all of these feature extraction and learning approaches which are proposed in this thesis are evaluated via several land-cover classification and scene category tasks.Specifically,the major topics of this thesis are listed as follows:1.To characterize the content of high-resolution SAR image,a semantic-contextual model(SCM)is developed to learning high-level features by integrating both the semantic and contextual information.Specifically,the semantic descriptor is utilized to capture the high-level and abstract information by modeling the intrinsic relation between terrain categories and low-level features.At the same time,the contextual information is utilized to model the co-occurrence relations with a superpixel segmentation on the specified high-resolution SAR image.The terrain labels of high-resolution SAR image can be obtained via a maximizing a posterior(MAP)procedure after utilized the superpixels as the basic operational units instead of simple pixels.It can be observed that the high-level features learned by SCM can effectively reduce the semantic gap between features and different terrain categories for land-cover classification.2.A novel discriminant feature learning method is proposed to extract high-level features of high-resolution SAR images by introducing the locality-constraint,which is widely utilized in the literature of computer vision.Firstly,a weighted discriminative filter(WDF)bank is learned from the large amount of labeled high-resolution SAR image patches to produce the low-level features.The content of high-resolution SAR image can be characterized discriminatively by these low-level features which are produced by WDF.Then,a locality structural constraint is introduced to formulate the high-level features in both the feature encoding and spatial pooling procedures.The superpixels are also used to be basic units instead of pixels,in which each superpixel is characterize by a hyper-feature based on some domain patterns which are learned from all of the high-level features of pixels.All of these hyper-features are feed to a classifier which is trained previously with some labeled superpixels to accomplish the final terrain classification task.3.With the increased amount of available high-resolution SAR images,it is a difficult task to capture enough discriminative information from the limited number of labeled high-resolution SAR image samples.A novel feature learning approach is proposed to address this difficulty,in which each high-resolution SAR image patch is characterized by a discriminant feature generated in a sparse ensemble based semi-supervised manner.In particular,a non-negative sparse coding procedure is applied on the given high-resolution SAR image patch set to generate the feature descriptors firstly.The set is combined with a limited number of labeled high-resolution SAR image patches and abundant number of unlabeled ones.Then,a semi-supervised sampling approach is proposed to construct a set of weak learners,in which each one is modeled by a logistic regression procedure.The discriminant information can be introduced by projecting high-resolution SAR image patch on each weak learner.Finally,the features of high-resolution SAR image patches is produced by a sparse ensemble procedure that can reduce the redundancy of multiple weak learners.4.Inspired by the prototype theory of cognitive science,the discriminative information of a high-resolution SAR image patch can be exploited by comparing to some prototypes that are learned from the training high-resolution SAR image patches.In this thesis,a novel feature learning approach that is called discriminant deep belief network(Dis DBN)is proposed to learning high-level features for high-resolution SAR image classification,in which the discriminant features are learned by combining ensemble learning with a deep belief network in an unsupervised manner.Firstly,some subsets of high-resolution SAR image patches are selected and marked with pseudolabels to train weak classifiers.Secondly,the specific high-resolution SAR image patch is characterized by a set of projection vectors that are obtained by projecting the high-resolution SAR image patch onto each weak decision space spanned by each weak classifier.Finally,the discriminant features are generated by feeding the projection vectors to a DBN for high-resolution SAR image classification.The experiment on land-cover classification illustrates that the proposed approach can fully exploit complementary information between all of the weak classifiers.5.Inspired by deep learning techniques and the probability mixture model,a generalized Gamma distribution based deep belief network(g?-DBN)is proposed in this thesis for high-resolution SAR image statistical modeling.Specifically,a generalized GammaBernoulli restricted Boltzmann machine(g?B-RBM)is proposed to model the highorder statistical relations of high-resolution SAR image patches.By stacking the g?BRBM and several standard binary RBMs in a hierarchy manner,a deeper structure,i.e.g?-DBN,is formulated to learn high-level features of SAR images.And lastly,a discriminative neural network is constructed by adding an additional predict layer for terrain label to the deep structure.With the learned high-level features,the proposed approach could capture meaningful structures which is effectively to characterize the content of high-resolution SAR images for terrain classification.