Sparse Representation And Dictionary Learning For Hyperspectral Remote Sensing Image Classification

Hyperspectral remote sensing technology is an important part of geoscience and remote sensing community.Hyperspectral remote sensing,which span the visible to infrared spectrum with hundreds of contiguous and spectral bands,can simultaneously obtain spatial information and fine spectral information.Owing to their subtle discriminative spectral characteristics,hyperspectral image support improved interpretation capabilities,and they can be used for various fields,such as ecological science,hydrological science,geological science,precision agriculture,and military applications.How to make full use of the rich spatial and spectral information in hyperspectral image and perform detailed interpretation of ground object is always a research hotspot and frontier problem in hyperspectral remote sensing field.Currently,as a new data mining technology,sparse representation has been widely applied to various kinds of hyperspectral image classification owe to its outstanding performance.However,single-feature representation-based methods can only reveal the sparse characteristics from one perspective,and can not fully exploit the sparse characteristics of hyperspectral imagery.Also,the spectral-spatial feature of hyperspectral imagery generally contains rich nonlinear structural information,which helps to improve the separability of ground objects.The dictionary directly constructed by stacking some labeled pixels is generally incomplete.Considering the characteristics of hyperspectral remote sensing image(HSI),this dissertation systematically study HSI sparse-based classification according to different levels of understanding of sparse characteristic from the perspective of feature diversity,model diversity and dictionary completeness,and proposed a sparse representation and dictionary learning classification framework for HSI,including multiple feature kernel sparse representation-based method,multiple kernel adaptive collaborative representation-based method,and spatially weighted dictionary learning.Five HSI data sets are adopted for validating the performance of the proposed classification framework.The main research conclusions and innovations are given as follows:1)The sparse discriminant characteristic from different spectral-spatial feature of hyperspectral image were revealed,a feature diversity sparse representation-based method was proposed,and sparse representation via multi-source heterogeneous feature fusion was realized.Different types of spectral-spatial features(such as,filter,texture,and shape)can describe the characteristics of HSI from different perspectives,and feature combination is an important way to improve the performance of HSI classification tasks.Taking each spectral-spatial feature as one task,multiple feature learning has been widely integrated into sparse multi-task learning model to obtain a multi-feature representation coefficients with stronger discriminant power.Generally,linear sparse representation-based model cannot handle HSI with highly nonlinear distribution.By adopting nonlinear mapping,the original feature space of HSI can be projected into a high or even infinite dimensional space.Hence,by integrating kernel principal component analysis(KPCA)into multi-feature-based sparse multi-task learning model,different kinds of spectral-spatial features can be effectively combined together in low-dimensional kernel space,and a multiple feature kernel sparse representation-based method was proposed to achieve feature diversity sparse representation-based learning.Experimental results demonstate that the proposed method can efficiently capture the sparse characteristics of different kinds of spectral-spatial feature of HSI,and the sparse optimization efficiency was improved by adding a dimensionality reduction stage in kernel space.2)By adopting multiple kernel learning technique to capture the nonlinear structural information from HSI,a model diversity sparse representation-based method was proposed,which solved the problem of multiple heterogeneous kernel model fusion.By embedding multiple kernel learning model into collaborative representation-based model,a multiple kernel adaptive collaborative representation-based method is proposed to capture different kinds of nonlinear structural information from different spectral-spatial feature.Multiple kernel patterns,e.g.,Naive,Multi-Metric,and Multi-Scale are adopted for the optimal set of basic kernels,which are helpful to capture useful information from different pixel distributions,kernel metric spaces and kernel scales.Moreover,by considering the different contribution of dictionary atoms,the adaptive representation strategy is applied to the multiple kernel learning framework via a dissimilarity-weighted regularizer to obtain a more robust representation of test pixels in fused kernel space.Experimental results confirm that the proposed method can efficiently capture the rich nonlinear structural information from HSI,outperform the other state-of-the-art classifiers.3)By adopting a joint sparse representation model with spatially adaptive strategy,a semantic sparse dictionary learning method was proposed.Based on typical supervised classifier,sparse representation-based classification with discriminant sparse coding feature was achieved for HSI.Based on the learned coding coefficients over fixed dictionary from the labeled pixels,a residual-based criterion is generally adopted to determine the class label of each pixel for representation-based methods.The dictionary that is constructed this way generally leads to the inefficient coding coefficient for the whole HSI when a large-sized HSI are available.Moreover,pixels are spatially close to each other generally have similar spectral characteristics,spatial information at neighboring locations is beneficial for HSI classification.To cope with the above limitations,we propose a spatially weighted dictionary learning method,which aims at learning a compact dictionary by obtaining the adaptive joint sparse representation of each spatial contextual patch over the whole HSI.By adopting the online learning method to learn the optimal dictionary from the whole HSI,the learned dictionary is generally more compact and can be used to learn the discriminative sparse coding feature for the classification task.Finally,different from the typical residual-based methods,a standard linear support vector machine(SVM)classifier is applied to the coding coefficients over the well-learned dictionary,and the corresponding classification map is then obtained.Experimental results confirms the proposed method outperforms the state-of-the-art competitors.Compared with feature diversity and model diversity methods,semantic sparse dictionary learning method can simultaneously consider classification accuracy and visual effect.The overall classification accuracy reaches to 96.26%for the Pavia University data set with 60 labeled pixels per class,which meets the application requirement for hyperspectral remote sensing interpretation.This dissertation aims at coping with the comprehension and expression problem of multi-level sparse characteristic by the hierarchical structure from aspects of attribute,model and dictionary update.The relevant theoretical and technical issues of sparse representation and dictionary learning methods are discussed,which provide a systematic solution for HSI sparse-based classification.