Research On Hyperspectral Image Classification Algorithm Based On Fused Multi-Classifers

Hyperspectral remote sensing image(HSI) classification is based on machine learning methods or other learning methods. In HSI classification computer is used to automatically distinguish tiny differences of target object in spectrum that traditional image could not handle.HSI Classification is practical and significant that many researcher has paid attention to this subject. This paper is based on the data feature and data charateristics of HSI figured out that traditional classification methods have weakness in accuracy, applicability and neglecting space information then research is done on the following three main aspects.Firstly, a survey of sparse representations classification(SRC), collaborative classification(CRC), support vector machines(SVM), nearest regularized subspace(NRS) basic classifiers algorithms was done and the mathematical theories and algorithms' advantages and disadvantages of these classifiers were confirmed. A theory study on possibility and feasibility of classifiers fusion in HSI classification was done. By analysed result, this paper considers to use effective basic cascade and parallel structure to fuse the multi-classifiers. The cluster reduction, redefintion and residuals fusion methods were used in classifier fusion.Secondly, SRC and NRS have limited in using space-related information and low accuracy, so an improvement was made, by using spatial joint representation SRC and NRS are developd to spatial joint sparse representation classification(JSRC) and spatial joint nearest regularized subspace(JNRS) so that accuracy can be upgraded. Facing the fact that classical method have adaptive problem, this paper proposes a novel hybrid approach combining JSRC and JNRS as the first stage and SVM as the second stage for HSI classification. The experiment results show that compared with single-stage classifier, the proposed two-state classifier not only inherit the advantages of both of the classical classifiers, but also improved the classification accuracy. It has better classification performance and the same level of time consume.Thirdly, the L1 norm is used by SRC which has shortage in selected atoms that too sparse to utilize the within-class information. However the L2 norm used by CRC also sufferd the shortage in selected atom that too collaborated to figure out the main parameter. This paper proposed to use a parallel structure fusion classifer that combine CRC and SRC based on residual fusion call residual Fusion Representation Classification(FRC). FRC use a balancing factor to adjust and achieve fusion. Experimental results of FRC fusion of classifiers have advantage in two ways. It achieved higher rates of correct classification, and maintained the same level of time comsume.Finnally, figured out their drawback in using space information and defect in building training dictionary, this paper proposed two improved algorithms.1. An spatial Joint residual Fusion Representation Classification(JFRC) is proposed to upgrade the utilization of homogeneous areas that enhanced the classification effect. Experiment result shows that JFRC gets satisfied accuracy.2. Non-local Dictionary Residual Fusion Representation Classification(NLD-FRC) uses the self-similarity oftesting pixel data and training dictionary. Results show that, NLD-FRC not only improved the image classification accuracy but also reduced training dictionary size which reduce the computational complexity of the algorithm. Compared with single-stage algorithm, the proposed methods gain better performance.