Research On Hyperspectral Remote Sensing Image Classification Based On LS-SVM

Abstract:In comparison with traditional multispectral remote sensing, hyperspectral remote sensing, characterized by combining the imaging and spectroscopy, contains more abundant spectral information. However, because of its numerous bands and narrower bandwidth, each band is highly correlated. From a certain perspective, the problem of data redundancy is caused by excessive dense and large amounts of bands. Meanwhile, it brings challenges and opportunities for hyperspectral remote sensing information acquisition and precise classification.Numerous domestic and international scholars have done a lot of researches in the precise classification field of hyperspectral remote sensing image. Especially, some of them have achieved satisfactory results by applying the support vector machine (SVM) to the hyperspectral remote sensing classification. The least squares support vector machine(LS-SVM) is an improvement of SVM algorithm. In this paper, the LS-SVM is applied to hyperspectral remote sensing classification. Related research and main works are as follows:(1) According to the features of LS-SVM and the drawbacks of the traditional multi-class classification method, this paper proposes the skew binary tree multi-class least squares support vector machines classifier. Then apply it to hyperspectral remote sensing image classification. The results show that the proposed method is effective and feasible, the classification accuracy and Kappa coefficient have been improved obviously.(2) According to the characteristics of hyperspectral remote sensing which owns many channels and large amounts of data, this paper proposes a feature extraction algorithm which is based on the subspace of bands (SOB) of ICA (independent component analysis). Then apply it to hyperspectral classification. The results of this experiment demonstrate that the ICA algorithm’s superiority of hyperspectral feature extraction. At the same time, the influence of image noise on classification accuracy is restrained by this algorithm, classification accuracy significantly excelled other feature extraction algorithm.(3) Due to the numerous bands of hyperspectral remote sensing data, not all the bands has a significant contribution to the hyperspectral remote sensing classification. According to the strong optimization of particle swarm optimization (PSO) algorithm, this paper uses PSO algorithm to optimize the parameters of Gaussian radial basis kernel function, uses binary PSO to select hyperspectral remote sensing data’s bands. Then apply this model to hyperspectral remote sensing classification. The results show that the hyperspectral remote sensing’s overall classification accuracy and Kappa coefficient can be improved.