Font Size: a A A

The Classification Of Hyperspectral Remote Sensing Image Based On Support Vector Machine

Posted on:2009-05-31Degree:MasterType:Thesis
Country:ChinaCandidate:L QiaoFull Text:PDF
GTID:2178360272480153Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
Hyper-spectral remote sensing image classification is a key technology in remote sensing applications. Rapid and high-accuracy hyper-spectral remote sensing image classification algorithm is the precondition of kinds of practical applications. Traditional pattern classification methods are based on the principle of experiential risk minimization, and they can achieve the best result, only when the number of samples approaches infinity. Unfortunately, in hyper-spectral image classification, training samples are usually limited. Taking into account the good generalization of support vector machines in small samples, nonlinearity and high dimension space, and according to features of hyper-spectra remote sensing image, this dissertation deeply studies support vector machine(SVM) and their application for hyper-spectral remote sensing image classification. The main contributions of this thesis are given as follows:First, kernel functions are the core of the support vector machine. They are divided into local kernel and the whole kernel two categories, different kernel functions can produce different classification affects. Common kernel function RBF only has a local characteristic, with the measurement of the difference in brightness according to the Euclidean distance between pixels of two types. This paper proposes the introduction of mixed spectral differences of pixel as a new measurement which also has whole characteristic. The new modular kernel function not only has a good learning ability, but also applies the full information which the high-dimensional spectra remote sensing image provided.Second, least squares support vector machine (LS-SVM) is the expansion of the standard support vector machines, in order to facilitate the mathematical solution. Chancing the original inequality constraints quadratic optimization problems into the linear equation restraint system, this processing greatly simplifies the calculation. However, this expansion has led to the loss of stability and the sparse ability. On the based of fully understanding the layered distribution characteristic of least squares support vector, this paper proposes a bilateral weighted least squares support vector machine algorithm specially treating such problem. This algorithm increases the threshold of non-support vectors, which are far away from the classification plane, reduces the weight of the vector, and thus weakens the impact of non-support vectors on the structure of optimal classification plane. The bilateral weighted method also compensates the influence of differences between the two categories, as part of the relaxation factor tends to zero, improves the sparse ability of the value matrix, and speeds up the classification rate.Last, this paper deeply analyzes the six existing typical systems of support vector machines classification construction, and discusses four aspects, e.g. the training speed, classification rate, selection and promotion capacity, and classification accuracy in detail. Especially for the high-spectral data structure, this paper proposes the introduction of super-ball structure of the support vector machine, to build the binary tree structure, based on separated distances from each category of all the samples. This algorithm balances the relationship among the distance between the two categories, the distance in individual category and the optimal classification plane, and it also decreases the accumulation of the top-down errors.
Keywords/Search Tags:hyper-spectral remote sensing image, support vector machine, kernel function, multi-classes classification
PDF Full Text Request
Related items