A Study And Implementation Of SVM Algorithm Applied To Pattern Classification Of TM Multispectral Image

Support Vector Machine (SVM) algorithm is a kind of pattern classificationalgorithm which bases on the theory of Statistical Learning Theory (SLT). Because ofits features of better computational efficiency, robustness and statistical stability, SVMgets rapid development in these recent years. It's widely used in many fields whichbelong to the application of pattern classification.The classification for land cover of TM multispectral Imaging is all along adifficult problem. Because there are all kinds of land covers and different kinds ofland covers will change along with the time's process and space's distributioncontinuously, land covers' reflection, multispectral image data will alter too. Inaddition, mutispectral data itself have the feature of "the same substance, the differentcomposition;the same composition, the different substance" Data will also be noisedduring collection process. All disadvantages we talk above make some traditionalmethod of classification perform poorly in aspect of computational complexity,classification accuracy, generalization performance.This thesis takes the area nearby HuaiRou reservoir of BeiJing as researchmaterial, builds the SVM algorithm framework of C-SVC type for solving landcover's classification problem of TM multispectral image which is composed of sixbands by BSQ data format. The algorithm uses three kind of kernel function: RadialBasis kernel, Polynomial kernel and Sigmoid kernel to classify the TM multispectralimage data. These classifications are all multi-class classification. The algorithm usesSequential Minimal Optimization (SMO) algorithm to solve the Quadratic Problem(QP) with restricted condition. This thesis analyses the theory basis of base SVMalgorithm, SMO algorithm and one against one multi-class classification method andimplements concretely these algorithms and classification software package with C++programming language. Otherwise, this thesis also implements another two kinds ofclassification algorithm: Minimum Distance and Maximum Likelihood Estimationalgorithms, aimed to compare SVM algorithm with them. This thesis compares themand proof that SVM algorithm has a better representation and predominance ataccuracy rate and generalization performance. The SVM algorithm can be used inland cover's classification of TM multispectral image practically and efficiently.