基于多核支持向量机的高光谱影像非线性混合像元分解

Abstract

As the mixed pixel decomposition based on linear spectrum models has lower decomposition accuracy and the nonlinear spectrum model is difficult to be established,a nonlinear mixed pixel decomposition method for the hyperspectral imagery was proposed based on the posterior probability of Multiple Kernel Support Vector Machine(MKSVM).On the basis of the SVM,the multiple kernel function formed by linear weighted combination was taken to replace the single kernel and the simple multiple kernel learning was used to solve the weights iteratively to achieve the classification.Then,the output values of the classifier were converted to pairwise coupling probabilities by thesigmoid function and then to the multi-class posterior probability.Finally,the hyperspectral imagery decomposition was achieved through the posterior probability.The results from experiments of two push-broom Hyperspectral Imagers(PHIs)show that the classification accuracies of hyperspectral imagery nonlinear mixed pixel decomposition based on MKSVM reach 95.62% and 91.51%,respectively,the Root Mean Square Errors(RMSEs)are reduced to 11.15%and 7.55%,and both are less than 15%.In conclusion,the influence of mixed pixel on hyperspectral imagery classification is eliminated,and the classification accuracy is increased.