Application of Adaptive Kernel Matching Pursuit to Estimate Mixture Pixel Proportion

An adaptive kernel matching pursuit (AKMP) algorithm to estimate mixture pixel proportion of remotely sensed image has been proposed. The AKMP algorithm applies greedy sparse approximation algorithm to the feature space induced by a nonlinear kernel function, and can therefore be able to capture nonlinear effects of image and performed better than conventional linear approaches. Moreover, it has the capability of adaptive selection of the kernel parameter before starting the greedy approximating procedure to avoid complex procedures of kernel function parameter selection. Experiments with ETM+ associated with IKONOS image have been carried out, and the result demonstrates that the proposed method can provide accurate proportion estimation. Comparisons with support vector regression (SVR) and linear mixture model (LMM) have also been done, and the experiments show that the proposed method outperform SVR and LMM in terms of RMSE.