Nearest regularized subspace based hyperspectral image classification with adaptive Markov Random Field and high confidence index rule

Spectral-spatial classification is very important in the field of remotely sensed hyperspectral imaging. The Markov Random Field (MRF) is usually used to provide the spatial constraint. However, the traditional construction of MRF uses a neighborhood and gives the same weight for all the pixels in the same neighborhood. Without the structure information in HSI, the performance of classification is not good enough. In this paper, we propose two improvements on MRF: the adaptive MRF is constructed according to spectral similarity of the spatial patch to reflect the structure information in HSI; a regularized method with a high confidence index (HCI) rule is presented to determine the refined probabilities of each pixel that promotes spatial continuity within each class. The effect of HCI rule is utilized to treat the probabilities discriminately, where the vectors with HCI are supposed to be well labeled. And the original probabilities of the pixels belong to each class are obtained by the spectral information using nearest regularized subspace (NRS) method in the first step. Experimental results on real hyperspectral data set demonstrate that the proposed method outperforms many existing methods in terms of the overall accuracy, average accuracy and kappa statistic.