Sparse multinomial logistic regression (SMLR) is an important supervised classification method for hyperspectral images (HSI). However

because the traditional SMLR based pixel-wise classifiers only use the spectral signatures

the good robustness and high classification accuracy are hardly achieved with a small number of samples without considering the spatial information of HSI. By using the kernel tricks

the kernel sparse multinomial logistic regression (KSMLR) method can partly overcome this limitation

however the resulted misclassification errors are still expected to be further reduced. According to the statistical analysis of classification errors resulted in KSMLR

we propose a novel two stage framework which combines KSMLR and error rejection for HSI classification. The proposed model

named KSMRL-TVL1

adopts the L1 norm to measure the heavy-tailed property of the classification errors so as to build the data fidelity term

and uses the total variation (TV) regularization term to measure the local spatial smoothness of the hidden probability field. The experiments on Indian Pines dataset and University of Pavia dataset show that the proposed method can better improve the robustness and classification accuracy.