A novel active learning approach for the classification of hyperspectral imagery using quasi-Newton multinomial logistic regression