Hyperspectral Image Classification Using Spectral-Spatial Features With Informative Samples

This paper proposes a new active-learning approach for multi-feature hyperspectral image classification. First, the extended multi-attribute morphological profiles (EMAPs) are introduced as features into the classifier of the multinomial logistic regression (MLR). Second, discontinuity preserving relaxation (DPR) is used to improve the precision of the labels predicted using the MLR classifier. Finally, in order to improve the efficiency of the training process using the EMAP-MLR-DPR classifier, we proposed selecting the informative training samples based on both the uncertainty and representativeness of the data. The breaking ties scheme is taken as the metric uncertainty of the samples, and the mean shift cluster is used to denote the representativeness of the unlabeled samples. The proposed method reasonably combines the spatial information and spectral information of hyperspectral data and effectively selects the key training samples with the most information. The effectiveness of the method is confirmed in the experiments on multiple hyperspectral data sets.