Ground truth labeling and samples selection for Hyperspectral Image Classification

Hyperspectral Image Classification (HSIC) has gained much attention for several real-world problems in which an appropriate number of labeled training samples are necessary to obtain a good performance which highly depends on the quality (spectral & spatial heterogeneity) of the training samples. For the above reason, this study first identifies a relationship between the misclassification rate and uncertainty of a classifier on a group of very few training samples. We then experimentally prove that uncertain samples significantly help to enhance generalization performance. Finally, we propose a divide and conquer strategy to split the samples based on low, mid, and high uncertain samples to strengthen the classification performance. The proposed pipeline is theoretically explained and experimentally analyzed on benchmark real Hyperspectral Datasets.