An novel active learning strategy for hyperspectral image classification

Manual training data selection for land cover classification is time consuming and expensive. Furthermore, it ignores the quality of the training set as well, which leads to worse classification results. Active learning methods have been proposed to solve this problem. However, most of the exiting methods ignore the spatial distribution of the hyperspectral images during the selection procedure. In this paper, the informative samples not only based on the spectral information but also on spatial distribution will be selected into training set. Furthermore, in order to reduce the workload of the manually labeling, the part of informative samples can be directly assigned labels by the proposed strategy. At last, we present experimental results that establish the superior performance of our proposed approach.