Unsupervised band selection for hyperspectral image classification using the Wasserstein metric-based configuration entropy

Band selection relies on the quantification of band information. Conventional measurements such as Shannon entropy only consider the composition information (e.g., types and ratios of pixels) but ignore the configuration information (e.g., the spatial distribution of pixels). The latter could be quantified by Boltzmann entropy. Among all the metrics of Boltzmann entropy, the Wasserstein metric-based configuration entropy (Wasserstein entropy for short) removes the redundant information of the continuous pixels. However, it is limited to 4-neighborhood. This article improves it to 8-neighborhood. Taking the hyperspectral images of Indian Pines and Italian Pavia University as examples, we used the difference of Wasserstein entropy to measure band correlation and then employed the unsupervised sub-optimal searching algorithm to determine the optimal band combination. We used the support vector machine classifier for image classification. Finally, we compared the accuracy of image classification based on the difference of Wasserstein entropy, mutual information, four types of normalized mutual information, and two variants of relative entropy. Results show that both the 4-neighborhood and 8-neighborhood Wasserstein entropy can be used for band selection of hyperspectral images, especially when few bands are considered. The 8-neighborhood Wasserstein entropy works better than 4-neighborhood. © 2021, Surveying and Mapping Press. All right reserved.