Band Selection for Hyperspectral Image Classification

Abstract

This research presents a band selection technique for hyperspectral image classification. This technique utilizes mutual information between bands and the ground truth reference map, and correlation coefficients between bands. Band selection in this research comprises two major steps. The first step is to define a set of informative bands, called a candidate set. This set can be identified via mutual information between each band and the ground truth reference map. The second step is to reduce the redundancy between bands using correlation coefficients between bands. Selected bands will be kept in a band selection set. This set comprises the first band presenting in the candidate set. The second band is chosen from the next band presenting in the candidate set that offers a correlation coefficient below the preset threshold. The next band selection can be done in the same criteria until the number of bands is met the requirement. The results show that the proposed technique is useful for band selection. When only 50% of all the bands are utilized for classification, the overall accuracy and average accuracy are greater than those from using all the bands, and kappa coefficients are equal, which are 81.08%, 82.93%, and 0.78, respectively. Moreover, if the number of bands used for classification is less than 50% of all the bands, the classification accuracy is not lower too much when compared with the decrease of the number of bands, such as when only 10% of all the bands are used, the overall accuracy, the average accuracy, and kappa coefficient are dropped to 69.39%, 66.32%, and 0.65, respectively.