Channel Selection by GA for Substance Discrimination Using Hyperspectral Images and Neural Network

Abstract

In this research, we aim to achieve better performance of substance discrimination using hyperspectral images than our latest work with PLS regression analysis. The PLS regression analysis necessitates training of classifiers whose total number is the same to the number of classes. Since training many classifiers is time-consuming, a neural network is applied to solve this problem. Moreover, genetic algorithm (GA) is introduced to obtain valid channels for the substance discrimination to avoid input of all the channels. Since this approach can optimize the number of valid channels and theirs types simultaneously, the proposed method is effective. In experiments, the proposed method and the previous method were compared with two image datasets. In the first dataset, the best intersection over union of the proposed method and previous method are 0.953 and 0.965. In the second dataset, the best accuracy of the proposed method is 0.859 while the previous method is 0.695.