Hyperspectral Band Selection with Convolutional Neural Network

Band selection is a kind of dimension reduction method, which tries to remove redundant bands and choose several pivotal bands to represent the entire hyperspectral image (HSI). Supervised band selection algorithms tend to perform well because of the introduction of prior information. However, The traditional methods are based on the entire image, without taking into account the differences in ground categories, and cannot figure out which band is discriminative for a specific category. In this paper, a supervised method is proposed based on the ground category with convolutional neural network (CNN). Firstly, we propose a structure called contribution map which can record discriminative feature location. Secondly, the contribution map is added to CNN to generate a new model called contribution map based CNN (CM-CNN). Thirdly, we apply CM-CNN for HSI classification with the whole bands. Then, we can get the contribution map which records discriminative bands location for each category. Finally, the contribution map guides us to select discriminative bands. We found that CM-CNN model can obtain a satisfactory classification result while preserving the position information of important bands. To verify the superiority of the proposed method, experiments are conducted on HSI classification. The results demonstrated the reliability of the proposed method in HSI classification.