A Global-Local Spectral Weight Network Based on Attention for Hyperspectral Band Selection

Hongqi Zhang, Xudong Sun, Yuan Zhu, Fengqiang Xu, Xianping Fu

Abstract: Band selection (BS) methods based on deep learning have achieved significant development. However, most existing band selection methods commonly utilize a fully connected neural network (FCN) or convolutional neural network (CNN) to explore the correlation among bands and rarely combine the two styles of the network to select bands. Moreover, almost all the methods employ the form of the combination of $L_{1}$ norm and Sigmoid to constitute attention model, which may lead to losing some informative band feature. To tackle these troubles, this letter proposes a novel band selection network using FCN and CNN, termed as global-local spectral weight network based on attention (GLSWA), in which the band features of each pixel is mined using the network of two types, and designing an attention-based scoring module (ASM) and a convolutional reconstruction module (CRM), respectively, so that each attention of band is adjusted by simultaneous considering the entire band features and successive one. Experimental results on three real hyperspectral image (HSI) datasets show that the proposed method achieves satisfactory accuracy than some state-of-the-art algorithms.