Band Weight Adaptive Classification Network for Hyperspectral Image Classification

Abstract

Hyperspectral images not only contain a large amount of ground object information in the plane space, but also carry a lot of spectral feature information in the spectral band dimension. Therefore, attention should be paid to both spatial and spectral features of hyperspectral images when classifying ground objects. Convolution neural network has remarkable ability of feature extraction, based on convolutional neural network, this paper proposes a dual-input convolutional neural network Band Weight Adaptive Classification Network (BWACN) and improves the Focal Loss Band Weight Adaptive Classification Network (FLBWACN). A branch of BWACN inputs two-dimensional spatial information of hyperspectral data, then extracts spatial features through 2D convolution, and maps spatial features to higher-dimensional space. The input of another branch of BWACN is the band information of hyperspectral images. The weight of each spectral band is initialized through the global average pooling layer, and then the weight of each band is trained implicitly by the attention mechanism in the subsequent training process, and the weight of each band is updated iteratively. BWACN can not only extract spatial information of hyperspectral image, but also extract band information, and can apply weight to spectral band through attention mechanism, so that the band containing important information plays a greater role in classification. Aiming at the problem of unbalanced sample number in actual data set, we improved FLBWACN model by using Focal Loss. Experiments show that the proposed model can achieve more than 93% accuracy when using only 5% of the training data.