Scale-Spectral-Spatial Attention Network for Hyperspectral Image Classification

Attention networks enable neural networks to focus on the most beneficial parts of their input. In the context of remote sensing image classification, studies about spatial, spectral and spatial-spectral attention networks have already been reported. In this paper, a network integrating a scale-based attention module, in addition to spatial-spectral attention is proposed. The scale-space has been produced via alpha-trees, in order for the network to focus on the most useful scales. It is tested with two real hyperspectral datasets, where it achieves a performance improvement.