Efficient Global Context Graph Convolution for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification has been widely used in remote sensing image analysis. Rich contextual information is beneficial to improve classification performance. Recently, plenty of attention-based CNN networks have been proposed for HSI classification due to the ability of the attention mechanism to perceive the global context. However, these methods have a high memory cost because the attention mechanism directly models each pair of pixel relationships, resulting in the size of the affinity matrix being very large. And it also failed to sufficiently leverage the relationship between pixels in HSI data. Regarding the problem, a lightweight and efficient end-to-end model is proposed for hyperspectral image classification in this paper. To efficiently extract global context, we propose a novel graph construction module, which shares the global attention map among all features and learns relations only between important features, thus saving computational overhead. In order to make better use of the context, we further aggregate the features relationships through the graph convolution module to achieve more accurate hyperspectral image classification. Experiments on three HSI datasets show it not only achieves better performance than other classification methods, but also improves calculation efficiency.