Progressive split-merge super resolution for hyperspectral imagery with group attention and gradient guidance

Abstract

Hyperspectral image (HSI) has been an important valuable information source for many remote sensing applications due to its rich spectral information. However, limited by existing imaging system, the spatial resolution of HSI greatly limits its practical applications. In this paper, in order to deal with the challenge of feature extraction and super-resolution of HSI, a progressive split-merge super-resolution (PSMSR) framework is proposed to overcome the inherent resolution limitations, which apply a multi-level split strategy from the task level, spectral level and feature level. Within this framework, a gradient-guided group-attention network (GGAN) is designed to extract the spatial-spectral features and reconstruct realistic texture, where group attention block (GAB) aims to increase the distinguishing ability of spectral channels, and gradient information is fully fused in the reconstruction process to promote sharp edges and realistic texture. Compared with current state-of-the-art CNN-based SR methods, the proposed method achieves significant improvement in six evaluation metrics and visual quality on multiple and diverse scenes, and it obtains a higher classification accuracy in classification experiment and gets better SR results on real data, which proves that our method can effectively improve the spatial resolution while preserving the spectral correlation. In addition, a series of ablation experiments prove the effectiveness of each component of our method.