Haze Thickness Prior-Injected Spectral-Spatial Dehazing Network for Hyperspectral Image

Abstract

Haze pollution leads to reduced spatial visibility and spectral distortion in remote sensing hyperspectral images (HSIs), reducing application capabilities. The distribution of haze in HSIs is spatially irregular and spectrally inhomogeneous, and the effective separation of the haze and the scene information is challenging. Current prior-based methods exhibit weak generalization, and many learning-based approaches introduce irrelevant redundant information, resulting in high computational costs and inadequate utilization of negative samples. To this end, we explore the potential of prior knowledge-guided learning of dehazing networks and propose a haze thickness prior (HTP)-injected spectral-spatial dehazing network for HSI. In particular, by analyzing the difference in haze sensitivity in the visible and infrared, the HTP and prior fusion attention module (PFAM) are proposed to bootstrap the learning of multiple layers of the network, thus enhancing the haze perception capability of the network. We aim to design the spectral-spatial dehazing module (SSDM), which enables the extraction of intrinsic features related to dehaze and reduces the redundancy of features and parameters. In addition, we introduce a multilevel contrastive regularization (MCR) combined with perceptual loss. By explicitly exploiting feature differences across multiple levels between positive and negative samples, this constraint guides the network to learn more discriminative representations and improves dehazing performance. Experimental results on the Gaofen-5 HSI dataset demonstrate that HTP-SSDN has efficient performance that outperforms state-of-the-art methods.