Transformer-Driven Inverse Problem Transform for Fast Blind Hyperspectral Image Dehazing

TL;DR

This work tackles hyperspectral image dehazing (HyDHZ) by reframing the challenging inverse problem into a spectral super-resolution (SSR) task via an inverse problem transform (IPT). It introduces the T$^2$HyDHZ framework, which integrates an Auto Band Selection (ABS) module, a Spectral Reconstruction (SR) module, and a Spectral-Spatial Enhancement (SSE) transformer to produce clean hyperspectral data without manual haze-region labeling. The model is trained with a combined loss that includes refined MRAE and an L1 sparsity term to encourage selecting informative long-wavelength bands, and it employs a novel spatial-spectral transformer to capture global dependencies. Across synthetic and real AVIRIS hazy data, T$^2$HyDHZ achieves state-of-the-art quantitative and qualitative performance while being significantly faster than prior methods, enabling practical deployment in remote sensing workflows. Overall, the IPT-motivated, transformer-based approach offers a robust, blind, and efficient solution for HyDHZ with strong implications for downstream spectral analysis.

Abstract

Hyperspectral dehazing (HyDHZ) has become a crucial signal processing technology to facilitate the subsequent identification and classification tasks, as the airborne visible/infrared imaging spectrometer (AVIRIS) data portal reports a massive portion of haze-corrupted areas in typical hyperspectral remote sensing images. The idea of inverse problem transform (IPT) has been proposed in recent remote sensing literature in order to reformulate a hardly tractable inverse problem (e.g., HyDHZ) into a relatively simple one. Considering the emerging spectral super-resolution (SSR) technique, which spectrally upsamples multispectral data to hyperspectral data, we aim to solve the challenging HyDHZ problem by reformulating it as an SSR problem. Roughly speaking, the proposed algorithm first automatically selects some uncorrupted/informative spectral bands, from which SSR is applied to spectrally upsample the selected bands in the feature space, thereby obtaining a clean hyperspectral image (HSI). The clean HSI is then further refined by a deep transformer network to obtain the final dehazed HSI, where a global attention mechanism is designed to capture nonlocal information. There are very few HyDHZ works in existing literature, and this article introduces the powerful spatial-spectral transformer into HyDHZ for the first time. Remarkably, the proposed transformer-driven IPT-based HyDHZ (T2HyDHZ) is a blind algorithm without requiring the user to manually select the corrupted region. Extensive experiments demonstrate the superiority of T2HyDHZ with less color distortion.

Figures (28)

• Figure 1: Graphical illustration of the idea of inverse problem transform (IPT), which solves the more challenging IP1 in another domain wherein a relatively easier IP2 can be more efficiently addressed.
• Figure 2: Graphical illustration of the proposed T$^2$HyDHZ network structure, which consists of three fundamental blocks, namely the auto band selection (ABS), spectral reconstruction (SR), and spectral-spatial enhancement (SSE) blocks. Particularly, the SSE block comprises spectral refinement (SpeR) and spatial refinement (SpaR), which adopt global self-attention in specific spectral and spatial dimensions (i.e., $\text{Att}_{\text{spe}}$ and $\text{Att}_{\text{spa}}$), respectively. Additionally, a feed-forward network (FFN) is employed to obtain higher-level features after the global attention mechanism in the refinement step.
• Figure 4: Visual presentation of bands selected by the ABS block. The bands are ordered according to their wavelengths, so the infrared bands are indexed by those $i>41$. One can see that those informative bands (i.e., with higher attention weights) do mostly concentrate on the relatively clean infrared region.
• Figure 5: Diagram of the global attention process in the transformer, where the global attention score is performed by each token (pixel vector) pair from the entire image, resulting in a global representation of contextual associations attentionisallyoueed.
• Figure 6: Qualitative study of AVIRIS real hyperspectral imagery from NASA over Kettle Moraine, USA.