HUTDNet: A joint unmixing and target detection network for underwater hyperspectral imagery

Abstract

Underwater hyperspectral target detection (HTD) technology holds pivotal value in enhancing maritime military power. However, the absorption and scattering properties of the water bodies result in the inevitable issue of mixed pixels in underwater hyperspectral images (HSIs). To address the issue, a joint hyperspectral unmixing and target detection network for underwater HSI is proposed, denoted as HUTDNet, which utilizes the material type and abundance information for downstream semantic tasks. Specifically, a nonlinear underwater unmixing network is designed to extract pure underwater endmembers and their associated abundance information, which is essential in assisting the subsequent target detection task. The network also extracts underwater virtual endmembers and their abundance values to reconstruct a more realistic underwater HSI. Then, the abundance weighting module determines the abundance weighting factor by calculating the spectral distance between a priori target spectra and the estimated underwater pure endmembers, generating a weighted abundance map. Finally, due to the inherent limitations in the characterization capabilities of abundance maps and endmembers, the detection network extracts key spectral feature maps from the input underwater HSI. These feature maps serve as complementary terms, fused with the original and weighted abundance maps. Subsequently, convolutional and fully connected layers are employed to extract deeper features and generate the target detection maps. Experiments on both real and synthetic datasets demonstrate the superior performance and efficiency of the proposed method in this paper compared to other state-of-the-art methods.

Keywords

• Hyperspectral imagery
• Hyperspectral unmixing
• Underwater target detection
• Deep learning