SVHAE: Spectral Variability-Aware Hybrid Autoencoder for Hyperspectral Underwater Target Detection

Hyperspectral imaging (HSI) has evolved as an important tool for many applications, including remote sensing, crime investigation, target detection, disease diagnosis, and anomaly detection. Among these, hyperspectral underwater target detection (HUTD) presents unique challenges due to spectral distortions caused by water absorption and scattering. Conventional methods often struggle with spectral variability, complicating detection accuracy. This paper introduces spectral variability-aware hybrid autoencoder (SVHAE) for HUTD, a novel autoencoder-based unmixing network incorporating parallel linear and nonlinear decoders to improve underwater target detection. Our method effectively reduces the effect of spectral distortions and addresses variability using a combined loss function integrating Kullback-Leibler divergence, mean squared error, and spectral angle distance. Experimental validation on real-world and simulated datasets demonstrates that our proposed SVHAE outperformed state-of-the-art methods by achieving superior AUC values. These advancements contribute to the progressing field of HUTD, making the way for robust solutions in marine exploration and detecting targets under the water.