Comparative evaluation of spectral variability-aware mixing models for marine mucilage monitoring

Marine mucilage poses ecological and economic concerns in coastal regions, prompting the need for timely and accurate monitoring solutions. Hyperspectral remote sensing provides rich spectral information that enables unsupervised mapping of mucilage without requiring labeled training data. Although recent studies have employed unmixing under the linear mixing model (LMM) for this purpose, LMM does not account for spectral variability, which can lead to inaccuracies in abundance estimation. We evaluate five spectral variability-aware hyperspectral unmixing models, namely the extended linear mixing model (ELMM), generalized linear mixing model (GLMM), perturbed linear mixing model (PLMM), modified augmented linear mixing model (mALMM), and additively tuned mixing model (ATMM) for improved marine mucilage monitoring, using PRISMA and Earth Surface Mineral Dust Source Investigation (EMIT) hyperspectral data, capturing mucilage events in the Sea of Marmara and the Adriatic Sea, respectively. Performance comparisons are conducted using reconstruction accuracy, spectral similarity to in situ spectra, and mapping accuracy based on pseudo ground truth maps derived from these spectra. Among the tested approaches, ATMM consistently provided balanced mucilage mapping performance across sensors and regions, whereas PLMM achieved the lowest reconstruction error. The findings provide insights into the effectiveness of spectral variability-aware unmixing methods, contributing to more robust and reliable monitoring approaches that can support early intervention and management of marine mucilage events.