Blind and endmember guided autoencoder model for unmixing the absorbance spectra of phytoplankton pigments

Abstract

Abstract Hyperspectral sensing of phytoplankton, free-living microscopic photosynthetic organisms, offers a comprehensive and scalable method for assessing water quality and monitoring changes in aquatic ecosystems. However, unmixing the intrinsic optical properties of phytoplankton from hyperspectral data is a complex challenge. This research addresses the problem of non-linear unmixing hyperspectral absorbance data of concentrated water samples using Blind (BAE) and Endmember Guided Autoencoder (EGAE). We show that spectral unmixing using the EGAE model with different objective functions can effectively estimate the abundance of different optical components in spectral data. The EGAE model demonstrated a higher correlation between unmixed endmember abundances and ground truth for chlorophyll-a (chl-a) and fucoxanthin (fx) biomarker pigment concentrations compared to the BAE model, effectively unmixed the absorbance spectrum of cyanobacterial pigment phycocyanin (pc) and was robust to changes in network architecture. It can adaptively unmix various endmembers without impacting the abundance estimates of other pigments. Our results demonstrate that EGAE provided stable abundance estimates and improved the accuracy and reliability of identifying and quantifying pigments, allowing for more precise unmixing of hyperspectral data into their constituent endmembers. We anticipate that our study will serve as a starting point for targeted unmixing of specific photosynthetic pigments using EGAE.