A New Version of an Endmember-Guided Autoencoder (EGAE-V2) with Improved Architecture and Regularization by Correlation Between Ground Truths and Latent Activations

Neural networks, particularly autoencoders, offer a promising approach for hyperspectral unmixing of photosynthetic pigments in natural water. Unmixing refers to separating the spectral signatures of substances and estimating their relative abundances. This process is challenging in suspensions such as water samples because dissolved organic matter, suspended particles, and various types of phytoplankton cause overlapping spectral signatures, making it difficult to distinguish specific pigments. The previously developed Endmember-guided autoencoder (EGAE) performs well in unmixing chlorophyll-a (chl-a) and phycocyanin (pc), which have distinct absorption peaks. However, unmixing fucoxan-thin (fx) is more difficult due to its less distinct absorption features that overlap with other pigments, water colour and non-algal particles. Our study showed that applying an improved model (EGAE-V2) with enhanced architecture and new regularization method using ground truth data improves unmixing of fucoxanthin without affecting the unmixing performance of other pigments.