Hyperspectral Retrieval Of Phytoplankton Absorption And Community Composition From Nasa’S Pace-Oci In Estuarine–Coastal Waters Using A Hybrid Framework Combining Mixture-Of-Experts And Variational Autoencoder

Authors

• Xingyu Bai, University of Delaware College of Engineering
• Bingqing Liu, Florida State University
• Jiang Li, University of Louisiana at Lafayette
• Yuanheng Xiong, The University of Southern Mississippi
• Eurico J. D'Sa, Louisiana State University
• Melissa M. Baustian, United States Geological Survey
• Xiaodong Zhang, The University of Southern Mississippi
• Brice K. Grunert, Cleveland State University
• Chisom O. Emeghiebo, University of Louisiana at Lafayette
• Cassie Glasspie, Louisiana State University
• Xu Yuan, University of Delaware College of Engineering

Document Type

Article

Publication Date

5-1-2026

School

Ocean Science and Engineering

Abstract

Retrieving the phytoplankton absorption coefficient (aphy[jls-end-space/]; m−1), one of the most spectrally rich inherent optical properties, remains challenging in optically complex coastal waters worldwide. Leveraging NASA's new hyperspectral mission, PACE, we introduce Hyper-MoE-VAE, a deep-learning architecture that integrates a Mixture-of-Experts with a Variational Autoencoder to retrieve high-dimensional aphy and subsequent estimation of phytoplankton community composition (PCC) from PACE-OCI hyperspectral remote sensing reflectance (Rrs[jls-end-space/]). Pre-trained on global hyperspectral bio-optical datasets and fine-tuned using regional field Rrs[jls-end-space/]–aphy pairings from inland– estuarine–coastal waters, Hyper-MoE-VAE demonstrated strong transferability and effective adaptation across regions. Validation with in-situ Rrs showed accurate aphy retrievals in Lake Erie (NRMSE = 0.12, ε = 17.10), Lake Pontchartrain (NRMSE = 0.11, ε = 37.12), and the Barataria–Terrebonne Estuary (NRMSE = 0.14, ε = 38.89). Using same-day PACE-OCI Level 2 Rrs[jls-end-space/], the model achieved comparable performance in Lake Erie (NRMSE = 0.19, ε = 55.19), Lake Pontchartrain (NRMSE = 0.14, ε = 51.39), and the Barataria–Terrebonne Estuary (NRMSE = 0.17, ε = 47.92). Hyper-MoE-VAE derived PACE-OCI hyperspectral aphy was further decomposed against mass-specific absorption spectra to estimate group-specific contributions to total chlorophyll a. The resulting PCC showed strong agreement with HPLC–CHEMTAX in Lake Erie (R2[jls-end-space/]= 0.692) and Gulf estuarine–coastal systems (R2 = 0.732). Monte Carlo noise experiments further revealed group-dependent sensitivities, with diatoms and dinoflagellates showing moderate susceptibility to noise, while cyanobacteria and cryptophytes exhibited narrow uncertainty distributions. These results demonstrate Hyper-MoE-VAE's capability for regional, operational water-quality monitoring with PACE-OCI and its adaptability to current and future hyperspectral missions.

Publication Title

Remote Sensing of Environment

Volume

337

Recommended Citation

Bai, X., Liu, B., Li, J., Xiong, Y., D'Sa, E., Baustian, M., Zhang, X., Grunert, B., Emeghiebo, C., Glasspie, C., Yuan, X. (2026). Hyperspectral Retrieval Of Phytoplankton Absorption And Community Composition From Nasa’S Pace-Oci In Estuarine–Coastal Waters Using A Hybrid Framework Combining Mixture-Of-Experts And Variational Autoencoder. Remote Sensing of Environment, 337.
Available at: https://aquila.usm.edu/fac_pubs/22025