Hyperspectral retrievals of phytoplankton absorption and chlorophyll-a in inland and nearshore coastal waters

Following more than two decades of research and developments made possible through various proof-of-concept hyperspectral remote sensing missions, it has been anticipated that hyperspectral imaging would enhance the accuracy of remotely sensed in-water products. This study investigates such expected improvements and demonstrates the utility of hyperspectral radiometric measurements for the retrieval of near-surface phytoplankton properties(1), i.e., phytoplankton absorption spectra (a(ph)) and biomass evaluated through examining the concentration of chlorophyll-a (Chla). Using hyperspectral data (409-800 nm at similar to 5 nm resolution) and a class of neural networks known as Mixture Density Networks (MDN) (Pahlevan et al., 2020), we show that the median error in a(ph) retrievals is reduced two-to-three times (N = 722) compared to that from heritage ocean color algorithms. The median error associated with our a(ph) retrieval across all the visible bands varies between 20 and 30%. Similarly, Chla retrievals exhibit significant improvements (i.e., more than two times; N = 1902), with respect to existing algorithms that rely on select spectral bands. Using an independent matchup dataset acquired near-concurrently with the acquisition of the Hyperspectral Imager for the Coastal Ocean (HICO) images, the models are found to perform well, but at reduced levels due to uncertainties in the atmospheric correction. The mapped spatial distribution of Chla maps and a(ph) spectra for selected HICO swaths further solidify MDNs as promising machine-learning models that have the potential to generate highly accurate aquatic remote sensing products in inland and coastal waters. For a(ph) retrieval to improve further, two immediate research avenues are recommended: a) the network architecture requires additional optimization to enable a simultaneous retrieval of multiple in-water parameters (e.g., a(ph), Chla, absorption by colored dissolved organic matter), and b) the training dataset should be extended to enhance model generalizability. This feasibility analysis using MDNs provides strong evidence that high-quality, global hyperspectral data will open new pathways toward a better understanding of biodiversity in aquatic ecosystems.