Resolving phytoplankton pigments from spectral images using convolutional neural networks

• DOI: 10.1002/lom3.10588
• Publication Date: 2023-11-06T09:15:27Z
• Authors: Pauliina Salmi; Ilkka Pölönen; Daniel Atton Beck...; Marco L. Calderini; Linda May; Justyna Olszewska; Laura Perozzi; Salli Pääkkönen; Sami Taipale; Peter Hunter
• Keywords: ta113; Computing, Information Technology and Mathematics; Akvaattiset tieteet; plankton; School of Resource Wisdom; 04 agricultural and veterinary sciences; Tietotekniikka; sisävedet; vedenlaatu; järvet; 6. Clean water; Computational Science; Resurssiviisausyhteisö; karotenoidit; Aquatic Sciences; 13. Climate action; Mathematical Information Technology; ta1181; 0401 agriculture, forestry, and fisheries; 14. Life underwater; Laskennallinen tiede

Abstract Motivated by the need for rapid and robust monitoring of phytoplankton in inland waters, this article introduces a protocol based on mobile spectral imager assessing pigments from water samples. The includes (1) sample concentrating; (2) imaging; (3) convolutional neural networks (CNNs) to resolve concentrations chlorophyll (Chl ), carotenoids, phycocyanin. was demonstrated with samples 20 lakes across Scotland, special emphasis Loch Leven where blooms cyanobacteria are frequent. In parallel, were prepared reference observations Chl carotenoids high‐performance liquid chromatography phycocyanin spectrophotometry. Robustness CNNs investigated excluding each lake model trainings one at time using excluded data as independent test data. For Leven, median absolute percentage difference (MAPD) 15% 36% carotenoids. MAPD estimated concentration high (102%); however, system able indicate possibility bloom. leave‐one‐out tests other lakes, 26% , 27% 75% higher error likely due variation distribution observations. It concluded that could support proxies biomass. Greater focus volume training would improve estimates.