Characterising retrieval uncertainty of chlorophyll-a algorithms in oligotrophic and mesotrophic lakes and reservoirs

Remote sensing product uncertainties for phytoplankton chlorophyll-a (chla) concentration in oligotrophic and mesotrophic lakes and reservoirs were characterised across 13 existing algorithms using an in situ dataset of water constituent concentrations, inherent optical properties (IOPs) and remote-sensing reflectance spectra Rrsλ collected from 53 lakes and reservoirs (346 observations; chla concentration < 10 mg m-3, dataset median 2.5 mg m-3). Substantial shortcomings in retrieval accuracy were evident with median absolute percentage differences (MAPD) > 37% and mean absolute differences (MAD) > 1.82 mg m-3. Using the Hyperspectral Imager for the Coastal Ocean (HICO) band configuration improved the accuracies by 10–20% compared to the Ocean and Land Colour Instrument (OLCI) configuration. Retrieval uncertainties were attributed to optical and biogeochemical properties using machine learning models through SHapley Additive exPlanations (SHAP). The chla retrieval uncertainty of most semi-analytical algorithms was primarily determined by phytoplankton absorption and composition. Machine learning chla algorithms showed relatively high sensitivity to light absorption by coloured dissolved organic matter (CDOM) and non-algal pigment particulates (NAP). In contrast, the uncertainties of red/near-infrared algorithms, which aim for lower uncertainty in the presence of CDOM and NAP, were primarily explained through the total absorption by phytoplankton at 673 nm (aϕ(673)) and variables related to backscatter. Based on these uncertainty characterisations we discuss the suitability of the evaluated algorithm formulations, and we make recommendations for chla estimation improvements in oligo- and mesotrophic lakes and reservoirs.