Remote retrieval of near-surface chlorophyll-a (Chla) concentration in small inland waters is challenging due to substantial optical interferences of various water constituents and uncertainties in the atmospheric correction (AC) process. Although various algorithms have been developed to estimate Chla from moderate-resolution terrestrial missions (∼10–60 m), the production of both accurate distribution maps and time series of Chla has proven challenging, limiting the use of remote analyses for lake monitoring. Here, we develop a support vector regression (SVR) model, which uses satellite-derived remote-sensing reflectance spectra (Rrsδ) from Sentinel-2 and Landsat-8 images as input for Chla retrieval in a representative eutrophic prairie lake, Buffalo Pound Lake (BPL), Saskatchewan, Canada. Validated against in situ Chla from seven ice-free seasons (N ∼ 200; 2014–2020), the SVR model outperformed both locally tuned, Rrsδ-fed empirical models (Normalized Difference Chlorophyll Index, 2- and 3-band, and OC3) and Mixture Density Networks (MDNs) by 15–65%, while exhibiting comparable performance to a locally trained MDN, with an error of ∼35%. Comparison of Chla retrieval models, AC processors (iCOR, ACOLITE), and radiometric products (Rayleigh-corrected, surface, and top-of-atmosphere reflectance) showed that the best Chla maps and optimal time series (up to 100 mg m−3) were produced using a coupled SVR-iCOR system.

Lake Erie, the shallowest of the five North American Laurentian Great Lakes, exhibits degraded water quality associated with recurrent phytoplankton blooms. Optical remote sensing of these optically complex inland waters is challenging due to the uncertainties stemming from atmospheric correction (AC) procedures. In this study, the accuracy of remote sensing reflectance (Rrs) derived from three different AC algorithms applied to Ocean and Land Colour Instrument (OLCI) observations of western Lake Erie (WLE) is evaluated through comparison to a regional radiometric dataset. The effects of uncertainties in Rrs products on the retrieval of near-surface concentration of pigments, including chlorophyll-a (Chla) and phycocyanin (PC), from Mixture Density Networks (MDNs) are subsequently investigated. Results show that iCOR contained the fewest number of processed (unflagged) days per pixel, compared to ACOLITE and POLYMER, for parts of the lake. Limiting results to the matchup dataset in common between the three AC algorithms shows that iCOR and ACOLITE performed closely at 665 nm, while outperforming POLYMER, with the Median Symmetric Accuracy (MdSA) of ∼30 %, 28 %, and 53 %, respectively. MDN applied to iCOR- and ACOLITE-corrected data (MdSA < 37 %) outperformed MDN applied to POLYMER-corrected data in estimating Chla. Large uncertainties in satellite-derived Rrs propagated to uncertainties ∼100 % in PC estimates, although the model was able to recover concentrations along the 1:1 line. Despite the need for improvements in its cloud-masking scheme, we conclude that iCOR combined with MDNs produces adequate OLCI pigment products for studying and monitoring Chla across WLE.

Cyanobacterial harmful algal blooms are an increasing threat to coastal and inland waters. These blooms can be detected using optical radiometers due to the presence of phycocyanin (PC) pigments. The spectral resolution of best-available multispectral sensors limits their ability to diagnostically detect PC in the presence of other photosynthetic pigments. To assess the role of spectral resolution in the determination of PC, a large ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$N =905$ </tex-math></inline-formula> ) database of colocated <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in situ</i> radiometric spectra and PC are employed. We first examine the performance of selected widely used machine-learning (ML) models against that of benchmark algorithms for hyperspectral remote sensing reflectance ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$R_{\mathrm {rs}}$ </tex-math></inline-formula> ) spectra resampled to the spectral configuration of the Hyperspectral Imager for the Coastal Ocean (HICO) with a full-width at half-maximum (FWHM) of < 6 nm. Results show that the multilayer perceptron (MLP) neural network applied to HICO spectral configurations (median errors < 65%) outperforms other ML models. This model is subsequently applied to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$R_{\mathrm {rs}}$ </tex-math></inline-formula> spectra resampled to the band configuration of existing satellite instruments and of the one proposed for the next Landsat sensor. These results confirm that employing MLP models to estimate PC from hyperspectral data delivers tangible improvements compared with retrievals from multispectral data and benchmark algorithms (with median errors between <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 73$ </tex-math></inline-formula> % and 126%) and shows promise for developing a globally applicable cyanobacteria measurement approach.