From sample to sonde to Sentinel-2: insights from a multi-scale chlorophyll-a monitoring effort in the Hudson River, New York

Abstract Monitoring cyanobacteria and other nuisance phytoplankton in the Hudson River is of great interest given its societal and ecological importance. Satellite remote sensing provides a cost-effective method to monitor chlorophyll- a (chl-a), a common proxy for algal biomass; however, the dynamic nature of rivers complicates approaches traditionally applied to lakes and oceans. During 2021–2023, we collected discrete samples for laboratory measurement of chl-a and measured in situ chl-a fluorescence during a series of longitudinal boat surveys along a 220-km reach of the lower Hudson River. Surveys were timed to coincide with Sentinel-2 satellite overpasses. We first investigated relations between laboratory-measured chl-a concentration and field-measured chl-a fluorescence, observing a weak correlation ( r 2 = 0.25) that improved substantially after splitting data by day (mean r 2 = 0.53). Separately, to estimate chl-a fluorescence using satellite data, we developed a series of random forest models leveraging the rich fluorescence dataset collected. We tested three model types: individual day models, leave-one-out models trained on all days except a holdout test day, and a single pooled model trained on all days. Generally, individual day models exhibited lowest error (mean of mean absolute error [MAE] = 0.16 relative fluorescence units [RFU]), followed by the single pooled model (MAE = 0.22 RFU). Daily holdout models showed highest error (mean MAE = 0.40 RFU); this approach was intended to represent model performance on a day unseen in the training set, providing a more conservative estimate of performance than the more traditional pooled approach. Findings from both analyses emphasize the importance of considering temporal variability when modeling riverine systems.