Pixel-scale satellite forecasting of cyanobacteria in Florida lakes

Rapid proliferation of potentially toxin-producing cyanobacteria (cyanobacterial harmful algal blooms, cHABs) is a significant challenge for water resource managers. The U.S. Harmful Algal Bloom and Hypoxia Research Control Act calls for robust approaches to forecasting cHABs in lakes and reservoirs. A previous national study developed a lake scale Bayesian spatiotemporal model to forecast weekly chlorophyll-a exceedance probability in 2192 satellite resolved lakes using Sentinel-3 Ocean Land Colour Instrument data. Building on this foundation, this study developed a machine learning based, spatially explicit forecasting model at the 300-m pixel scale for Sentinel-3 resolvable lakes, conditioned on environmental variables from 2017 - 2024. Three machine learning models were constructed for Sentinel-3 resolvable lakes across Florida to generate ensemble forecasts of chlorophyll-a concentration and World Health Organization Alert Level categories with multi-week leads. The pixel scale metric significantly increased the number of resolvable bloom events, recording 4682 (33%) more exceedance events than the lake scale metric. The best performing model (random forest) achieved 88.2% accuracy at Alert Level 1 and 92.2% accuracy at Alert Level 2. Overall R² was 0.56 and mean absolute error was 5.4 μgL^-1 The best performing model had notable skill at 1-2 week leads but diminished at a 4-week horizon. Probabilistic forecast performance shows that the model predicted observed chlorophyll concentrations and WHO Alert Levels with skill. Direct comparison showed similar skill between the lake and pixel scale forecast.