Monitoring harmful algal blooms in surface water bodies across U.S. using Sentinel-2 images and Machine Learning

Abstract

Harmful algal blooms (HABs) are an increasing threat to surface water quality. Chlorophyll-a (Chl-a), a proxy for phytoplankton biomass, and Karenia brevis (K. brevis), the dominant species responsible for red tide events in coastal regions, are two key indicators for monitoring HABs. Traditional monitoring methods are accurate but inadequate, resulting in spatially and temporally sparse measurements. We developed a framework that integrates Sentinel-2 surface reflectance and machine learning (ML) to estimate Chl-a in U.S. rivers and K. brevis in the Gulf of Mexico. Four ML algorithms: XGBoost Linear, regularized random forest (RRF), K-nearest neighbors, and XGBoost Tree were developed, and model estimates were bias-corrected using tree-based binning. Based on the test results at 71 Chl-a monitoring stations and 75 K. brevis locations, we found that the XGBoost Linear was the best model for Chl-a (NSE: 0.87, RMSE:1.42 µg l−1, MAE: 0.72, Pbias: −3.2%), whereas RRF was the best for K. brevis (log-NSE: 0.77, log-RMSE: 1.22, log-MAE: 0.87, Pbias: 11.2%). More than 78% of gauge stations for Chl-a and 81% of K. brevis showed positive NSE. Site-wise evaluation shows that model performances are correlated with the observed Chl-a range (ρ = 0.47, p-value < 0.001) and K. brevis range (ρ = 0.61, p-value < 0.0001), indicating that sites with higher in situ variability tend to exhibit higher performances. This research highlights the feasibility of using Sentinel-2 images and ML for HAB monitoring across surface water bodies.