Near-term forecasting of cyanobacteria and harmful algal blooms in lakes using simple univariate methods with satellite remote sensing data

Abstract

Near-term forecasting of cyanobacteria and harmful algal blooms (HABs) in lakes is essential to reduce risks to human and animal health and water treatment. Cyanobacteria forecasting models are typically complex, requiring input of biophysical and chemical measurements or DNA sequencing in situ. Satellite imagery presents a unique opportunity to estimate cyanobacteria concentration directly at low cost and over wide spatial and long timescales. This study explores the hypothesis that simple univariate forecasting methods can reliably forecast cyanobacterial blooms in the near-term (1 week ahead) detected using satellite remote sensing. A simple univariate model based on logical decomposition with a moving average and seasonal component was developed to forecast chlorophyll a concentrations from cyanobacteria and algal blooms in lakes using spatially aggregated satellite remotely sensed data. A small test set of 15 spatially distributed waterbodies was used to assess forecast performance on 1-week, 2-week, and 4-week forecast horizons using a year-long hold-out time series. For a 1-week time horizon, cyanobacterial blooms posing a high health risk could be forecast with 80% accuracy. The 2-week and 4-week forecast accuracy dropped to 71% and 69%, respectively. Forecast performance was only weakly influenced by lake size, suggesting that the spatial-aggregation approach may be valid even for large lakes. Additionally, longer time series reduced the observed forecast error, presumably because of better seasonal characterization. This study is the first to demonstrate that simple univariate models with remotely sensed time series can forecast cyanobacteria and HABs with almost the same reliability as complex models.