Abstract:
Eutrophication is a worldwide environmental problem affecting many freshwater systems that are seeing an increased risks of harmful algal blooms (HABs). Water management agencies are under pressure to monitor these systems to avert the adverse impacts of HABs. Traditional monitoring programs that are based on collecting field data have become increasingly difficult to rely on given their poor spatio-temporal coverage and mounting costs. This study aims to assess the potential of using Bayesian Networks (BNs) to predict and assess HAB related water quality impairments in lakes across the contiguous United States, using remotely sensed data collected from the family of Landsat satellites. The performance of Landsat 5, 7 and 8 individual bands and band ratios was assessed for predicting Chlorophyll-a (Chl-a) concentrations and Secchi Disk Depths (SDDs). The results underscored the significant role that the Green-to-Blue ratio and the green band had when predicting Chl-a levels. Meanwhile, our results showed that the Blue-to-Red ratio and the red band were the most informative when estimating SDD levels, especially in mesotrophic, eutrophic and hypereutrophic lakes. Moreover, the results showed significant differences in model performance across trophic state and ecoregions. The performance of the discrete and hybrid Bayesian Networks were found to be largely identical for both water quality parameters. Nevertheless, the latter is preferred given that the discrete models required the discretization of all continuous variables; a process that is known to introduce errors.
