Examining the effect of physicochemical and meteorological variables on water quality indicators of harmful algal blooms in a shallow hypereutrophic lake using machine learning techniques

Two independent machine learning techniques, boosted regression trees and artificial neural networks, were used to examine the physicochemical and meteorological variables that affect the seasonal growth and decline of harmful algal blooms (HABs) in a shallow, hypereutrophic lake in southern Oregon. High temporal resolution data collected at four monitoring locations were aggregated into daily timesteps to create two response variables: (1) daily maximum pH (pH_max), representing HAB growth, and (2) daily minimum dissolved oxygen (DO_min), representing HAB decline. Predictors included meteorological and physical data, estimates of external phosphorus loading, and previous-year average nutrient concentrations, and excluded HAB biomass and internal phosphorus loading. The predictors that captured seasonal changes in both pH_max and DO_min were temperature, inflows, lake-surface elevation, and external phosphorus loading, while short-term changes were captured by measures of stratification, temperature, and wind speed. The pH_max models had similar fits with leave-one-year-out cross-validation (LOYO-CV) R² values of 0.2–0.43 (median = 0.40). The DO_min models for the deeper locations had LOYO-CV R² values of 0.27–0.43 compared to 0.1–0.25 for the shallower locations. Model performance was affected by variability due to patchiness of HABs, measurement uncertainty, and advection.