Multi-model approach to predict phytoplankton biomass and composition dynamics in a eutrophic shallow lake governed by extreme meteorological events

- We combined powerful statistical tools and physically-based models to analyze and predict key processes of phytoplankton dynamics in a eutrophic shallow lake destined for drinking water.
- An 11-year time series of meteorological, hydrological, physicochemical and biological variables, continuously measured by water managers was considered.
- The models were able to explain a significant part of the variability of phytoplankton biomass and composition and presented acceptable predictive accuracy.
- In the current water quality situation, predictive models are crucial to anticipate non-desirable water quality conditions.

A multi-model approach to predict phytoplankton biomass and composition was performed in a eutrophic Uruguayan shallow lake which is the second drinking water source of the country. We combined statistical (spectral analysis and Machine learning techniques) and physically based models to generate, for the first time in this system, a predictive tool of phytoplankton biomass (chlorophyll-a) and composition (morphology-based functional groups). The results, based on a 11-year time series, revealed two alternating phases in the temporal dynamics of phytoplankton biomass. One phase is characterized by high inorganic turbidity and low phytoplankton biomass, and the other by low inorganic turbidity and variable (low and high) phytoplankton biomass. A threshold of turbidity (29 TNU), above which phytoplankton remains with low biomass (<15-20 ug/l) was established. The periods of high turbidity, which in total cover 30% of the time series, start abruptly and are related to external forcing. Meteorological conditions associated with the beginning of these periods were modeled through a regression tree analysis. These conditions consist of moderate to high wind intensities from the SW direction, in some cases combined with high antecedent precipitation or low water level. The results from the physically-based modeling indicated that the long decaying time-scale of turbidity and intermediate resuspension events could explain the prolonged length of the high turbidity periods (∼1.5 years). Random Forests models for the prediction of phytoplankton biomass and composition in periods of low turbidity resulted in a proportion of explained variance and a classification error over a test sample of 0.46 and 0.34 respectively. Turbidity, conductivity, temperature and water level were within the most important model predictors. The development and improvement of this type of modeling is needed to provide management tools to water managers in the current water supply situation.