A data-driven approach to predict algal blooms using satellite-derived water quality and hydrometeorological drivers
The present work leverages simulated hydrometeorological factors and satellite-derived chlorophyll-a to predict phytoplankton dynamics for Mulargia reservoir (Sardinia, Italy). A Random Forest (RF) model was (a) calibrated to minimize out-of-bag errors of chlorophyll-a predictions for a 5-year-long period (2015-2019), and (b) benchmarked against a naïve predicting alternative for multiple forecasting horizons. Calibration and benchmarking revealed that the RF model can predict the temporal dynamics of phytoplankton growth accurately up to ten days in advance (mean absolute scaled errors ranged from 0.5 to 0.9). Permutation variable importance metrics and the individual conditional expectation plots revealed that moderate temperatures, high soluble phosphorus loads, and low light intensities favor the occurrence of phytoplankton growth in Mulargia. This finding is consistent with the dominance of Planktothrix sp. that has been observed in the reservoir. Conclusively, this work lays the foundation for an operational forecast model to help local stakeholders with the present and future reservoir management.
Machine learning; forecasting; phytoplankton blooms; remote sensing; hydrometeorological predictions
Kandris K., Romas E., Tzimas A., Bresciani M., Giardino C., Bauer P., Pechlivanidis I., Dessena M., 17th International Conference on Environmental Science and Technology, CEST2021, 1-4 September 2021, Athens, Greece. https://doi.org/10.30955/gnc2021.00420