Leveraging High-Frequency Water Quality Data and Machine Learning for Monitoring Harmful Algal Blooms

Freshwater management is one of the most critical resources on the earth due to the plethora of water use and its limited availability. Increased algae proliferation is one of the significant problems of freshwater bodies that is triggered by nutrient enrichment and enabling conditions such as light and warm temperatures. This algal proliferation is toxic to the ecosystem through their biomass and potential toxin production that can cause hypoxic conditions and is often referred to as Harmful Algal Blooms (HABs). Current monitoring approaches include laboratory analysis of water samples to observe algal cells, monitoring of water quality parameters using water quality sensors, use of biosensors to monitor algae cells, use of remotely sensed data to monitor HABs expansion, process-based models to simulate algal dynamics, and use of data-driven models to predict HABs occurrence. However, microscopic and other laboratory techniques are time-consuming and require high taxonomic skills. Likewise, remote sensing methods are limited by the inability to monitor HABs beneath water surfaces since reflectance values are often based on images capturing upper layers of water surfaces. Process-based models for HABs monitoring are also limited by the uncertainty of kinetic rate coefficients for different algal species and the complexity of currently existing models that necessitates data-driven models to understand the dynamics of HABs. This research utilizes high-frequency water quality data and different machine learning models to understand the dynamics of HABs in Boyd Millpond using chlorophyll-a concentration as index. The results indicate that accurate three-day-ahead predictions of HABs can be reliably achieved using a daily frequency of observations, while longer-range predictions, such as those extending to seven days or more, are more accurate with hourly observation data. In addition, the research showed that prediction errors increase as data assimilation frequency reduces with the highest accuracy obtained when new observations are assimilated daily. The result provides HABs investigators with valuable insights into methods to reduce the impact of measurement and model uncertainty when employing ML models for decision-making relating to HABs management. The outcome of this research will ensure timely and accurate HABs predictions to support effective management strategies and decision-making efforts.