Using machine learning to model dead zones in lakes

Aquatic ecosystems are complex environments that can be affected by many variables, including weather, the biological activities of the organisms living in them, and anthropogenic nutrient pollution. The impact of these variables on aquatic ecosystems can also depend on the properties of the water body, such as temperature and depth. These interconnected processes can become unbalanced, with devastating consequences.

To anticipate these consequences, a group of UConn researchers has developed a versatile computer modeling method that uses machine learning to improve existing lake water quality monitoring and prediction efforts. The method was recently published in Environmental Modeling & Software.

Marina Astitha, Associate Professor at the Faculty of Civil and Environmental Engineering and Head of the Atmospheric and Air Quality Modelling Group, explains that the research took five years and was carried out in collaboration with a former student, Christina Feng Chang (Dr. 22), as part of her dissertation, and Professor Penny Vlahos, Head of the Environmental Chemistry and Geochemistry Research Group at the Faculty of Marine Sciences.

Aquatic environments are prone to eutrophication, a process triggered by excess nutrients, primarily caused by fertilizer runoff from agricultural activities, entering aquatic ecosystems and causing algal blooms. The increased growth and eventual decomposition of these plant-like materials consumes much or all of the available oxygen, which has detrimental effects on other organisms in the environment. Low-oxygen or hypoxic areas are called “dead zones” and can lead to fish kills, water quality problems, and other harmful ecological and economic impacts. Astitha explains that these eutrophication events are expected to increase with climate change, and that models like this will become more important for monitoring and forecasting purposes.

The researchers focused their study on the central basin of Lake Erie, which has experienced seasonal algal blooms and eutrophication events for decades. The lake’s proximity to large agricultural areas where fertilizers are used and to metropolitan areas where air pollution is a problem presents a unique set of challenges that the team wanted to investigate.

Because millions of people rely on Lake Erie for their water supply, modeling was and remains critical to monitoring water quality, Astitha says.

“Currently, forecast models produce daily forecasts, which is particularly important for people in these areas because they are large population centers. Water is not just for recreation; people use it in their daily lives.”

However, Astitha says no single model can account for all the variables that affect water quality. To solve this problem, they started building machine learning models to integrate data from different sources and train machine learning algorithms with observations in the lake.

Astitha says her first publication using this method focused on machine learning modeling of chlorophyll a, an indicator of algal biomass and eutrophication, and a second paper used the same methodology but looked at nutrient pollution from rivers and streams. This latest work examines physical and biological processes encapsulated in a physics-based model to understand the dynamic processes associated with eutrophication events.

Astitha says they have to start from scratch in building models for each process studied, but it is necessary to evaluate the various physical, biological, weather and human processes that affect eutrophication.

Chang explains that eutrophication processes begin in the spring, when fertilizers on agricultural land and subsequent rainfall can wash nutrients into the lake. In the summer, Lake Erie’s waters form three layers: a warmer layer closer to the surface called the epilimnion, an intermediate layer that experiences the most drastic water temperature fluctuations called the metalimnion, and a deeper, cooler layer called the hypolimnion. The metalimnion layer hosts the thermocline, where temperature changes abruptly. In the summer, during stratification, there is little to no mixing between the epilimnion and hypolimnion layers, meaning the deepest water becomes increasingly oxygen-poor as the summer progresses.

The central basin of the lake is prone to the most severe hypoxic events. To study these events and understand their causes, Astitha says the model was designed to predict dissolved oxygen (DO), which is an indicator of hypoxia in the water, and apparent oxygen utilization (AOU), which is an indicator of biological activity in the aquatic ecosystem. To train the model, they used 15 years of data collected between 2002 and 2017.

The results were good, says Astitha, and the model accurately predicted the observed DO and AOU conditions. The model also found that thermal stratification, or the separate temperature layers in the water column, was the most influential variable driving eutrophication in their study area.

“It was a good proof of concept because there are only a few data points in the lake,” says Astitha. “Ideally, any model would need more comprehensive lake coverage, which is not available. With the point observations we have, that is not feasible. Nevertheless, the model worked very well.”

Models like this are becoming increasingly important for monitoring water quality and supporting decision-making in the face of advancing climate change. Astitha said they expect conditions such as temperature increases to increase stratification, while extreme precipitation events caused by climate change may increase the amount of nutrients entering the lake.

“What we see in hypoxia is that in this natural system, nitrogen and phosphorus are already present, but when hundreds of hectares of land are fertilized, some of this fertilizer seeps into the water. This depends on the mixing or stratification of the lake, and weather conditions affect this. Conceptually, we assume that climate change will make the situation worse, and we can now use the model to play out hypothetical future scenarios under the conditions of climate simulations.”

Astitha says future research will include applying the method to other freshwater or marine ecosystems and conducting a more thorough analysis using different climate change projection data to examine the impacts of climate change scenarios on the water quality of these systems.

“From my perspective, we wanted to develop a tool that complements the models that already do these important predictions and monitoring. In the age of machine learning and artificial intelligence, we are trying to bring that element in and see how helpful it is, which is what motivated me to start and continue this work.”