Unraveling Temporal Dynamics of Lake Water Quality with Modular Compositional Learning

Water quality is integral to lakes’ ecosystem services but very challenging to predict due to the complexity of lakes at multiple spatial and temporal scales. Lake metabolism, the balance of primary production and respiration within the ecosystem, is a critical ecosystem-scale process driving essential water quality metrics, such as oxygen availability, organic matter turnover, and water clarity. Additional fluxes of oxygen and organic matter including atmospheric exchange and matter transport are also included in quantifying lake metabolism. Due to the modular design of process-based metabolism models, they are well suited to modular compositional learning (MCL), a novel framework within the Ecology Knowledge-Guided Machine Learning paradigm. In the MCL methodology, uncertain process descriptions in process-based modules can be surrogated by machine learning modules to improve overall performance. By comparing the performance of different combinations of process-based and machine learning modules, we can gain a greater understanding of how well each module represents its associated ecosystem process. We have created an MCL model coupling lake physics and metabolism modules to model dynamics of dissolved oxygen, dissolved organic carbon, and particulate organic carbon along the water column at hourly intervals over the course of 5 years within Lake Mendota (Wisconsin, USA). The metabolism modules include atmospheric gas exchange, net primary production, and ecosystem respiration, with transport of organic matter within the water column as an additional module. Physics modules are derived from previous work modeling lake thermal structure. Different combinations of process-based and machine learning models were created by swapping out modules and their performance was compared using observed water quality parameters as target variables. In our previous work using this approach to simulate lake physics only, replacing the process-based diffusive transport module with a machine learning module in a hybrid MCL framework improved predictions of lake thermal structure relative to purely process-based and purely machine learning approaches. For this new work on lake metabolism, preliminary results indicate that biological processes, production and respiration, are areas of potential advancement of understanding via machine learning.