Coupling Bayesian modeling of lake oxygen dynamics with machine learning to advance aquatic ecosystem understanding

Abstract:

During summer stratification, freshwater lakes can experience severe hypolimnetic hypoxia due to a combination of reduced vertical transport of dissolved oxygen (DO) and high rates of water column and sediment oxygen demand. Hypolimnetic hypoxia can severely reduce water quality and reduce habitability for valuable coldwater fish species. Accurate predictions of lake DO at both seasonal and decadal scales are a high priority for lake managers. However, dynamics are challenging to predict because hypolimnetic hypoxia is an ecosystem-scale property that emerges from the interactions between exogenous drivers and internal lake physical and biological processes, and the relative influence of those drivers may be time scale dependent. Recently, machine learning algorithms in combination with information from process models (known as knowledge-guided machine learning, KGML) were successfully applied to substantially improve water temperature predictions and simulations of lake phosphorus cycling. We aim to use the synergistic KGML approach to improve the simulations of long-term DO time-series for 8 intensively monitored lakes in Wisconsin, USA, that belong to a diverse set of catchments and represent different environmental gradients.

We set up a two-layer Bayesian model to simulate dissolved oxygen dynamics as a set of discrete first order linear difference equations that are solved using a Forward Euler approximation. The DO models used meteorological data from NLDAS-2, observed oxygen data from NTL-LTER, and modeled lake water temperatures from calibrated process-based models (GLM-AED2) as driver data. The output of each lake’s DO model was used to guide predictions of a neural network that also used a broad suite of additional driving variables assumed to be relevant to lake oxygen dynamics. Using this approach, we were able to identify additional catchment-specific drivers and properties not included in the DO model that are important for understanding long-term oxygen depletion dynamics. These additional variables helped differentiate control of lake oxygen dynamics in lakes with varying morphometries, trophic states and catchments with contrasting land use.