PREDICTING FLOODPLAIN DISSOLVED OXYGEN WITH A RECURRENT NEURAL NETWORK

Dissolved Oxygen (DO) concentrations in waters overlying floodplains regulate soil reduction-oxidation processes, and therefore have a significant impact on the internal cycling of nutrients between floodplain soils and floodwaters. As DO concentrations decrease, conditions become more favorable for the internal release of iron-bound phosphorus (P) from floodplain soils, which has important water quality implications. DO dynamics in floodplains are a function of complex, non-linear processes affected by floodplain soil chemistry and biology, incoming riverine water quality and quantity, and floodplain hydrology. Advancements in sensor technology have allowed for near-real time monitoring of DO concentrations, but sensors are spatially limited and expensive, taking significant time and resources to maintain. Therefore, accurate modeling of DO concentrations is of interest to researchers, restoration professionals, water quality managers, and other stakeholders working to protect water quality and restore and/or preserve floodplain habitat. Due to the complex, non-linear behavior of DO concentrations in floodplains, traditional statistical methods struggle to accurately predict DO concentrations. In this work, we illustrate the development and testing of a recurrent neural network (RNN) model to predict DO concentrations from continuous sensor data from two floodplains located in the Lake Champlain Basin of Vermont during the Spring of 2022. Our findings demonstrate that RNNs show great promise for predicting DO concentrations up to seven days into the future, based solely on water level, water temperature, and event inundation time as model inputs. A model capable of accurately predicting DO concentrations has important implications for modeling other complex biogeochemical processes in floodplains, such as internal release of sediment-bound P, that could help us to better understand whether floodplains will act as a source or sink of P for given storm events.