Machine learning-based prediction of seasonal hypoxia in eutrophic estuary using capacitive potentiometric sensor

Abstract

A hypoxia occurred in eutrophic estuary was predicted using long short-term memory (LSTM) model with prediction time steps (PTSs) of 0, 1, 12, and 24 h. A capacitive potential (CP), which provides quantitative information on dissolved oxygen (DO) concentration, was used as a predictor along with precipitation, tide level, salinity, and water temperature. First, annual changes in DO concentration were clustered in three phases of annual DO trends (oversaturation, depletion, and stable) using k-means clustering. CP was the most influential variable in clustering the DO phases. The LSTM was implemented to predict the DO phases and hypoxia occurrences. In the simultaneous prediction of the depletion phase and hypoxia occurrence with a 12 h PTS, the accuracy was 92.1% using CP along with other variables; it was 3.3% higher than that achieved using variables other than CP. In the case of predicting the depletion phase and hypoxia non-occurrence using CP along with other variables, the accuracy was 61.1%, which was 5.5% higher than that when CP was not used. When using CP along with other variables, the total accuracy was highest for all PTS. Overall, the utilization of CP and machine learning techniques enables accurate predictions of both short-term and long-term hypoxia occurrences, providing us with the opportunity to proactively respond to disasters in aquaculture and environmental management due to hypoxia.