Forecasting Abrupt Depletion of Dissolved Oxygen in Urban Streams Using Discontinuously Measured Hourly Time-Series Data

Depletion of dissolved oxygen (DO) is a major cause of fish kills in urban streams. Although forecasting short-term DO concentrations in streams prior to hypoxic events is necessary, such efforts have been rarely made. In this study, 24-h forecasting models were developed for DO concentrations in three urban streams of South Korea. To forecast the DO concentrations at the outlet sites, which coincide with fish kill hot spot areas, water quality parameters at the lower reaches and hydrometeorological parameters were used as input variables. The monitoring data were measured hourly between 2017 and 2018 and divided into training and test sets at a ratio of 8:2. Tenfold cross validation was performed for hyperparameter optimization. Due to the dynamic characteristics of DO concentrations and the discontinuity in time-series data, a long short-term memory (LSTM) neural network modeling approach was selected. Overall, a high degree of accuracy was recorded for all study streams. Although hypoxic events were forecast with lower accuracy, the timing and magnitude of abrupt DO depletion were well captured. Water temperature and DO concentrations at the lower reaches and 24-h cumulative precipitation were important variables for forecasting DO concentrations at all stream outlets. In particular, the importance of cumulative precipitation across all streams indicated that the effects of nonpoint sources were critical in depleting DO in urban streams. Monitoring of both lower reaches and outlets in conjunction with a variable importance analysis enhanced interpretability of the LSTM model outputs. This study improves our understanding of precursors of hypoxic events in urban streams.