Deep learning-based prediction of exceeding the criteria for river chlorophyll a concentrations using high-frequency data from a sensor network

Sensor networks enable the collection of high-frequency, large water quality datasets that provide valuable information for managing eutrophication, such as chlorophyll a (Chl-a) concentration. Deep learning models have been successfully applied to derive useful insights from large-scale environmental data. However, sensor data often contain missing values, presenting challenges for applying deep learning models. Therefore, we employed the reverse time attention model with a decay mechanism (RETAIN-D) to simultaneously conduct feature engineering, prediction, and interpretation within a single model structure. Various environmental, hydrological, and meteorological variables were utilized as input features to predict the exceedance of Chl-a criteria. Data were collected from 2018 to 2022 at four monitoring sites along the Geum River, South Korea. RETAIN-D demonstrated strong prediction performance (accuracy = 0.84–0.90, AUC = 0.69–0.91, F-measure = 0.89–0.90 on the test set) across varying Chl-a criteria. Environmental variables were more important than hydrological and meteorological for predicting the exceedance of Chl-a criteria. The contribution of input features to the model prediction was generally higher in more recent time steps when the Chl-a criterion of the target site was applied. These results highlight the effectiveness of RETAIN-D in analyzing high-frequency time series data from sensor networks.