Harmful algal blooms (HABs) are a critical issue in water quality and resource management. The accelerating climate crisis and anthropogenic eutrophication have led to a global increase in the frequency, duration, and magnitude of HABs. Therefore, pro...
Harmful algal blooms (HABs) are a critical issue in water quality and resource management. The accelerating climate crisis and anthropogenic eutrophication have led to a global increase in the frequency, duration, and magnitude of HABs. Therefore, proactive management is necessary to minimize socio-economic consequences associated with HABs, and to sustainably utilize water resources. A predictive model that can quantify the relationship between cyanobacteria abundance and environmental factors, and accurately forecast HABs prior to their occurrence, can be a useful tool for timely implementation of proactive measures.
Deep learning approaches show great applicability in forecasting HABs, which are interactively influenced by numerous environmental drivers. However, further efforts are required to improve limitations that hamper the benefits of adopting deep learning approaches as decision-support tools for proactive HAB management. In this study, hybrid deep learning models were developed to address the lack of explainability in black-box deep learning models, the limited temporal resolution due to insufficient monitoring frequency, and the difficulty in expanding spatial resolution due to varying monitoring schemes among different sites. The usefulness of the developed hybrid models was demonstrated by their applications at the Gangjeong·Goryeong weir section in the Nakdong River, South Korea, where HABs management is necessary for safe water supply.
Explainability is crucial to gain more insights for decision making in HABs management. In this study, reverse time attention (RETAIN) was adopted to improve explainability of HAB forecasts. RETAIN was applied to one-week forecasts of HABs for the study site using weekly basis data. Input features encompassing water quality, meteorological, and hydrological factors were used to forecast cyanobacteria cell counts. In addition, long short-term memory (LSTM) and gated recurrent unit (GRU) were adopted for comparisons of predictability with RETAIN.
RETAIN yielded a high degree of prediction accuracy with test R2 of 0.79, NSE of 0.78, RMSE of 1.93, and MAE of 1.41 on natural log scales. Despite similar predictability to LSTM and GRU, RETAIN could provide relative importance of environmental factors at both time- and feature-levels without the need for post-hoc explanation techniques. Cyanobacteria abundance, water temperature, and total nitrogen from the past week were most important features in forecasting HABs. Increases in mean contributions with increasing water temperature and weekly residence time were pronounced in summer than other seasons. Furthermore, the temporal patterns of mean contributions and their corresponding individual forecasts were highly correlated. The study results demonstrate the utility of RETAIN in terms of both predictability and explainability.
Insufficient monitoring frequency hindered the applicability of deep learning approach in forecasting HABs. To improve temporal resolution of HAB forecasts, this study developed a novel hybrid deep learning models by combining multiple feature engineering and attention mechanism. The hybrid deep learning models were applied to HABs forecast on a daily scale at the study site with varying forecast horizons (i.e., 7, 14, and 28 days). Input feature pertaining to meteorological, upstream and tributary water quality, and hydrological factors used to forecast cyanobacteria cell counts. Regardless of feature engineering and attention mechanisms, 7-day forecasts (NSE = 0.64–0.73, RMSE = 2.11–2.46; on natural log scale) were more accurate than 14-day (NSE = 0.55~0.62, RMSE = 2.51~2.72; on natural log scale) and 28-day (NSE = 0.48–0.57, RMSE = 2.68–2.94; on natural log scale). The prediction accuracy of hybrid models was higher when both feature engineering and attention mechanism explored the input sequence in the same direction.
Despite similar performance exhibited by RETAIN with backward recurrent imputation of time series (RETAIN–RITS_B) and dual-stage attention based recurrent neural network with forward RITS, explanations derived by former had clearer distinctions in the relative importance of different input features and time steps. The difference in explainability is consistent for 7-day forecasts in which both models yielded highest prediction accuracy. Therefore, RETAIN–RITS_B exhibited great suitability in supporting proactive HAB management. Moreover, the comprehensive investigation of the various mechanisms based on real-world application provided in this study can serve as a guideline for wide range of domains that could benefit from forecasting with enhanced temporal resolution.
Multi-site forecasts of HABs for upstream to downstream sections of river can effectively provide a useful basis for bloom management. In this study, the segment-wise node embedding mechanism was proposed to improve spatial resolution of HAB forecasts while reflecting hierarchical structure of river. Here, mechanisms from previously developed models were used as modules of the proposed embedding mechanism to overcome different monitoring schemes among sites. Consequently, node embedding network structure that encompassing water quality, algae monitoring sites and waste water treatment plant (WWTP) in study area were constructed. The node embedding network model was applied to 4-day forecasts of HABs for two sites in study area.
The model exhibited high prediction performance for forecast target sites with R2 of 0.64–0.71, NSE of 0.63–0.69, RMSE of 2.23–2.26, and MSE of 1.38~1.62 for test data on natural log scales. The node embedding network model can provide the importance of input features from linked nodes. The features linked from nearby nodes for segment consistently exhibited higher importance for all forecasts target sites. These results indicate that proposed embedding mechanism successfully distinguish the hierarchical structure of the upstream and downstream segment. Furthermore, the node embedding network model could provide simulations of cyanobacteria abundance on a daily scale according to the point source and weir operation scenarios based on network structure. The results of scenario analysis suggested that reduction of effluent total nitrogen concentration of WWTP and combined operation of upstream and downstream weirs were effective in managing HABs in the study area. These results demonstrate the remarkable usefulness of segment-wise node embedding mechanism in HAB management. Moreover, the proposed node embedding mechanism is applicable to various water bodies that require integrated analysis reflecting spatial characteristics.
The hybrid deep learning models developed in this study successfully addressed the limitations of current deep learning-based HAB forecasts. I expect that proposed hybrid deep learning models would be applicable to wide variety of water quality and resource management domains that can benefit from improved explainability, temporal resolution, and spatial resolution.