Harmful algal blooms (HABs) are a serious threat to ecosystems and human health. The accurate prediction of HABs is crucial for their proactive preparation and management. While mechanism-based numerical modeling, such as the Environmental Fluid Dynamics Code (EFDC), has been widely used in the past, the recent development of machine learning technology with data-based processing capabilities has opened up new possibilities for HABs prediction. In this study, we developed and evaluated two types of machine learning-based models for HABs prediction: Gradient Boosting models (XGBoost, LightGBM, CatBoost) and attention-based CNN-LSTM models. We used Bayesian optimization techniques for hyperparameter tuning, and applied bagging and stacking ensemble techniques to obtain the final prediction results. The final prediction result was derived by applying the optimal hyperparameter and bagging and stacking ensemble techniques, and the applicability of prediction to HABs was evaluated. When predicting HABs with an ensemble technique, it is judged that the overall prediction performance can be improved by complementing the advantages of each model and averaging errors such as overfitting of individual models. Our study highlights the potential of machine learning-based models for HABs prediction and emphasizes the need to incorporate the latest technology into this important field.

River water quality is one of the main challenges that societies face during the 21st century. Accurate and reliable real-time prediction of water quality is an effective adaptation measure to counteract water quality issues such as accidental spill and harmful algae blooms. To improve accuracy and skill of water quality forecasts along the Yeongsan River in South Korea three different ensemble data assimilation (DA) methods have been investigated: the traditional Ensemble Kalman Filter (EnKF) and two related algorithms (Dud-EnKF and EnKF-GS) that offer either possibilities to improve initial conditions for non-linear models or reduce computation time (important for real-time forecasting) by using a (smaller) time-lagged ensemble to estimate the Kalman gain. Twin experiments, assimilating synthetic observations of three algae species and phosphate concentrations, with relatively small ensemble sizes showed that all three DA methods improved forecast accuracy and skill with only subtle difference between the methods. They all improved the model accuracy at downstream locations with very similar performances but due to spurious correlation, the accuracy at upstream locations was somewhat deteriorated. The experiments also showed no clear trend of improvement by increasing the ensemble size from 8 to 64. The real world experiments, assimilating real observations of three algae species and phosphate concentrations, showed that less improvement was achieved compared to the twin experiments. Further improvement of the model accuracy may be achieved with different state variable definitions, use of different perturbation and error modelling settings and/or better calibration of the deterministic water quality model.

Hyperspectral imagery (HSI) provides substantial information on optical features of water bodies that is usually applicable to water quality monitoring. However, it generates considerable uncertainties in assessments of spatial and temporal variation in water quality. Thus, this study explored the influence of different optical methods on the spatial distribution and concentration of phycocyanin (PC), chlorophyll-a (Chl-a), and total suspended solids (TSSs) and evaluated the dependence of algal distribution on flow velocity. Four ground-based and airborne monitoring campaigns were conducted to measure water surface reflectance. The actual concentrations of PC, Chl-a, and TSSs were also determined, while four bio-optical algorithms were calibrated to estimate the PC and Chl-a concentrations. Artificial neural network atmospheric correction achieved Nash-Sutcliffe Efficiency (NSE) values of 0.80 and 0.76 for the training and validation steps, respectively. Moderate resolution atmospheric transmission 6 (MODTRAN 6) showed an NSE value >0.8; whereas, atmospheric and topographic correction 4 (ATCOR 4) yielded a negative NSE value. The MODTRAN 6 correction led to the highest R² values and lowest root mean square error values for all algorithms in terms of PC and Chl-a. The PC:Chl-a distribution generated using HSI proved to be negatively dependent on flow velocity (p-value = 0.003) and successfully indicated cyanobacteria risk regions in the study area.