In 1998, South Korea implemented an algae alert system to ensure the continuous supply of clean and safe tap water and protect the public from the risks of algal toxins that may occur during waterside recreational activities. This system, established under the Water Environment Conservation Act, involves regular monitoring of cyanobacterial density (cells/mL), which serves as an indicator of HAB occurrences. HAB alerts are issued when the cyanobacterial density exceeds predetermined threshold values: alerts and warnings are issued at threshold values of 1,000 and 10,000 cells/mL, respectively, for areas with drinking water sources; and at 20,000 and 100,000 cells/mL, respectively, for areas where water is used for recreational activities; an outbreak is declared at a value of 1,000,000 cells/mL for areas with drinking water sources.

In this study, four algal alert sites within the water source areas along the mainstream Nakdong River (Fig. 1), which records the highest number of HAB alert issuance days in South Korea, were selected. The study sites Haepyeong (HP), Gangjeong–Goryeong (GG), Chilseo (CS), and Mulgeum–Maeri (MM) are located from upstream to downstream and serve as drinking water sources. Their approximate distances from weirs, which were constructed as a part of the Four Major Rivers Restoration Project, are as follows: HP is located 6 km downstream of the Gumi Weir; GG is 8 km upstream of the Gangjeong–Goryeong Weir; CS is 10 km upstream of the Changnyeong–Haman Weir; and MM is 42 km downstream of the Changnyeong–Haman Weir.

Data on cyanobacterial density and water quality were sourced from the Water Environment Information System (http://water.nier.go.kr). Weekly measurements of cyanobacterial density (cells/mL) for each study site were obtained from the algae information section in this system. For instances where measurements exceeded warning-level figures, data were collected twice a week or more frequently. Cyanobacterial counts were determined by observing a water sample under a microscope. The water sample was a composite sample obtained from three points: within 50 cm from the water surface as well as at 1/3 depth and 2/3 depth from the water surface at the maximum depth point of the study sites. When the cyanobacterial density exceeded warning-level figures, the sample was taken as a mixture from these three depths at three distinct points of the river—the middle, left, and right sides. Data were collected in 2016–2023 at the HP, GG, and CS sites and in 2020–2023 at the MM site. Water quality data comprised weekly measurements of biochemical oxygen demand (BOD), Chl-a content, chemical oxygen demand (COD), dissolved oxygen (DO), electrical conductivity (EC), pH, suspended solids (SS), total nitrogen (TN), total phosphorus (TP), and water temperature. These parameters were measured at water quality monitoring stations adjacent to the study sites (HP, GG, CS, and MM).

Water flow rate and velocity data were obtained from the Water Resources Management Information System (http://www.wamis.go.kr). An acoustic Doppler current profiler was used to obtain these data. Using these data, we determined formulas to describe the relationship between water flow rate and velocity, allowing the conversion of daily water flow rate measurements into velocity values for each study site. Daily atmospheric temperature data were collected from Automated Synoptic Observing Systems located adjacent to each study site. These data were provided by the Open MET Data Portal operated by the Korea Meteorological Administration (https://data.kma.go.kr).

Because daily water velocity data were not directly collected at the study sites, they were derived from daily water flow rate measurements. First, records of water flow rates, including corresponding velocities, were obtained from the Water Resources Management Information System. The specific observation stations designated for each study site were as follows: Gumi Bridge (for HP), Seongju Bridge (for GG), Gaenae-ri (for CS), and Samrangjin Bridge (for MM). A regression analysis—either linear or nonlinear (e.g.., polynomial, exponential, or logarithmic)—was conducted to examine the relationship between water flow rate and velocity for the periods before and after the implementation of the Four Major Rivers Restoration Project (i.e., 2001–2010 and 2013–2019, respectively). The regression model with the best goodness-of-fit, based on the coefficient of determination, is shown as a result (Fig. S1). Subsequently, daily water flow rate data were collected from each observation station, and water flow rate–velocity conversion formulas were utilized to estimate the water velocity values at the study sites. Differences in water velocity values before and after the implementation of the Four Major Rivers Restoration Project were statistically evaluated using the Mann–Whitney U test.

This study utilized both the traditional yet effective multiple linear regression (MLR) model and the ensemble-based robust prediction model, random forest (RF), as machine learning methods. These models are widely used among various machine learning algorithms [40, 41] as they offer the advantage of higher performance with smaller datasets compared to artificial neural network models, which require large datasets because of their numerous parameters. These models were trained using data from each study site, and the model with the highest validation accuracy for simulating water velocity changes was selected. The RF model was configured with 100 trees, and two settings of hyperparameters—the number of features to consider when looking for the best split (being “sqrt” and five)—were employed for model training. The Python Scikit-learn library was used for machine learning modeling, with training, validation, prediction, and simulation processes implemented in Python. Data preprocessing, input variable configuration, and modeling were based on the previous methodology reported [14]. The dataset was constructed by setting the log-transformed cyanobacterial counts (log cells/mL) as the dependent variable and various water quality indicators measured on the same date as the independent variables (11 indicators: BOD, Chl-a content, COD, DO, EC, pH, SS, TN, TP, water temperature, and water velocity). Only data from the April–November period, when algal blooms mainly occur, were included. Data with cyanobacterial counts below 100 cells/mL or containing missing values were omitted.

In this study, the leave-one-out cross-validation method was employed for model validation. This is a highly precise validation technique that keeps one data point as the validation set while training the model on the remaining data. The model’s prediction performance is then evaluated using the excluded data point. This process is repeated for each data point, with every point used once for validation. This method yields an overall assessment of model performance across the entire dataset, making it suitable for long-term evaluations. To assess model prediction performance, the correlation coefficient (r) and mean squared error (MSE) between the measured and predicted values were calculated, which reflect the prediction accuracy and error, respectively. These statistical metrics were employed to evaluate how effectively each model captured the actual data—aiming to minimize model error and maximize prediction accuracy.

Future HCBs resulting from climate change were simulated using projected temperature increments for the year 2100 considering the worst-case scenario: Representative Concentration Pathway 8.5 (RCP 8.5). RCP 8.5 assumes that carbon dioxide emissions will continue to rise, reaching a concentration of 936 ppm by 2100, with global temperatures increasing by up to 4.8°C compared to current levels [26]. Monthly air temperature projections for 2091–2100 at each study site (HP, GG, CS, and MM), under RCP 8.5, were obtained from the Climate Information Portal (climate. go.kr/home/) operated by the Korea Meteorological Administration.

Future water temperatures under climate change scenarios were estimated using these projected air temperatures at each study site and an empirical air temperature–water temperature relationship (water temperature (°C) = 0.8806 × air temperature (°C) + 4.3015; R² = 0.901). These projected water temperature values were incorporated into the dataset as an input variable set in the trained machine learning models to predict future HCBs under the RCP 8.5 scenario. Monthly increased water temperatures were applied to the models for each study site. To minimize bias from infrequent data points—specifically, those with water temperatures exceeding the 75th percentile—only data below this threshold were used when adjusting input values for increased water temperatures. This approach ensured more reliable predictions by avoiding extrapolations beyond the range of the training dataset. In other words, the model would not perform well if it were applied to water temperatures exceeding the maximum values seen during training. The 75th percentile was selected to retain a sufficiently large dataset while excluding infrequent data points that exceed 3–4 °C (the projected range of water temperature rise by RCP 8.5) below the maximum observed water temperature. Subsequently, proportional relationships between predicted cyanobacterial densities (according to temperature increments) and the outputs of the trained model (before temperature increases) were derived. Specifically, linear regression without an intercept (i.e., y = ax) was performed for each site using the output values of cyanobacterial abundance (log cells/mL) before (x) and after (y) the water temperature adjustments, based on 75% of the dataset. The resulting relationship was then used to convert all modeled (current) cyanobacterial abundance values (x) into projected future values (y). For comparing future simulated HCBs and the current ones, predicted future cyanobacterial abundances were compared with the outputs of models trained using collected observational data.

To simulate changes in HCBs due to weir operation, datasets comprising independent and dependent variables were used to train the machine learning models (Section 2.3). The model with the highest validation performance was then employed to determine variations in cyanobacterial densities resulting from changes in water velocity (−20%, +20%, +40%, +60%, and an increase resulting from weir demolition). Water velocity under scenarios where the weirs were demolished was estimated by comparing water velocities acquired before and after weir construction at each site. For scenarios involving water velocity variations, model outputs were generated by adjusting the input water velocity values. In particular, for scenarios with decreased water velocity, data wherein the water velocity exceeded the 50^th percentile was used, and for scenarios with increased water velocity, data wherein the water velocity was below the 75^th percentile was used. This approach was adopted because of the limited number of samples with high or low water velocities, which could have led to potential biases in model’s predictive capability. Subsequently, changes in cyanobacterial densities due to water velocity changes were quantitatively identified through a linear relationship. Changes in predicted future cyanobacterial densities based on water velocity changes were then simulated using this linear relationship and predicted cyanobacterial densities under climate change scenarios.

A regression analysis between water flow rate and velocity revealed remarkably high coefficients of determination, indicating the goodness of fit of the water flow–velocity relationship: 0.891 and 0.977 for HP, 0.905 and 0.950 for GG, 0.805 and 0.994 for CS, and 0.952 and 0.999 for MM, for the periods before and after implementing the Four Major Rivers Restoration Project, respectively (Fig. S1). Generally, an analysis based on data obtained after implementing the project showed a more linear water flow rate–velocity relationship with higher coefficients of determination than that based on data obtained before implementing the project, suggesting that weir construction reduced the variability in water flow rate and velocity. These water flow rate–velocity relationships for each site facilitated the conversion of daily water flow rates into water velocities, which were then used as one of the independent variables in subsequent machine learning modeling in the succeeding sections.

Our analysis also indicated a positive skewness in the water velocity distribution after project implementation, in contrast to the more uniform distribution observed before project implementation (Fig. 2). Weir construction under the Four Major Rivers Restoration Project led to a significant reduction in average water velocity by factors of 8.2, 3.7, 3.3, and 5.6 compared to initial values at the HP, GG, CS, and MM sites, respectively (p < 0.001). This finding is consistent with that of a previous study that indicated a significant reduction in discharge after dam construction [42]. Before implementing the project, the average water velocities were as follows: HP (0.467 m/s) > GG (0.435 m/s) > CS (0.356 m/s) > MM (0.270 m/s). Meanwhile, the water velocities recorded after implementing the project were GG (0.116 m/s) > CS (0.108 m/s) > HP (0.057 m/s) > MM (0.048 m/s). The decrease in water velocity at HP was the highest because of its proximity to a nearby weir: the site is located only 5 km away from the weir. Meanwhile, despite being the farthest from a weir (~42 km away), MM experienced a significant decrease in water velocity, the magnitude of which was greater than those for GG and CS. This reduction was attributed to the enhanced capacity of the river owing to an increase in its cross-sectional area following dredging as well as the decreased water discharge at the Nakdong River Estuary Bank (located ~32 km downstream of MM) after weir construction. The mean water discharge dropped from 484 m³/s between 2008 and 2010 to 243 m³/s between 2013 and 2015.

After project implementation, the mean retention times at HP, GG, CS, and MM were estimated to be 5.54, 2.50, 4.60, and 18.25 days, respectively. Water with a hydraulic retention time exceeding 7 days is considered a lake or reservoir [43]. Consequently, some areas of the Nakdong River, including MM, and occasionally HP and CS, can be classified as lakes. Limnologists suggest that sites with retention times of >7 days favor phytoplankton accumulation [44], indicating that water quality management at these sites should be prioritized.

The variability in water velocities before weir construction seemed to be influenced by precipitation levels while water velocities became notably stable after the project implementation, generally regulated to <0.1 m/s. These findings indicated that the overall hydraulic retention time in the Nakdong River increased threefold to eightfold due to weir construction, and this retention time is anticipated to decrease by the same degree upon the demolition of the weirs.

The validation results of the machine learning models predicting cyanobacterial abundance are summarized in Table 1, and the predicted versus actual values from the best-performing model are shown in Fig. S2. Across all sites, the RF model exhibited higher correlation coefficients and lower MSE values than the MLR model, indicating its superior reliability. At CS and MM, the performance gap between the MLR and RF models was relatively narrow, with both models showing a high performance (r > 0.7). Differences in performances due to the adjustment of hyperparameters in the RF model—namely using two different settings (“sqrt” and five)—were found to be minimal. These results are consistent with those of a recent study that also reported an MSE of approximately 0.280 when predicting cyanobacterial abundance using the RF model [45]. The RF model, which achieved the best prediction performance for each site, was used to simulate cyanobacterial abundance in response to changes in temperatures under RCP 8.5 and alterations in water velocity. The significant variation in predictive performance among study sites could be attributed to differences in the distribution of cyanobacterial density values. Site HP, which exhibited the poorest performance, had highly skewed data, resulting in overall underprediction. In contrast, site MM, which showed the best performance, had a dataset resembling a normal distribution (Fig. S2).

Table 2 presents the training performance of the MLR model along with the statistics of the independent variables. Water temperature was shown to be a highly influential factor, corroborating previous studies [46, 47] and supporting the assumption of this study that HCBs are significantly affected by temperature. The MLR model for the MM site, which exhibited the highest training and validation performance, revealed that pH, TP, EC, and TN were statistically significant independent variables (p < 0.05). All these variables, except for TN, were positively correlated with cyanobacterial abundance. pH was found to be strongly associated with algal growth [48] and may increase due to cyanobacterial photosynthesis. TP is a well-established key driver of HCBs, influenced by both external loading and internal cycling processes [49, 50]. High EC values likely indicate the inflow of pollutants [51] and an increase in total dissolved solids [52].

The negative effect of TN on cyanobacterial abundance may stem from seasonal nitrogen dynamics. A correlation coefficient of −0.47 was observed between water temperature and TN at the study sites. This could be attributed to nitrification, denitrification, nitrogen assimilation by cyanobacteria, and seasonal fertilizer application timing [53, 54]. The regression coefficient for water velocity at MM approached significance (p = 0.065), indicating a potential relationship with cyanobacterial abundance, while BOD and SS showed the least significant relationship with cyanobacterial abundance. Additionally, COD was identified as a significant variable at other sites and serves as an indicator of organic compounds in water that can promote cyanobacterial proliferation [55].

Across all sites, except for GG, a negative correlation between water velocity and cyanobacterial abundance was observed. In particular, the MLR model—which demonstrated higher performances at the CS and MM sites—confirmed the negative correlation between water velocity and cyanobacterial abundance. This finding supports the hypothesis that reduced water flow is conducive to cyanobacterial proliferation. Additionally, cyanobacterial abundance decreased with increased water velocity, particularly up to 0.2 m/s at the study sites, after which the rate of decrease diminished when water velocities exceeded 0.2 m/s (Fig. S3). This suggests that the reduction in water velocity due to weir construction has contributed to an overall increase in cyanobacterial abundance.

Annual air temperatures in 2091–2100 under RCP 8.5 were projected to be 16.7 °C, 17.1 °C, 17.3 °C, and 18.1 °C at HP, GG, CS, and MM, respectively, marking an increase from the current temperatures of 12.9 °C, 13.0 °C, 13.4 °C, and 14.4 °C, respectively. The prediction of cyanobacterial abundance under the worst-case climate change scenario made using the RF model indicated a significant increase at all sites along the Nakdong River, particularly downstream and at higher cell densities (Fig. S4). Notably, GG and CS were expected to experience the largest increases in cyanobacterial abundance due to climate change. Currently, GG and CS show cyanobacterial densities of 4,000–15,000 and 10,000–33,000 cells/mL, respectively, which are anticipated to increase to 10,000–40,000 and 32,000–140,000 cells/mL, respectively, in the future. The maximum predicted cyanobacterial abundance can reach nearly 1 million cells/mL at CS and exceed 1 million cells/mL at MM, which is the threshold for “outbreak.”

The predicted distributions of cyanobacterial abundance under future temperature increments and after weir demolition are illustrated in Fig. 3. Given the water velocity differences before and after weir construction, water velocities at the HP, GG, CS, and MM sites were assumed to increase by factors of 8.2, 3.7, 3.3, and 5.6, respectively, if the weirs were demolished or fully opened. At all sites, an increase in cyanobacterial abundance due to rising temperatures under RCP 8.5 in the future (2091–2100) was shown to decrease because of increased water velocity resulting from weir demolition, leading to more positively skewed histograms. However, even with a decrease in cyanobacterial abundance following weir demolition, the overall cyanobacterial abundance was not expected to fall below the current levels, when temperature rise was not considered. In particular, at MM, cyanobacterial abundance reached outbreak conditions (1,000,000 cells/mL) with a low likelihood even with increased water velocity resulting from weir demolition. The likelihood of such an incident at MM was estimated to be equivalent to 2.8, 8.7, and 5.3 weeks of outbreak during 10 years under current, future, and future-with-no-weir scenarios, respectively. Opening the weirs could not only reduce HABs but also may promote algal growth due to increased sunlight irradiation on the water with reduced depth and the spreading of the overall concentration of nutrients in the river [24]. Light availability is another key factor influencing HCBs [56, 57], and related data could be integrated into further studies.

The frequencies of cyanobacterial occurrences exceeding the four thresholds (10,000, 20,000, 100,000, and 200,000 cells/mL), including the “warning” and “outbreak” thresholds for surface water sources and recreational water activity areas, were predicted under various weir operation scenarios (Fig. 4). Under future conditions with increased temperatures (RCP 8.5), a decrease in water velocity hardly affected the occurrence frequency of HCBs likely because water velocity at the study sites was already very low, showing minimal response to velocity decreases. An increase in water velocity reduced the future occurrence frequency of HCBs. As the water velocity gradually increased, the occurrence frequency for all the four thresholds progressively decreased for all sites. Previous studies also reported improved water quality following large-scale water discharges [58, 59]. Moreover, the findings of this study align with earlier evaluations highlighting the significant influence of upstream dam operations on phytoplankton communities and aquatic ecosystems [60]. Dam removal leads to substantial ecological transitions, including changes in sediment transport, channel morphology, and habitat availability. For instance, removals on the Elwha River have demonstrated that sediment dynamics and vegetation patterns can undergo substantial shifts following dam deconstruction [31, 61, 62].

By the year 2100, the annual frequency of events exceeding 10,000 cells/mL threshold for drinking water sources under increased water velocity by 20% relative to the present value showed decreases in the following order: CS (8.8 weeks; a 0.9-week decrease compared with that under RCP 8.5) > MM (7.2 weeks; same as that under RCP 8.5) > GG (6.0 weeks; a 0.2-week decrease compared with that under RCP 8.5) > HP (1.4 weeks; a 0.1-week decrease compared with that under RCP 8.5). When water velocity was increased by 60%, the order was as follows: CS (8.1 weeks; a 1.6-week decrease compared with that under RCP 8.5) > MM (7.0 weeks; a 0.2-week decrease compared with that under RCP 8.5) > GG (5.4 weeks; a 0.8-week decrease compared with that under RCP 8.5) > HP (1.2 weeks; a 0.3-week decrease compared with that under RCP 8.5). Meanwhile, the annual frequency of events exceeding the 100,000 cells/mL threshold for waters used for recreational activities under increased water velocity by 20% showed decreases in the following order: MM (2.7 weeks; a 0.8-week decrease compared with that under RCP 8.5) > CS (2.6 weeks; a 0.9-week decrease compared with that under RCP 8.5) > GG (0.8 weeks; the same as that under RCP 8.5) > HP (0 weeks; same as that under RCP 8.5). Concomitantly, for an increase of 60% in water velocity, the order was as follows: MM (2.5 weeks; a 0.5-week decrease compared with that under RCP 8.5) > CS (1.5 weeks; a 3.0-week decrease compared with that under RCP 8.5) > GG (0.7 weeks; a 0.1-week decrease compared with that under RCP 8.5) > HP (0 weeks; the same as that under RCP 8.5). As the water velocity increased, the occurrence frequency under climate change tended to decrease, approaching the current occurrence frequency levels (without temperature increase); the decrease was particularly slower at the MM site. Given its low water velocity prior to the project implementation and minimal changes in water velocity after weir construction, there was minimal impact of water velocity on cyanobacterial changes at this site possibly because of relatively high temperatures at the site. Thus, it was inferred that sites closer to a weir, such as HP, GG, and CS, effectively offset increased cyanobacterial proliferation and HCB frequency resulting from climate change through dam operation.

This study simulated changes in cyanobacterial abundance following decreases in the hydraulic retention time resulting from the opening or demolition of weirs, providing insights into how effective weir operation can mitigate the intensity and frequency of HCBs. In summary, operating the weirs in the Nakdong River so as to increase the water velocity three to eight times (increased water velocities before weir construction) can reduce cyanobacterial density up to 20%–50%, while increasing the velocity by more than 60% can reduce the frequency of HCBs by more than 20%. The simulation focused on predicting cyanobacterial responses to changes in water velocity under climate change scenarios.

However, this study has several limitations. It assumes that future nutrient levels will remain within current ranges and considers temperature as the sole climate change variable. Additionally, uncertainties may arise from the quality of long-term monitoring data, which could be affected by measurement errors. Future precipitation patterns and water quality changes due to pollution sources, among other factors, must be included for more comprehensive predictions of future HCB patterns resulting from climate change and changes in hydraulic retention time. Trade-offs between ecological responses, such as HABs, and securing water resources using engineered structures must be carefully evaluated before undertaking heavy civil works. A variety of physical and chemical techniques can be applied for algal removal in surface water bodies [63]. However, such methods are often impractical for large-scale, continuously flowing water systems. Prioritizing the management of nutrient pollution sources and implementing strategic adjustments to river hydrology may offer more effective and sustainable approaches to mitigating HCBs.