The analysis of spatiotemporal effects of environmental factors on harmful algal blooms in a bloom-prone river using partial least squares structural equation modeling Open Access

Rivers are a major water source and living space for humans. However, the river environment is disturbed by various factors, such as pollution sources, river structures, weather, and their interactions (Yi et al. 2018; Sin & Lee 2020; Park et al. 2021; Liao et al. 2023; Shih et al. 2022). Harmful algal blooms (HABs), which involve the excessive proliferation of phytoplankton, are a representative condition of river ecosystem disturbance (Su et al. 2022). They have negative impacts on the environment, economy, and human health. HABs significantly reduce light penetration into the water, inhibiting photosynthesis. Furthermore, the decomposition of HABs biomass leads to substantial oxygen depletion, potentially causing hypoxic conditions in affected water bodies (Liu et al. 2022a, b; Sun et al. 2022). In addition, HABs disrupt the stable utilization of water resources, making it difficult to operate water purification plants and restricting recreational activities (Huisman et al. 2018). HABs also release various toxic substances, such as hepatotoxins and neurotoxins, posing a threat to human health (Kubickova et al. 2019; Chia et al. 2018).

Recently, the occurrence of HABs has been increasing due to changes in environmental conditions resulting from climate change, urbanization, and other factors. This necessitates the preparation of countermeasures (Lee et al. 2021). Climate change has increased water temperatures and created favorable conditions for algal growth. Moreover, it has increased rainfall, leading to a higher influx of nutrients that are essential for algal growth (Ho et al. 2020). Additionally, land management practices, such as agricultural and urban development, have created an environment where nutrients can easily run off into water systems (Weber et al. 2020; Fernandes et al. 2019). HABs have increased due to both natural and anthropogenic factors, and as a result, they are expected to cause significant damage. Therefore, it is essential to thoroughly analyze the key environmental factors that influence HABs in order to mitigate the damage they cause.

Numerous studies have analyzed the effects of environmental factors on HABs using various statistical techniques (Kozak et al. 2019; Ho et al. 2020; Park et al. 2021). Kozak et al. (2019) analyzed the response of cyanobacteria species to environmental factors using canonical correspondence analysis and evaluated the effects of habitat and catchment condition based on the Mann–Whitney test. To explore both the direct and indirect effects of rainfall on algal blooms, Ho et al. (2020) developed multiple linear regression models for HABs and nutrients separately. Park et al. (2021) used Spearman's rho correlation and redundancy analysis to analyze the key environmental factors affecting the spatiotemporal variability of harmful cyanobacterial blooms dominated by Microcystis spp. and Aphanizomenon spp. Previous studies have identified the relationship between environmental factors and indicators of HABs using statistical techniques. However, research considering the spatiotemporal characteristics and interspecies relationships of HABs, which could have additional impacts on these relationships, has been rare. While studies such as Ho et al. (2020) and Park et al. (2021) have made notable contributions in analyzing the spatiotemporal characteristics and interspecies relationships of HABs, they were constrained by the inability to integrate these complex interactions within a unified statistical analysis. Structural equation modeling (SEM) is a multivariate statistical technique that integrates various statistical analysis methods, such as factor analysis, multiple regression analysis, and path analysis, to understand and estimate cause-and-effect relationships between factors (Hair et al. 2021). Partial least squares structural equation modeling (PLS-SEM), a specific type of SEM, exhibits greater versatility in relationship analysis as it is less sensitive to conditions like normal distribution and sample size (Jia et al. 2020; Duan et al. 2022; Li et al. 2021; Wang et al. 2023; Na et al. 2024; Shi et al. 2024).

Using various monitoring systems within the study site, HAB factors, water quality variables, hydrological variables, and meteorological variables were collected spanning the years 2013–2022 (Supplementary Table S1). HAB factors included cell count (cell/mL) of four cyanobacteria genera: Microcystis (Mic), Anabaena (Ana), Oscillatoria (Osc), and Aphanizomenon (Aph), as well as the total cyanobacterial cell counts, which represented the sum of the four genera. The data for HAB factors were obtained from the station Dasa (DS) (https://water.nier.go.kr/), which served as the representative monitoring station of the GG weir. Additionally, the water quality factors included total phosphorus (TP, mg/L) and total nitrogen (TN, mg/L). The data for water quality were obtained from the four stations: Baekcheon (BC; 35°52′45.20″, 128°21′57.40″) and Habincheon (HB; 35°51′38.16″, 128°24′17.28″), located in the two tributaries, and Seongju (SJ; 35°52′42.63″, 128°23′28.26″), located upstream of the GG weir and DS station (https://water.nier.go.kr/). The meteorological factor, daily average air temperature (Atemp, °C), was obtained from the nearest automated surface observing system station (https://data.kma.go.kr/). The hydrological factor, retention time (RT, day), was calculated based on the total discharge and storage data obtained from the GG weir station (https://water.or.kr/).

Each monitoring data with a different time resolution was converted to a consistent weekly time resolution, which is the lowest time resolution. The time resolution was standardized to weekly intervals, aligning with the week in which HAB factors were measured, by utilizing the weekly mean values of each variable. To consider the temporal lag in the influence of environmental factors on HAB factors, the data related to water quality, meteorological factors, and hydrological factors were merged 1, 2, and 3 weeks prior to the week in which HAB factors were measured. Data from 4 weeks prior were not selected as a variable because its correlation coefficients were lower than those of weeks 1–3, and the average RT at GG weir is approximately 22 days (Table 1, Supplementary Table S2). Any missing values within the merged dataset were imputed by Kalman filtering. Although PLS-SEM has a tolerance for distributional assumptions, log normalization was applied to variables that did not conform to the normal distribution to alleviate any potential impact arising from the distribution.

PLS-SEM comprises two fundamental components: the measurement model, also known as the outer model, and the structural model, also referred to as the inner model (Zhao et al. 2023). The measurement model focuses on the relationship between the latent variables and the indicator variables, which contain the raw data. It is important to note that the measurement model can be of two types: reflective or formative. In a reflective measurement model, the latent variables explain the indicator variables. On the other hand, in a formative measurement model, the indicator variables explain the latent variables.

Moving on to the structural model, its purpose is to explore the relationship between the latent variables. In this model, the exogenous latent variable represents the influencing factor, while the endogenous latent variables represent the affected factors. The relationships between these latent variables are defined by path coefficients. In our study, we used path coefficients to compare and analyze the influence of environmental factors on HABs.

The appropriateness of the PLS-SEM model is evaluated using different criteria for reflective measurement models, formative measurement models, and structural models (Table 1). For the reflective measurement model, we evaluated ‘internal consistency reliability’, ‘convergent validity’, and ‘discriminant validity’. Internal consistency reliability measures the reliability through consistency among the indicator variables, with Cronbach's α serving as the evaluation index. A Cronbach's α value of 0.6 or higher indicates satisfactory internal consistency reliability. Convergent validity is evaluated to determine the degree of correlation between indicator variables connected to the same latent variable. We use the ‘average variance extracted’ (AVE) as an evaluation index. If the AVE value for a latent variable is 0.5 or higher, the latent variable satisfies convergent validity. Discriminant validity evaluates the degree of correlation between indicator variables associated with different latent variables. We use the ‘heterotrait-monotrait ratio’ (HTMT) as an evaluation index. As the most conservative criterion, we evaluate the measurement model satisfying discriminant validity when the HTMT value between each pair of latent variables is less than 0.85.

For the formative measurement model, we evaluate the collinearity and reliability of the indicator variables. In this study, the ‘variance inflation factor’ (VIF) is used as a collinearity evaluation index, and VIF values less than 10 indicate the absence of collinearity. Reliability is evaluated by examining the significance level at 0.05 after applying a bootstrapping procedure to the relationship between the indicator variables and the latent variable.

In the structural model, we evaluate collinearity, R², and f². Collinearity is evaluated in the same way as the formative measurement model. The R² value, the ratio of variance in the endogenous latent variable explained by the linked latent variable, is evaluated as weak (0.25), moderate (0.5), or large (0.75) explanatory power. The contribution of the connected latent variables to the R² value of the endogenous latent variable, f², is evaluated as a small (0.02), medium (0.15), or large (0.35) effect size. In this study, PLS-SEM is implemented using the ‘seminr 2.3.2’ package in R 4.2.3.

Each water quality monitoring station represented distinct data characteristics (Table 2). Stations SJ and DS, located in the mainstream, represented similar average values for TP, which were lower than the average value at stations BC and HB, located in tributaries. This pattern seems to result from the influx of large amounts of nutrients from agricultural lands and industrial wastewater near BC and HB (Jung et al. 2020). In contrast, for TN, the difference between tributary stations (BC and HB) and mainstream stations (SJ and DS) was not distinct. Atemp and RT variables have higher variance than water quality variables. During the warm season, TP and Atemp showed higher values compared with other seasons, while TN and RT exhibited lower values. RT is inferred to be relatively low due to the discharge of reservoir water, resulting from concentrated precipitation during the warm season. Following the Shapiro–Wilk test, it was determined that none of the environmental variables and HABs variables conformed to a normal distribution. Therefore, log normalization was applied to all variables, except for Atemp, which contained negative values.

Table 2

Evaluation index of PLS-SEM

.	Evaluation target .	Evaluation index .	Index criteria (satisfaction) .
Reflective measurement model	Internal consistency reliability	Cronbach's	Above 0.6
	Convergent validity	AVE	Above 0.5
	Discriminant validity	HTMT	Below 0.85
Formative measurement model	Collinearity	VIF	Above 5
	Significance	t-value	Above 1.96
Structural model	Collinearity	VIF	Above 10
	Explanatory power	R2	–
		f2	–

.	Evaluation target .	Evaluation index .	Index criteria (satisfaction) .
Reflective measurement model	Internal consistency reliability	Cronbach's	Above 0.6
	Convergent validity	AVE	Above 0.5
	Discriminant validity	HTMT	Below 0.85
Formative measurement model	Collinearity	VIF	Above 5
	Significance	t-value	Above 1.96
Structural model	Collinearity	VIF	Above 10
	Explanatory power	R2	–
		f2	–

View Large

The appropriateness of both the measurement model and the structural model was evaluated for three PLS-SEM models with different structures. In all three models, both the reflective and formative measurement models satisfied the criteria of Cronbach's, AVE, HTMT, VIF, and t-statistics (Tables 3–5). However, unlike the measurement models, the structural model showed different results for each PLS-SEM. In M1, all latent variables satisfied the VIF criteria, and the R² for HABs latent variable was 0.584 (Figure 3). According to the f² values, the TN and Atemp latent variables had moderate effect sizes on R², with values of 0.295 and 0.179, respectively. Similarly, all latent variables in M2 satisfied the VIF criteria, and the R² for HABs latent variable was 0.553 (Figure 4). Based on the f² value, both Atemp and DS latent variables had effect sizes between small and medium, with values of 0.088 and 0.089, respectively. In M3, exogenous latent variables and the remaining three endogenous latent variables for each of the four cyanobacteria genus latent variables satisfied the VIF criteria. Among the four cyanobacteria genera, the highest R² value was observed for Mic at 0.637, and the effect size of the latent variables connected to Mic, including TN, Atemp, TP latent variables, was relatively high. The R² values for other harmful cyanobacteria genera followed the order of Ana, Aph, and Osc (0.402, 0.284, and 0.07) (Figure 5). Among the effect sizes of the latent variables connected to Ana, the Aph, Atemp, and Osc latent variables were relatively high. In the case of Aph, the effect size of Ana, Mic, and TN latent variables was relatively high, and in the case of Osc, only Ana latent variable had an effect size above the criteria for a small effect size.

Table 3

Evaluation of measurement model and structural model on M1

Latent variable .	Reflective measurement model .					Structural model .
	Cronbach's .	AVE .	HTMT .	VIF (max) .	t-value (min) .	VIF .	f2 .
TP	–	–	–	1.412	2.926	1.588	0.001
TN	–	–	–	7.960	1.976	1.596	0.295
Atemp	0.983	0.968	0.674	–	–	2.312	0.179
RT	0.851	0.771	0.609	–	–	1.441	0.001

Latent variable .	Reflective measurement model .					Structural model .
	Cronbach's .	AVE .	HTMT .	VIF (max) .	t-value (min) .	VIF .	f2 .
TP	–	–	–	1.412	2.926	1.588	0.001
TN	–	–	–	7.960	1.976	1.596	0.295
Atemp	0.983	0.968	0.674	–	–	2.312	0.179
RT	0.851	0.771	0.609	–	–	1.441	0.001

View Large

In M1, the TN latent variable had a significant negative relationship (−0.443, t-statistics = −13.055) with HABs latent variable, while the Atemp latent variable had a significant positive relationship (0.415, t-statistics = 10.18) with HABs latent variable. RT and TP latent variables had positive relationships with HABs latent variable, but these were not statistically significant. The concentration of TN is high in winter and low in summer, but HABs have mainly occurred in summer in the Nakdong River, so TN has a negative relationship with HABs (Park et al. 2021). All latent variables exhibited indicator variables with a statistically significant positive relationship. In particular, TN and TP latent variables, representing water quality factors, had the highest outer weights (0.496, 0.542) with the indicator variables from 1 week prior to the HABs latent variable. On the other hand, the meteorological factor Atemp and the hydrological factor RT latent variable had the highest outer loading (0.990, 0.918) with the indicator variables from 2 weeks prior to the HABs latent variable, but the difference between each outer loading was slight.

In M2, the DS latent variable had a significant negative relationship (−0.443, t-statistics = 6.642) with the HABs latent variable, whereas the Atemp latent variable had a significant positive relationship (0.343, t-statistics = 6.615) with the HABs latent variable. The RT latent variable and other water quality latent variables (SJ, BC, and HB) did not have a statistically significant positive relationship. However, SJ, BC, and HB, which flow into DS, had a significant positive relationship with DS. This means that the relationship between water quality variables was less affected by the distance between the water quality stations and the algal alert system station. However, in the case of HABs, which are influenced by various environmental factors, the relationship between each water quality latent variable and HABs appears to be greatly affected by the distance. For SJ, BC, HB, and DS, the TP indicator variable had a negative weight, while the TN indicator variable had a positive weight at each station. This appears to be because, unlike TN, the seasonal variation of TP concentration, which is high in summer and low in winter, is similar to that of HABs (Park et al. 2021). Overall, TN prevailed over TP in terms of the absolute weight for water quality latent variables at all stations.

In M3, the Atemp latent variable had a significant positive relationship (0.322, t-statistics = 3.418; 0.242, t-statistics = 2.133) with Mic and Ana latent variables. The TN latent variable had a significant negative relationship (−0.300, t-statistics = −3.581; −0.232, t-statistics = −2.152) with Mic and Aph latent variables. The TP latent variable had a significant positive relationship (0.224, t-statistics = 3.618) with Mic latent variable. In contrast, the RT latent variable exhibited a statistically insignificant relationship with any of the latent variables related to HABs. This result may have arisen in part from the confounding effects of the meteorological factor. Due to the seasonality of these variables (Table 1), the effect of Atemp on HABs may decrease the effect of RT in the M3. The relationships between the Mic, Ana, Aph, and Osc latent variables were mostly insignificant. Only the relationships between the Ana and Aph latent variables were statistically significant. These results are caused by the differences in the optimal growth water temperature of each cyanobacterial species. The optimal growth temperature for Anabaena is about 20–25 °C, and Aphanizomenon has a similar optimal growth temperature of around 20 °C. In contrast, Microcystis has a higher optimal growth temperature of 25–35 °C (Park et al. 2021).

To ensure the health of aquatic ecosystems and the sustainable use of water resources, the management of weir-constructed river sections is important. It is essential to prepare management strategies for HABs that cause various negative effects. For this purpose, various environmental factors affecting HABs need to be analyzed from multiple perspectives. Therefore, we employed PLS-SEM to quantify the causal relationship between HABs and environmental factors and verified three hypotheses.

The study results showed that due to the environmental characteristics of the Nakdong River basin, TN had a significant negative relationship with HABs. The water quality latent variables, such as TN and TP, were significantly influenced by the indicator variables taken 1 week prior to the HABs data. Thus, since we confirmed that environmental variable data with the shortest time lag had the strongest relationship, during periods when algae are dominant, monitoring should be conducted more frequently than weekly intervals to prepare for HABs in response to environmental changes. When constructing the PLS-SEM by imitating the structure of the river, the water quality latent variables of the upstream had a significant effect on those of the downstream but did not significantly affect the downstream HABs. Therefore, it is determined that considering environmental variables not only from bloom-prone regions but also from the upstream of the mainstream and inflowing tributaries will help in responding to HABs.

The analysis of the relationship between the dominant species of HABs and environmental variables revealed that the Ana and Aph latent variables significantly influenced each other. However, not all relationships in the model were significant. This is a result of PLS-SEM's inability to identify nonlinear relationships between environmental variables and HABs variables. While deep learning-based algorithms have recently emerged to consider nonlinear relationships, they require large amounts of data and have black-box characteristics that make it difficult to intuitively understand relationships between variables. However, PLS-SEM can be used with small amounts of data and allows for intuitive relationship analysis, which is expected to be highly beneficial for decision-makers.

Furthermore, the relationships between environmental variables and HABs derived from this model have limitations in interpreting causality due to factors such as the seasonality of environmental variables. Therefore, it is proposed that future studies focus exclusively on data from the warm season while considering the time-lag, spatial characteristics, and interspecies influences as examined in this study. This approach is expected to facilitate a more comprehensive interpretation of the relationships between environmental variables and HABs.

This research was funded by the National Institute of Environmental Research (NIER) grant funded by the Ministry of Environment (MoE) of the Republic of Korea (NIER-SP2023-383).

All relevant data are available from an online repository or repositories.

The authors declare there is no conflict.