Ashtamudi estuary is one among the Ramsar sites and is renowned for its fish diversity. Located between 8°53'N to 9°2'N latitude and 76°31'E to 76°41'E longitude, the estuary is connected to the Arabian Sea through a perennial opening (Fig. 1). The estuary receives freshwater discharge from the Kallada River, which originates from the Kulathupuzha hills of the Western Ghats at a distance of 120 km upstream of the Ashtamudi estuary. The estuary has a maximum depth of over 6 m and is the deepest estuary in the State of Kerala. The annual average rainfall prevailing in the area is about 2400 mm and the estuary acts as flood storage for the city of Kollam. The estuary is rich in various mangrove species like Avicennia officinalis and supports over 57 species of avifauna. The estuary is a habitat for phytoplankton of different genera.

The estuary is stressed with anthropogenic activities including over-fishing, solid waste dumping (Priya et al. 2023), waste water discharge and cage fish farming. Being a hot tourist spot, the water quality is affected adversely due to unscientific practices in the catchment area. The excessive usage of chemical fertilizers in agricultural areas near the estuary has led to the inflow of nutrients, leading to eutrophication. The estuary has been classified as eutrophic as reported in previous studies (Sruthy et al. 2021). This necessitates the development of prediction models for the most prominent parameter that represents eutrophication state namely, chlorophyll-a. For this, data pertaining to various water quality parameters such as pH, turbidity, salinity, DO, TN, TP, chlorophyll-a were collected from the government agencies during the periods 2012 to 2018 over a monthly basis. The location of sample station was at mid-estuarine reaches and the data from one single station alone could be considered for the present study due to lack of data from spatial stations. However, the data analysis of one station shall help in identifying the effect of salinity on chlorophyll-a at a particular location, and hence, salinity can be used as an indirect parameter for representing estuarine mixing. The raw data were cleaned and checked for outliers and then normalized to enhance the performance of the AI models. The data were set back to its original form after network training.

Figure 1. 

Location map of the study area.

The performance of prediction models is dependent on the input variables. Thus, the significant variables affecting chlorophyll-a was selected appropriately for enabling the models to accurately predict chlorophyll-a concentration. In order to identify the significant influencing parameters, correlation coefficient between chlorophyll-a and water quality variables were estimated. The parameters that have exhibited a significant correlation of at least by 95% confidence level were selected as input parameters for model development.

The chlorophyll-a prediction models were developed using Artificial Neural Networks (ANN) and Adaptive Neuro Fuzzy Inference System (ANFIS) modelling approaches. Of the total data set, 70% of the data were used for training the models and the remaining 30% of the data were used for testing and validation purposes.

2.3.1 ANN models

In the present study, three ANN models were developed and tested: Back Propagation Neural Network (BPNN), Generalized Regression Neural Network (GRNN), and Recurrent Neural Network (RNN). BPNN is among the most commonly adopted NNs for solving environmental problems. It generally consists of an input layer that contains the input parameters, a hidden layer that has several neurons, the number of which is fixed using a trial-and-error procedure, and an output layer that contains the output variable, chlorophyll-a. The error in the prediction of chlorophyll-a is minimized by adjusting the weights during the training process and is done by distributing the errors using the back propagation algorithm (Tian et al. 2017). The BPNN architecture used for the study is shown in Fig. 2a. GRNN is a type of radial basis function (RBF) neural network that has fast learning and rapid convergence capabilities (Chen and Liu 2015). They are commonly used to deal with non-linear problems. They consist of a feed-forward network having an input layer, one hidden layer and one output layer. The common GRNN formulation can be expressed as

$y=\sum _{i=1}^n{w}_i{\phi }_i\left(x\right)$ (1)

Where y is the output, w_i is the weight between hidden and output layer, φ_i (x) is the output value of the neuron in the hidden layer after transfer by GRNN.

RNN is a type of deep learning neural network model that works through parameter sharing and state transformation. It consists of an input and output layer and node groups, collectively called RNN cell. The model parameters are updated during training phase. A typical structure of RNN adopted for the study is shown in Fig. 2b.

Figure 2. 

(a) BPNN architecture; (b) RNN architecture; (c) ANFIS architecture; and (d) membership functions for ANFIS model adopted for the study.

In order to identify the optimal chlorophyll-a prediction model, the performance indicators namely, coefficient of determination (R²) and root-mean-square error (RMSE) were used. The higher R² and lower RMSE were used as the criteria for identifying the best prediction models for chlorophyll-a.

$R^2=1-\frac{{\sum }_{i=1}^n(x_i-y_i{)}^2}{{\sum }_{i=1}^n(x_i-\overline{x}{)}^2}$

$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n(y_i-x_i{)}^2}$ (2) (3)

Where n is the number of observations, x_i is the observed chlorophyll-a concentration, y_i is the predicted chlorophyll-a concentration, x̄ is the mean value of observed chlorophyll-a.

The water quality parameters namely, pH, turbidity, salinity, dissolved oxygen, total nitrogen, total phosphorous and chlorophyll-a were used for the study. Table 1 gives the statistical summary of the parameters. The turbidity variations were between 0.251 and 19.98 NTU, a higher turbidity can indirectly represent higher chlorophyll-a concentration. Salinity variations were much higher in the estuarine environment and is much dependent on the freshwater flow and tidal flow. A minimum of 1.25‰ was observed during monsoon period, while salinity as high as 30‰ has been detected during the pre-monsoon, when the freshwater influx was negligible. There was a wide variation in pH and dissolved oxygen, the estuarine waters were predominantly alkaline in nature with satisfactory to good dissolved oxygen for supporting aquatic life. The mixing and circulation within the estuarine reaches has been the main reason for saturated dissolved oxygen in the estuary, except for a depletion during pre-monsoon period. The total nitrogen and total phosphorous varied from 110 to 950 μg/L and 15 to 490 μg/L respectively, with higher values observed during pre-monsoon season. A mean value of 7.6 μg/L of chlorophyll-a was observed with variations between 0.9 and 17.9 μg/L over a temporal scale.

The chlorophyll-a variations are influenced by many parameters like availability of sunlight, nutrients, mixing conditions and these parameters are represented by measurable variables like turbidity, salinity, total nitrogen and total phosphorous. The concentration of chlorophyll-a can also affects the water quality parameters such as pH and dissolved oxygen or in other terms, the variations in these parameters can indirectly represent the levels of chlorophyll-a. Therefore, to identify the factors influencing chlorophyll-a, its correlation with the water quality parameters were estimated using correlation coefficient, which is a commonly adopted indicator of the relation between two parameters. Table 2 represents that chlorophyll-a has the maximum correlation with total nitrogen and total phosphorous, with a higher significance of 99% confidence level, followed by turbidity and salinity, whose correlation with chlorophyll-a is significant at 95% confidence level. Dissolved oxygen was observed to have a negative weak correlation, while pH had a weak positive correlation, but were not significant. Accordingly, total nitrogen, total phosphorous, turbidity and salinity were considered for developing prediction models for chlorophyll-a.

The water quality parameters namely, turbidity, salinity, total nitrogen and total phosphorous showed significant correlation with chlorophyll-a. It is well recognized that total nitrogen and total phosphorous are causative parameters of chlorophyll-a, while turbidity can be considered as a response variable, because higher chlorophyll concentrations lead to the enhancement in water turbidity levels. Many studies that have developed prediction models for chlorophyll-a have used cause variables like nitrogen and phosphorous (Vollenweider et al. 1998; Gupta 2014; Luo et al. 2019) and water quality parameters that act as response variables too (Alves et al. 2013; O’Boyle et al. 2013; Chen and Liu 2015; Santos 2015). However, the salinity has been identified to have a significant correlation with chlorophyll-a in the Ashtamudi estuary. Therefore, to analyse the effect of salinity on the prediction of chlorophyll-a, two scenarios have been considered:

• Scenario I:

Chl-a = f(Turb, TN, TP)

• Scenario II:

Chl-a = f(Turb, Sal, TN, TP)

Where Chl-a is the chlorophyll-a concentration, Turb is the turbidity, Sal is the salinity, TN is the total nitrogen and TP is the total phosphorous. In order to further identify the best fit model, seven models were trained and tested using the available data: three neural network models namely, BPNN, GRNN, RNN; and four ANFIS models with triangular, trapezoidal, Gbell and Gaussian membership functions. Table 3 gives the model performance indicators for various models. Fig. 3 shows the predictability of chlorophyll-a using various models for the two scenarios considered for the study. It is observed that of the two scenarios, the one which considered salinity along with cause and response variables has improved the predictability of chlorophyll-a. GRNN model has been identified to be best suited for predicting chlorophyll-a for both the scenarios; by incorporating salinity, the predictability has enhanced marginally, leading to a prediction with an R² of 0.9976 for training and 0.9982 for testing, nearly approaching 1. The RNN model was identified to have the worst performance in predicting chlorophyll-a for scenario I with an R² of 0.37 and 0.21 respectively while training and testing and thus exhibiting drastic enhancement in its performance. Similarly, the performance of BPNN model was very poor for scenario I, especially during the testing phase, with an RMSE as high as 0.85, while its performance considerably improved for scenario II with 90.6% reduction in RMSE. These results highlight that ANN and RNN are much sensitive to the input parameters, while GRNN is more stable and can be applied in estuarine environments at various salinity conditions with less sensitiveness for the predicted chlorophyll-a with respect to the input variables. However, mostly all the ANFIS models were likely sensitive to the input variables almost to the same degree. Of the four ANFIS models, the models with Gaussian MF and Generalized bell MF exhibited good performance in chlorophyll-a predictions, especially for scenario II, followed by trapezoidal and triangular MF models. Surprisingly, they showed higher R² during testing phase than training phase.

Fig. 4 shows the validation of various models for the two scenarios studied. It is evident that ANN model over-predicted chlorophyll-a values at low chlorophyll-a levels, while under-predicted at higher chlorophyll-a levels for scenario I, while this trend was altered for scenario II, wherein the over-prediction of chlorophyll-a was considerably reduced at lower levels. In the case of GRNN, the model marginally under-predicted chlorophyll-a concentration at higher concentration levels, while there was a perfect fit of predicted chlorophyll-a with the observed values for scenario II. Even though there were wide under-predictions in chlorophyll-a for the RNN model, a good fit was obtained for scenario II. The predictability of all of the ANFIS models was more or less similar for scenario I with chlorophyll-a under-predictions at higher concentration, while the inclusion of salinity as an input parameter enabled the models to accurately predict the observed chlorophyll-a concentrations, thereby highlighting the role of salinity in the prediction of chlorophyll-a in estuarine environments.

Figure 3. 

Predicted chlorophyll-a for scenario I using (a) BPNN; (b) GRNN; (c) RNN; (d) ANFIS – triangular MF; (e) ANFIS – trapezoidal MF; (f) ANFIS – Gbell MF; (g) ANFIS – Gauss MF; and for scenario II using (h) GPNN; (i) GRNN; (j) RNN; (k) ANFIS – triangular MF; (l) ANFIS – trapezoidal MF; (m) ANFIS – Gbell MF; (n) ANFIS – Gauss MF.

Figure 4. 

Validation of predicted chlorophyll-a for scenario I using (a) BPNN; (b) GRNN; (c) RNN; (d) ANFIS – triangular MF; (e) ANFIS – trapezoidal MF; (f) ANFIS – Gbell MF; (g) ANFIS – Gauss MF; and for scenario II using (h) GPNN; (i) GRNN; (j) RNN; (k) ANFIS – triangular MF; (l) ANFIS – trapezoidal MF; (m) ANFIS – Gbell MF; (n) ANFIS – Gauss MF.

The salinity was identified to play a crucial role in chloroplyll-a dynamics in the Ashtamudi estuary. The incorporation of salinity as an input parameter significantly enhanced the predictability of chlorophyll-a. The variations of chlorophyll-a were observed to have a diversity at various salinity values as shown in Fig. 5a. The variations of chlorophyll-a were rather more pronounced at higher salinity values. At salinity values less than 5‰, there was an overall decreasing trend for chlorophyll-a with salinity, but the relation was not much significant (Fig. 5b). However, when the salinity ranged between 5 and 30‰, there was a pronounced increase in chlorophyll-a with an increase in salinity. As a larger % of the data set fall in the higher salinity category (of >5‰), the correlation between salinity and chlorophyll-a was obtained as positive. Even though the chlorophyll-a had a non-linear variation with salinity, the AI models were capable to replicate these variations and was able to predict the chlorophyll-a concentration at all salinity levels. At lower salinity levels, the freshwater flow is appreciable and this will create salinity gradients, causing the estuarine reaches to act as partially-stratified or stratified zones. In these zones, mixing is relatively less. A higher near-surface salinity represents higher mixing in the water column and these zones act as well-mixed zones. Higher mixing can promote the penetration of light into deeper zones, thereby enhancing primary productivity. This will have implications on the enhancement of chlorophyll-a concentrations. Thus, the study brings out the hypothesis that well-mixed conditions in estuarine reaches are associated with higher chlorophyll-a concentration and that salinity can be used as an indirect representation of mixing conditions that plays crucial role for chlorophyll-a prediction. Studies have to be extended to analyse the performance of the models on a temporal and spatial basis and for various tidal conditions. Further, an in-depth analysis on the influence of other water quality parameters on the chlorophyll-a dynamics and prediction is also to be performed. The validation of this hypothesis has to be extended to spatial attributes of the Ashtamudi estuary and also needs validation in other world estuaries.

Figure 5. 

Variation of chlorophyll-a with salinity (a) for salinity between 0 and 30‰; (b) for salinity between 0 and 5‰; and (c) for salinity between 5‰ and 30‰.

The authors have declared that no competing interests exist.

No ethical statement was reported.

No funding was reported.

Conceptualization: KLP. Data curation: SH. Formal analysis: JCCC. Investigation: SH, KLP. Writing - original draft: JCCC. Writing - review and editing: KLP, SH.

K. L. Priya https://orcid.org/0000-0002-8433-2920

Joan Cecilia C. Casila https://orcid.org/0000-0001-6319-8999

All of the data that support the findings of this study are available in the main text.