Introduction

Eutrophication of surface waters such as rivers is one of the most common environmental problems in the world in terms of water quality. The main consequences of the excess of nutrients in lakes, coastal waters, large rivers or small streams are hypoxia/anoxia, increase of primary production (abundant algae growth) followed by increase of turbidity, disturbance of the aquatic ecosystems equilibrium, among others which are direct consequences of eutrophication (Gibson et al. 2000; Yasin et al. 2010; Chen et al. 2016). The European Environmental Agency defines eutrophication as an increase in the rate of organic matter supply to an ecosystem, which most commonly is related to nutrient enrichment enhancing the primary production in the system (EEA 2001). Often, rivers are important conduit of nutrients but algal production may be very low due to light limitation (from turbidity or forest cover), or high river velocity upstream. In contrast, eutrophication is more pronounced in rivers’ delta, estuaries, and in lakes. The trophic state of lakes and tropical rivers is usually classified into six grades based on chlorophyll-a concentrations: 0–1.6 μg/L representing oligotrophic, 1.6–10 μg/L mesotrophic, 10–26 μg/L light-eutrophic, 26–64 μg/L mid-eutrophic, 64–160 μg/L high-eutrophic, and > 160 μg/L hypereutrophic (Jin and Tu 1990; Huo et al. 2013). The increasing high anthropogenic inputs of nutrients in river watersheds are well known as the main culprit of eutrophication (Aguiar et al. 2011; N’goran et al. 2019). Therefore, it is necessary to monitor the evolution of eutrophication for the preservation of the freshwater quality.

Many indices such as the trophic indexes (TRIX, TLI, and TSI) have been used for both the monitoring of the surface water quality and classification of river and lake trophic levels using the chlorophyll-a, nutrients concentrations along with a few physicochemical parameters (Pesce and Wunderlin 2000; Moscuzza et al. 2007; Pettine et al. 2007; Cunha et al. 2013; Huo et al. 2013; Ota et al. 2015; Singh and Singh 2015; Wu et al. 2018). Through linear regression, these indexes provide an insight on how nutrient (i.e., dissolved inorganic nitrogen, reactive phosphorus, total phosphorus), light availability and other factors (e.g., dissolved oxygen, potassium permanganate) stimulate algal biomass development (usually measured as chlorophyll-a, Chl-a) and contribute to the increase of the aquatic systems enrichment condition (Primpas and Karydis 2010; Cunha et al. 2013; Li et al. 2017).

However, the prediction of eutrophication from these indices remains difficult given the interdependence and complexity of climatic, geographical and ecological factors affecting this phenomenon. Artificial Neural Network (ANN) is an increasingly popular alternative in environmental modeling because of its high potential for predicting complex relationships with precise accuracy (Sudheer et al. 2002; Sudheer and Jain 2004; Lohani et al. 2011; Wu et al. 2013; Wu et al. 2014; Huang and Gao 2017; Chen et al. 2018). ANNs are mathematical tools whose functioning is inspired by that of the human brain. Like their biological counterpart, neurons in layers receive, treat (by weighted summation), and transfer information generally via a nonlinear function (Assidjo et al. 2008). It has been demonstrated that ANNs are better than many linear and mechanistic regression models because of their ability to simulate non-linear phenomena including algal bloom dynamics and dissolved oxygen (DO) from water quality monitoring data. For example, Muller and Muller (2015) forecasted estuarine hypoxia in the main-stem of the Chesapeake Bay using a wavelet based ANN model. Chang et al. (2015) used ANNs to successfully forecast the low flow velocity regime in a constructed wetland in the Florida Everglades. Huo et al. (2013) predicted eutrophication with indicators such as DO, total nitrogen (TN), chlorophyll-a (Chl-a) and secchi disk depth (SD) with reasonable accuracy in Lake Fuxian, the deepest lake of southwest China, while Huang and Gao (2017) simulated chlorophyll-a in Lake Poyang in China using ANN models. However, studies in tropical regions are concentrated primarily in Asia, and limited data is available for West African rivers (Awad 2014; Wang et al. 2015; Tian et al. 2017; Hao et al. 2019). As an example, in Côte d’Ivoire, studies on eutrophication modeling have been conducted only in lagoons (Yao et al. 2017). Given that river waters are vital for people in rural areas, prediction of eutrophication in these waters is necessary. The aim of this study was to model the spatio-temporal evolution of chlorophyll-a, the main indicator of eutrophication using Artificial Neural Network in three typical rivers of Côte d’Ivoire (The Comoé, Bandama and Bia Rivers) which are used for drinking, fishing, bathing, and irrigation by the local populations without any treatment. One of the main detrimental effects of eutrophication is the increased occurrence of harmful algae, especially Cyanobacteria which can present chronic health risks to humans (Le Moal et al. 2019). For example, liver cancer was observed in a population from central Serbia following the consumption of water contaminated by Cyanobacteria (Svirčev et al. 2009). Secondly, eutrophication of these three rivers could result in socio-economic impacts such as cost of water treatment and financial problems for fishing communities. From field data, simulations for different rivers were made; then, the adequacy between predicted and target data was discussed.

Materials and Methods

Study Area and Samplings

The present study was conducted on three main rivers: the Comoé, Bandama and Bia Rivers which irrigate southeastern Côte d’Ivoire (Fig. 1). The Comoé River originates from the Banfora region in Burkina Faso and traverses 1160 km in Côte d’Ivoire before discharging into Ebrié Lagoon and the Gulf of Guinea. It has an annual average flow of about 106 m3/s and a drainage basin of 82,408 km2. The Bandama River takes its source in the northern Côte d’Ivoire, between Korhogo and Boundiali at an altitude of 480 m and flows into Grand-Lahou Lagoon and the Gulf of Guinea in the South. With a length of 1050 km, its catchment area covers 97,500 km2 with an annual average discharge about 263 m3/s.

Fig. 1
figure 1

Sampling stations along the Comoé, Bandama and Bia Rivers

Full size image

The hydrological regime of the Comoé and Bandama Rivers is essentially a tropical transitional regime with a single flood from August to October, and a long period of low water from January to May (Fig. 2).

Fig. 2
figure 2

Annual discharges of the Comoé, Bandama and Bia Rivers

Full size image

The Bia River is a small river located in the southeast corner of Côte d’Ivoire. It originates from Ghana in north of Chemraso, flows southward and discharges into Aby Lagoon. The length of the Bia River is about 120 km in Côte d’Ivoire, and the annual average discharge is 104 m3/s. The Bia River is linked to an equatorial transition regime marked by two annual floods. The first (usually the strongest) and second floods, occur during the June–July period and the October–November period, respectively (Durand et al. 1994; Girard et al. 1970).

Four climatic seasons occur in the study area: a long dry season (December–March), a long rainy season (April–July), a short dry season (August-middle September) and a short rainy season (Middle September–November).

Six sampling campaigns were conducted in the Comoé, Bandama and Bia Rivers from March 2016 to November 2017 during the dry, rainy and flood seasons. Five stations were sampled in the Comoé River from station CO1 to station CO5, five in the Bandama River from station BA1 to station BA5, and five in the Bia River from station BIA1 to station BIA5 (Table 1).

Table 1 Geographical locations of sampling stations and sampling period
Full size table

Procedure, Reagents and Quality Assurance

A Multiparameter Meter with GPS - HANNA HI 9828 was used to measure in situ parameters including temperature, pH, salinity, conductivity and dissolved oxygen. Water samples were collected with a 2.5 L Niskin bottle at 0.3 m below the surface of the streams. Water samples were filtered through a 0.45 μm Whatman GF/C filter and then stored in polyethylene flasks of 1 L volume (previously washed with 10% HCl and rinsed three times with ultrapure water). Subsequently, these flasks were acidified with sulfuric acid or hydrochloric acid, then stored at 4 °C and protected from light in a cooler. Analyses were performed within 72 h in the laboratory. Nutrients and chlorophyll-a analyses were performed according to Rodier et al. (2009) reference method. A portable spectrophotometer (HACH model DR/2400) was used for analysis (Koroleff 1970 for ammonium ion measurements; Grasshoff et al. 1999 for nitrates and Murphy and Riley 1962 for phosphates). Chlorophyll-a was extracted with 90% acetone in the dark and analyzed by Lorenzen’s method (Lorenzen 1967). Blanks were analyzed in each batch of water samples throughout the entire analytical procedure. The accuracy and precision of the results were checked by triplicate measurements. The detection limits were 0.004 mg/L for PO43−, 0.002 mg/L for NH4+, 0.005 mg/L for NO3− and 0.1 μg/L for Chl-a. The errors in analyses in terms of standard deviations of triplicate samples were found to be 0.001, 0.005, 0.02 mg/L and 0.38 μg/L for PO43−, NH4+, NO3− and Chl-a, respectively.

Data Set

The time-series data collected during the 2 years allowed for a characterization of water physico- chemical parameters and the chlorophyll-a (Chl-a) values in the Comoé, Bandama and Bia Rivers (Table 2). The data set consisted of 150 measurements per parameter. These parameters (predictors) can interact with the chlorophyll-a dynamics in rivers. We investigated the presence of associations between both predictor and response variables by calculating Pearson and Spearman correlation coefficients between all possible combinations of variables. The correlation coefficients estimated between predictors and the response variable (Chl-a) demonstrated that many of the predictor-response associations were nonlinear, as evidenced by the values of the Pearson correlation scores being lower than the Spearman correlation scores (Tables S1 to S3). This result suggested that such associations between predictor variables and the response should not be modeled with linear models. Therefore, ANNs are an ideal alternative to model these relationships as they are better suited to account for the non-linear associations between the response and the interactions among predictors (Li et al. 2013).

Table 2 Range, mean, standard deviation and relative standard deviation for water quality data, meteorological data and chlorophyll-a in the three sampled rivers between 2016 and 2017
Full size table

Neural Network

Artificial Neural Network (ANN) Model Development

Finding neural network models includes computing appropriate weights and biases that minimize the discrepancy between observed and simulated data. Development of artificial neural networks were performed based on the Matlab neural network toolbox R2018a Software. During all simulations, the Levenberg-Marquardt algorithm (Levenberg 1944; Marquardt 1963) was used to accelerate the training step. Artificial neural network models were built to predict chlorophyll-a concentration. The simulations were carried out initially with as predictors, all the ten parameters measured (Temperature, pH, EC, DO, salinity, PO43−, NO3−, NH4+, river discharge, and pluviometry). Two approaches were considered in this study (Fig. 3). First of all, the static modeling of the chlorophyll-a was performed using ten predictors constituting the input parameters Xi and measured chlorophyll-a values in the river as the output parameter Yi. Then, the dynamic models of chlorophyll-a were performed adding the time, namely the month of the season during which the sampling was conducted, to the ten previous predictors. The response Yi(t) represents the forecasting chlorophyll-a concentration in the river for the current season while the response Yi(t + 1) represents the chlorophyll-a concentration for the next season in the river. It should be noted that the static modeling approach forecasts the present values of chlorophyll-a without including the time (month) as an input parameter, while the dynamics modeling approaches forecast the future values of chlorophyll-a (Fig. 3).

Fig. 3
figure 3

Structures of the Artificial Neural Network (ANN)

Full size image

To form the network, data were randomly and evenly divided into three subsets by the calculation algorithm: training (50% of data), testing (25% of data) and validation (25% of data). The training subset is used for computing and updating the network weights and biases (Assidjo et al. 2008). The error in the testing set is monitored during the training process. The testing error will normally decrease during the initial phase of training, as does the training set error. However, when the network begins to overfit the data, the error in the testing set will typically begin to increase. When the testing error increases for a specified number of iterations, the training is stopped and the weights and biases of the minimum testing error are returned. The validation data set are only used to test the final solution in order to assess the performance of the network (Assidjo et al. 2008).

All inputs were normalized using the following formula:

$$ {x}_{ni}=\frac{2\left({x}_i-{x}_{min}\right)}{\left({x}_{max}-{x}_{min}\right)}-1 $$
(1)

Where xni is the normalized data ranging between −1 and 1, xi is the initial data, and xmin and xmax are minimum and maximum values of the data set.

The network architecture has been optimized by varying the hidden layer neurons number from 1 to 15 taking into account the highest correlation coefficients of training (Rapp) and validation (Rval) phases, but also the lowest values of the mean squared errors (MSE). For each hidden neuron, simulations were carried out 1500 × 5 times by randomly generating weights and the best result of the corresponding network architecture has been registered. The transfer function of hidden layer nodes was the tanh function and the linear function for the output layer (Lee et al. 2016). Every hidden neuron produces an output y′j according to Eq. 2:

$$ y{\prime}_j=\tanh \left(\sum {w}_{ij}{x}_i+{b}_j\right) $$
(2)

Where wij represents the weights, xi the inputs, bj the bias associated with the output y′j and tanh is the transfer function between the input layer and the hidden neuron.

Each output neuron produces an output value Yi which represents the predicted value of chlorophyll-a according to Eq. 3:

$$ {Y}_i=\sum \left({v}_{ij}y{\prime}_j\right)+{b}_i $$
(3)

Where vij represents the weights, y′j the hidden neurons and bi the bias associated with the output Yi (Assidjo et al. 2008).

ANNs models obtained appeared quite complex with sometimes up to 13 neurons in the hidden layers. It was necessary to carry out tests to ensure that there was no overfitting. Thus, the relevance of the input variables of the ANN models was estimated according to the Olden method (Olden et al. 2004) for chlorophyll-a prediction in each river. Connection weight approach of Olden et al. (2004) calculates the product of the raw input-hidden and hidden-output connection weights between each input neuron and output neuron and sums the products across all hidden neurons (Plate 1 in supplementary files). This approach assigned importance values to the predictors that were proportional to their influence on the output of the model. These values were above zero if the predictor had a positive association with the response variable and below zero if the predictor had a negative association with the response variable (Coutinho et al. 2019). Networks with values of importance assigned to predictors below −0.5 or above +0.5 were considered valid, in order to exclude overly complex models. This filtering criteria gives the best results and reduces to a maximum to 6, the number of neurons on the hidden layer in the final models. When two predictor values presented close importance or a linear correlation (Pearson), a preferential choice was made for the one with the highest importance. This would reduce the networks size at most six predictors on the input layer, and that would provide a nice trade-off between predictive powers while avoiding overfitting. Thus, the simulations were redone 1500 × 5 times with the predictors selected at the input by randomly generating new weights, and the best result of the corresponding network architecture was registered. The raw data was provided in Tables S4 to S6 (supplementary files).

The principal component analysis (PCA) was also, applied to data from the Comoé, Bandama and Bia Rivers (1650 measurements per river) to examine potential factors driving the physical and chemical parameters. The PCA was performed using Statistica® version 7 software. All graphs were performed using Sigmaplot® version 12.0 and 14.0.

Model Performance

Evaluation of the model performance consists in judging its ability to predict chlorophyll-a dynamics in the Comoé, Bandama and Bia Rivers. The ANN program calculated the correlation coefficients for the training (Rapp) and validation (Rval) sets and the mean square errors for the training (MSEapp) and validation (MSEval) sets. The best networks were selected among those that achieve the highest performance (i.e., high Rapp and Rval with low MSEval values) without overfitting (i.e., adequate fits for both training and validation sets). In all, 225 ANNs were built for each river after the simulations. Networks were considered valid if they were characterized by Rapp > 0.5; Rval > 0.5 and MSE < 1 (Coutinho et al. 2019).

Another approach is determination of Schwarz’s Bayesian information criterion (BIC) (Schwarz 1978; Assidjo et al. 2008) obtained as follows:

$$ BIC=\log (MSE)+p\ \frac{\log (n)}{n} $$
(4)

Where MSE is mean squared errors, n the number of training patterns and p the total number of network weights. So the smallest BIC would give the best generalizable model.

The mean relative error (MRE) is also introduced to evaluate the performance of ANN models and to estimate whether the obtained model is generalizable (Xu et al. 2015).

$$ MRE=\frac{1}{n}\sum \limits_{i=1}^n\left|\frac{Y_{pr}-{Y}_{obs}}{Y_{obs}}\right|\times 100 $$
(5)

Where n is the number of the samples in test set.

From eq. (5), it is obvious that the smaller the MRE is, the better generalization is achieved (Xu et al. 2015).

Results and Discussion

Chlorophyll-a Dynamics

The spatio-temporal distributions of the chlorophyll-a in the Comoé, Bandama, and Bia Rivers during the dry, rainy and flood seasons are shown in Fig. 4. Thematic maps were computed using the ArcGIS/ArcMap environment based on Inverse Distance Weighting (IDW) interpolation. The seasonal variation of phytoplankton biomass distributions, expressed as chlorophyll-a showed no clear trend over the study period (March 2016 to November 2017) at the three sites as a result of urban and agricultural inputs. Very high chlorophyll-a concentration (164 μg/L) was observed during the flood season at the mouth of the Comoé River (CO5), while a very low value (6.20 μg/L) was obtained in the same river at Bonoua station (CO4) during the flood season. Strong proliferation of phytoplankton had been noted in the estuarine part of the three rivers during the three seasons because rivers are often important conduit of nutrients to downstream habitats (Yasin et al. 2010). Yet, this proliferation was not pronounced upstream especially during the dry and flood seasons, except at the Manzan station (CO2) where rainwater runoffs loaded in nitrates led to an increase in photosynthetic activity (de Sousa Barroso et al. 2016). Pearson correlation analyses were performed to test for associations between chlorophyll-a abundance and measured environmental parameters, especially nitrate and phosphate. (Tables S1 to S3). Chlorophyll-a correlated positively with phosphates in the Bandama and Bia Rivers and with nitrates in the Comoé River. The river discharge and dissolved oxygen (DO) correlated negatively with chlorophyll-a in the Bandama and Bia Rivers, respectively. The lower yet significant Pearson correlation scores observed (range: - 0.236 to +0.433) suggest that the variation of these nutrients and physicochemical parameters could linearly account for the chlorophyll-a values. The chlorophyll-a values increased with phosphate inputs in the Bandama and Bia Rivers, and nitrate increased with chlorophyll-a in the Comoé River mainly in the rainy season. This result could indicate a limitation of nitrogen in the Comoé River and phosphorus in the Bandama and Bia Rivers, respectively, during the rainy season. Overall, the map distributions and the presence of associations between the chlorophyll-a abundance with the environmental parameters measured, show that chlorophyll-a dynamics in the rivers could be dictated by nitrates and phosphates availability, point sources and diffuse sources as well as the river flow.

Fig. 4
figure 4

Spatio-temporal chlorophyll-a dynamics in the Comoé, Bandama and Bia Rivers from March 2016 to November 2017. Data plotted are average concentrations obtained from March 2016 and March 2017 during the dry season, June 2016 and June 2017 during rainy season, and October 2016 and November 2017 during flood season

Full size image

Eutrophication State

The trophic status of the rivers was assessed using the classification scheme for Chl-a biomass (EEA 2001). The water quality of the Comoé, Bandama and Bia Rivers was most often under the mesotrophic state during the three sampling seasons (Table 3). This finding confirms our rationale that high nutrient concentrations result in sometimes a hypereutrophic state of the rivers, especially during the flood season as revealed by the significant correlations obtained between nutrients and algae (Chl-a). In addition, during the dry and rainy seasons, abundant sunlight, water temperature and aquatic chemistry create an ideal environment for algae growth. Algae blooms in rivers in summers were related to nutrient enrichment, especially in nitrogen and phosphorus from primarily agriculture non-point pollution (Wang et al. 2004; Zhang et al. 2011).

Table 3 The classification scheme for Chl-a biomass in the Comoé, Bandama and Bia Rivers according to the classification of the European Environmental Agency (EEA 2001). Data are average concentrations obtained from March 2016 and March 2017 during the dry season, June 2016 and June 2017 during rainy season, and October 2016 and November 2017 during flood season
Full size table

Predictive Power

The relative importance of the predictor values in the best performing ANNs was estimated using the Olden method (Olden et al. 2004) (Fig.5). The further away from zero, the more important the predictor was, regardless of sign. Given the performance of static modeling of Chl-a in the three rivers, the high scores Rval (range: 0.67 to 0.86) and the low MSEval (range: 0.05 to 0.15) testified to the accuracy of ANN networks obtained with at most six predictors such as water salinity, EC, temperature, PO43−, pluviometry, NO3− (Table 4). Although the predictor importance varied among these networks, the overall ranking of the predictors was conserved. Four to six predictors forecasted the measured Chl-a values using the dynamic approaches. The high scores Rval (range: 0.63 to 0.95) and the low MSEval (range: 0.07 to 0.22) supported the accuracy of the obtained ANN networks.

Fig. 5
figure 5

Bar plot depicting the relative importance of predictors for the response estimated from ANN weights through the Olden method

Full size image
Table 4 Performance and variable importance for the best performing networks for chlorophyll-a prediction. The rank of predictor variable importance is included above the first and the last variables
Full size table

A PCA analysis was performed on the samples collected from all rivers and seasons (the dry, rainy and flood seasons) to examine factors driving the distribution of the studied parameters in the rivers (Table 5). The results showed three significant components for each season, accounting for 71.3, 72.7 and 71.2%, respectively, of the total variances of information contained in the original dataset. Water EC, salinity, pluviometry and river discharge showed high loadings (absolute weight > 0.6) on factor 1 during the three seasons. Moreover, the Pearson correlations showed linear associations between river discharge with pluviometry, EC and water salinity, but there was no linear association of these parameters with Chl-a. The factor 1 shows that pluviometry (rainfall) may drive the distribution of water EC, salinity, and river discharge in the rivers.

Table 5 Sorted rotated factor loadings (Varimax normalized) of data in the three principal factors derived from the principal component analysis (PCA) for the Comoé, Bandama and Bia Rivers
Full size table

Factor 2 showed high loadings (absolute weight > 0.6) with pH and DO during the dry season, with PO43− and NO3− during the rainy season, and with NO3− and temperature during the flood season. Moreover, the Pearson correlations showed linear associations between PO43− and Chl-a in the Bandama (r = 0.29) and Bia (r = 0.43) Rivers, while NO3− and Chl-a had linear association in the Comoé River (r = 0.37). The rainy and flood seasons caused excess phosphorus and nitrate inputs probably through soil erosion; this may lead to hyperfertilization of the environment and an increase in primary production (Gladyshev and Gubelitc 2019). So, factor 2 shows that PO43− and NO3− impact chlorophyll-a abundance.

Factor 3 showed high loadings (absolute weight > 0.6) with NO3− during the dry season; with NH4+ during the rainy season, and with DO during the flood season. This factor shows that DO and NO3− and NH4+ ions may impact chlorophyll-a abundance. Algal growth in the aquatic system was also accompanied by a drop in oxygen demand, as example in the Bia River where a negative linear correlation between Chl-a and DO was obtained (r = − 0.18). Overall, the PCA results show that rainfall and the seasons play a key role on the distribution of the physical and chemical parameters in the rivers.

Chlorophyll-a Prediction

The highest Rapp and Rval values corresponding to the lowest MSEval and BICs of each river are recorded in Table 6. The correlation coefficient Rapp varied from 0.85 to 0.98 for the training subset, while the correlation coefficient Rval varied from 0.63 to 0.95 for the validation subset. Further, the choice of potentially generalizable ANN models was made taking into account the lower values of MSEval (0.05 to 0.22) and the BIC (< − 0.5). These results allowed to identify the models of the network architecture (Ngoran et al. 2009; Yao et al. 2017) that can simulate the dynamics of chlorophyll-a in the Comoé, Bandama and Bia Rivers, according to both approaches (i.e., static and dynamic). Yet, since the best network model is a compromise between the results obtained during the computing process (training and validation) and the adequacy of the simulated and observed values, regression plots were constructed to compare experimental and predicted data of chlorophyll-a in the three rivers (Fig. 6). Yobs is the target variable (i.e., measured chlorophyll-a) and Ypr is the simulated variable (i.e., predicted chlorophyll-a). Each graph consists of one dashed line called a best-fit line (y = T, i.e., predicted = experimental data) with a correlation value (Rall) on the plot. The best generalizable models are the ones that predict more than 50% (Rall > 0.5) of the chlorophyll-a value (Mandal et al. 2015).

Table 6 Characteristics of the models obtained for each river
Full size table
Fig. 6
figure 6

Comparison of experimental and predicted data obtained from ANN

Full size image

As shown in Figs. 6 and 7, the 5–5 - 1 model predicted 76% of the static evolution of chlorophyll-a, while the 4–6 - 1 and 5–3 - 1 models predicted 82% and 76%, respectively, of the dynamic evolution of chlorophyll-a at season (t) and the next season (t + 1), in the Comoé River. Overall, the high Rapp (> 0.85) and Rval (> 0.60) scores presented in Table 6 and the low discrepancies between expected and observed values displayed in Fig. 7 are indicative of a good prediction of the chlorophyll-a in this river by ANNs models.

Fig. 7
figure 7

Measured and predicted values of Chlorophyll-a from static and dynamic approaches in the Comoé River

Full size image

With regard to the Bandama River, the 6–5 - 1 model predicted 76% of the static evolution of chlorophyll-a, while the 5–3 - 1 and 6–6 - 1 models predicted 79% and 73%, respectively, of the dynamic evolution of chlorophyll-a at season (t) and the next season (t + 1). This is supported by Fig. 8 on which the static and dynamics (t and t + 1) approach curves showed very similar values between predicted and field measurements. Moreover the high Rapp (> 0.85) and Rval (> 0.65) scores presented in Table 6 and the low discrepancies between expected and observed values displayed in Fig. 6 are indicative of the good prediction of the Chlorophyll-a in this river by ANNs models.

Fig. 8
figure 8

Measured and predicted values of Chlorophyll-a from static and dynamic approaches in the Bandama River

Full size image

As for the Bia River, the 4–4 - 1 model predicted 85% of the static evolution of chlorophyll-a, while the 6–5 - 1 and 5–4 - 1 models predicted 84% and 79% of the dynamics of chlorophyll-a at season (t) and the next season (t + 1), respectively. As seen on Fig. 9, the static and dynamic approach curves predicted chlorophyll-a values very close to the field data. Moreover the high Rapp (> 0.85) and Rval (> 0.70) scores presented in Table 6 and the low discrepancies between expected and observed values displayed in Fig. 6 are indicative of the good prediction of the Chlorophyll-a in this river by ANNs models.

Fig. 9
figure 9

Measured and predicted values of Chlorophyll-a from static and dynamic approaches in the Bia River

Full size image

In general, the ANN models showed a high degree of consistency between the field and prediction data.

The Bia River MREs were 3.90, 3.66, and 5.27%, respectively, for the static and dynamic approaches (t and t + 1), indicating low difference between the outputs of the ANN models and field data. Therefore, the 4–4 - 1, 6–5 - 1 and 5–4 - 1 models are generalizable for the Bia River (Table 7). The same was true for the 6–5 - 1, 5–3 - 1 and 6–6 - 1 models in the Bandama River with MRE values of 4.58, 3.95 and 3.70%, respectively, and for the 5–5 - 1, 4–6 - 1 and 5–3 - 1 models in the Comoé River with MRE values of 7.71, 5.51 and 3.22%, respectively.

Table 7 Performance of the ANN models obtained
Full size table

This study revealed nine best neural networks for forecasting Chl-a concentrations in the Comoé, Bandama and Bia Rivers. The neural networks were successful in modeling between 73 and 85% of the response (chlorophyll-a concentration) with at most six predictor variables. The network performance was relatively comparable to that of ANNs built by Coutinho et al. (2019) (82%) to model Chl-a abundance in Guanabara Bay, Rio de Janeiro using water temperature, transparency, total nitrogen, total phosphorus, and salinity as relevant predictors and by Flombaum et al. (2013) (68%) to model microbial abundance in the Global Ocean. Therefore, the neural network approach is a successful method of modeling such complex and non-linear phenomena as algal blooms in freshwater systems with different environmental conditions. Thus, the hypothesis that ecosystem processes are best modeled when non-linearity of relationships is assumed is retained (Recknagel et al. 1997; Wilson and Recknagel 2001). Yet, Recknagel et al. (2002) and Park et al. (2015) showed that artificial neural network models can be powerful short-term predictors of the timing of algal bloom events but are difficult to generalize and do not provide explicit explanation of underlying processes. The differences in the precision observed for the nine final networks in the present study could have been due to the need to include biotic predictors in modeling the response (Tromas et al. 2017).

Environmental Significance

Results from this study demonstrate that the Comoé, Bandama and Bia Rivers are in an advanced eutrophication state, based on Chl-a observations over 2 years. A number of environmental parameters such as water temperature, pH, EC, salinity, PO43−, and river discharge were found as significant drivers of chlorophyll-a concentration depending on the river.

Overall, the ANN model simulations data fitted accurately the field measurements; therefore, they can be useful as alert tool for decision-makers in the monitoring of chlorophyll-a (thus, the eutrophication phenomenon) evolution in the Comoé, Bia and Bandama Rivers.

Conclusion

The monitoring of riverine water eutrophication state in West Africa represents a major challenge for the leaders as these waters are vital for the subsistence of populations. This study revealed that chlorophyll-a dynamic in the Comoé, Bandama and Bia Rivers was characterized by higher levels upstream during the rainy season as well as very high values downstream in the estuarine parts during the dry, rainy and flood seasons. Chlorophyll-a concentrations in the rivers resulted mainly from anthropogenic inputs such as agricultural and urban runoffs, leading to a progressive deterioration of the waters. Artificial neural network (ANN) models predicted up to 85% of the chlorophyll-a concentrations in the rivers with satisfactory accuracy. Therefore, the result of this study demonstrates that the degree of eutrophication in tropical rivers can be predicted successfully using ANN models with few data in the field, which could benefit river management with less cost for water quality monitoring. It should be appropriate to raise awareness of the populations on the moderate use of agricultural inputs in order to reduce water eutrophication.