Estimation of zooplankton density with artificial neural networks (a new statistical approach) method, Elazığ-Türkiye

Cip reservoir was built in 1965 for irrigation purposes on the Cip Stream in Elazığ. The body volume of the its, which is an earth body fill type, is 446,000 m³, and its height from the stream bed is 23 meters. The reservoir volume is 7 hm³ at normal water level, and the reservoir area is 1.10 km² at normal water level. It provides irrigation to an area of 1,100 hectares. The plankton samples were collected monthly from Cip Reservoir in 2021- 2022, using a standard plankton net (55-μm mesh size) from three stations (Figure 1). The samples were fixed in 4% formalin, analysed under an inverted microscope (GMBH D-6330 diavert inverted microscope, Earnst Leitz Ltd., Canada) and identified under a compound microscope (Nikon Eclipse E 100, Nikon Instruments Inc., Japan). Water temperature, dissolved oxygen, pH, electrical conductivity and secchi disk were measured in-situ with the YSI professional plus brand meter. Alkalinity was also determined by the titrimetric method. Total phosphorus and total nitrogen were determined by analyses using Merck test kits (spectrophotometric) and the Water and Wastewater Analysis Photometer Merck Spectroquant Nova 60 A. Counting of zooplankton species was done in petri dishes with 5-ml sub-samples. A minimum of 200 individuals were quantified per replicate, and the final density was converted to individuals per cubic metre. Monthly changes of total zooplankton at stations were recorded.

Figure 1

Zooplanktonic organisms are a food source for invertebrates, fish, and birds. Nearly all freshwater fish are planktivorous in their early life. Planktivorous fish feed both small zooplankton and large phytoplankton (Horne & Goldman 1994). Factors affecting the horizontal and vertical distribution of zooplankton in water are the physical and chemical properties of water, wind, currents, streams entering the lake, depth, season, heat, light, nutrients and predators. When the light is high, the plankton descend from the surface and rise to the surface when the light is low. Since they cause the viscosity and density of water to change with temperature, living things are the best for themselves. Zooplankton gravitate towards the layer with the appropriate temperature (Tanyolaç 2009). Zooplankton density can be mentioned as an indicator of healthy water quality (Karjalainen et al. 1996; Moss et al. 1997; Muylaert et al. 2006).

This is connected with the transfer of energy from zooplankton producers, which represent the second trophic level in the food web, to heterotrophs with higher trophic levels. (Deivanai et al. 2004; Ismail & Mohd Adnan 2016). They respond quickly to physical and chemical changes in their environment. Previous studies have shown that different groups of zooplankton are good indicators of eutrophication: Attayde & Bozelli (1998); Burns & Galbraith (2007); Pinel-Alloul et al. (1990); Sousa et al (2008); Saler (2017); Bulut and Saler (2018); Bulut and Saler (2020); Bulut and Saler (2019). Studies of the effects of environmental factors and the density of zooplankton taxa provide information about the functioning of water systems (Bulut & Saler 2020).

Zooplankton are very sensitive to changes in their environment (Legendre & Demers 1984). Therefore, some important predictive parameters that directly or indirectly affect the zooplankton habitat in the Cip reservoir were selected. The factors that directly affect zooplankton in their habitat are water temperature and dissolved oxygen. In some seasons, due to the increase in flooding, water dilution, changes in nutrient and oxygen availability occur, and accordingly, the reproduction and metabolic rate of zooplankton are affected (Loverde et al. 2009). Electrical conductivity and pH, which are measures of production and decomposition processes, are variables that indirectly affect zooplankton density.

ANNs methodology, which provides significant advantages thanks to many features, is widely used in the field of predictive modeling as in other fields. Artificial Neural Networks (ANNs) are artificial information processing models, created by imitating the work of the human brain and taking advantage of the physiology of the brain. ANNs are some of the most successful new approaches in solving problems in recent years (Haykin 1994; Sagıroglu et al. 2003).

The learning feature of ANNs is one of the most important features that attract the attention of researchers because the ability to produce solutions for events that have never been seen before by learning the relationship between inputs and outputs about any event, whether linear or not, from existing examples, which forms the basis of intelligent behaviour in ANNs. A neural network stores information, makes it useful and consists of simple units. It is a parallel distributed processor. Artificial neurons are simply clustered in ANNs. This clustering is done in layers, and then these layers are related to one another. Basically, all neural networks have a similar structure. Some neurons are connected to the outside to receive inputs and some neurons to transmit outputs. All the remaining neurons are in the hidden layers, that is, they only have connections within the network (Anderson & McNeill 1992).

MATLAB’s Neural Network Toolbox (Ver R2016a) was used for ANNs. ANNs created in MATLAB software consist of three parts. These are “training”, “testing” and “validation”. The proposed model is divided into three layers: input, output and hidden layers. In addition, the model consisting of 31 neurons (8 in the input layer, 3 in the output layer and 20 in the hidden layer) is designed as fully connected feed-forward-feedback. The model structure of the system is shown in Figure 2.

Figure 2

In forward propagation, the temperature, pH, dissolved oxygen, conductivity, secchi disk, alkalinity, total nitrogen, total phosphorus data received in the input layer are taken from the input layer and transmitted to the hidden layer by passing through the activation function. In the hidden layer, new values from the previous layer are reactivated and transmitted to the output layer. Error rates are calculated between the targeted values and the new values obtained in the output layer. If the calculated error rate is greater than 1e^-7, the back propagation algorithm is run. If the errors are reflected in the weight values in the hidden layer and input layers, new weight values are created. Weight values and bias values are updated according to Equation 1 and Equation 2. The mathematical equation of the neuron model is as follows (Eq. 3) (Krenker et al. 2011): (yi (k) is the output value at discrete time k; F is a transfer function; wi(k) is the weight value at discrete time k, where i goes from 0 to m; xi(k) is where i goes from 0 to m input value at discrete time k, b bias). MAPE was used to compare ANNs and other methods. The smaller the MAPE values, the closer the estimated values to the true values (Benzer et al. 2017). MAPE is as in the following equation (Eq. 4). Yi is the actual observation value; ei is the difference between the true value and the predicted value; n is the total number of observations.Eq. 1 w:w−ε∂C∂w $$w:w - \varepsilon {{\partial C} \over {\partial w}}$$ Eq. 2 b:b−ε∂C∂b $$b:b - \varepsilon {{\partial C} \over {\partial b}}$$ Eq. 3 Y(k)=F∑i=0mwi(k)⋅xi(k)+b $$Y(k) = F\left[ {\sum\nolimits_{i = 0}^m {wi(k) \cdot xi(k) + b} } \right]$$ Eq. 4 MAPE=1n∑i=1neiYi×100……… $$MAPE{\rm{\;}} = {1 \over n}\sum\nolimits_{i = 1}^n {\left| {{{ei} \over {Yi}}} \right| \times 100 \ldots \ldots \ldots } $$

In the study, some water quality parameters such as water temperature, pH, dissolved oxygen, electrical conductivity, secchi disk, alkalinity, total nitrogen, and total phosphorus were measured during the sampling in the field. Accordingly, it was determined that the highest water temperature was 27.8°C in summer at the 2nd station, and the lowest was 3.2°C in the winter at the 2nd station. The highest pH was 8.5 in winter; the lowest 7.4 in summer at the second station. The maximum dissolved oxygen concentration was 8.8 mg l^-1 at the 1st station in winter, while the lowest concentration was 6.1 (mg l^-1) at the 2nd station. Conductivity in the maximum reading was 536 (μS cm^-1) in summer at the 3rd station; the lowest reading was 305 (μS cm^-1) in winter at the 2nd station. The highest secchi disk was 48.3 (cm) in summer at the 1st station; the lowest was 38.2 (cm) in winter at the 3rd station. The alkalinity was 320 (mg CaCo₃ l^-1) in summer, while the lowest was 120 (mg CaCo₃ l^-1) in autumn in the 3rd station. The highest total nitrogen was 3.4 (mg N l^-1) in summer at the 1st station; the lowest was 0.2 (mg N l^-1) in winter at the 3rd station. The maximum total phosphorus was 1.6 mg l^-1 at the 1st station in summer, while the lowest value was 0.01 (mg l^-1) at the 3rd station in winter (Figure 3).

Figure 3

Zooplankton structure also showed temporal changes. On average, rotifers were the major component in all the periods studied; cladocerans and copepods were determined during the seasons (Figure 4).

Figure 4

Actual values of the 1st, 2nd and 3rd station zooplankton density (rotifera, cladocera, copepoda) and artificial neural networks values are given according to the months in Table 1, Table 2, and Table 3. The actual values of zooplankton density and the results obtained from the artificial neural networks were compared. These values were calculated one by one. Mean absolute percent error (MAPE) values were calculated with actual values and ANNs values. ANNs values were determined to be close to the real data. When Table 1 is examined, MAPE (%) value was determined as 1.143 for Rotifer, 0.118 for cladocera, and 0.141 for copepoda. When Table 2 is examined, MAPE (%) value was determined as 0.941 for Rotifer, 0.377 for cladocera, and 0.185 for copepoda. When Table 3 is examined, MAPE (%) value was determined as 0.342 for Rotifer, 0.557 for cladocera, and 0.301 for copepoda.

ANNs results were obtained using training data and artificial neural networks. ANNs results regarding zooplankton densities for the 1st station, 2nd station, and 3rd station are given in Figure 5, Figure 6, and Figure 7. The best fit between targets and outputs is determined by the linear regression line. The R value determines the relationship between these targets and outputs. The targeted output R value is calculated as 0.99716 for training, 0.92436 for validation, 0.99642 for testing, and 0.94196 for all in the 1st station. The targeted output R value is calculated as 0.95697 for training, 0.99126 for validation, 0.52772 for testing, and 0.95631 for all in the 2nd station. The targeted output R value is calculated as 0.91557 for training, 0.93203 for validation, 0.98536 for testing, and 0.92585 for all in the 3rd station. R values close to 1 indicate the best results of the training (Figure 5-7).

Figure 5

Figure 6

Figure 7

Figures (8-10) show the corresponding validation checks and the gradient of epochs. For the training state of artificial neural networks model, the validation checks were attained as 6, at epoch 6 and gradient = 1668541.3655, and at epoch 10 in the 1st station. For the training state of artificial neural networks model, the validation checks were attained as 6, at epoch 6 and gradient = 455785.4862, at epoch 6 in the 2nd station. For the training state of artificial neural networks model, the validation checks was attained as 6, at epoch 6 and gradient = 3483291.2888, and at epoch 6 in the 3rd station.

Figure 8

Figure 9

Figure 10