Prediction of Nitrate Concentrations Using Neural Networks in Traditional Wells Capturing the Shallow Groundwater in M’Bahiakro Municipality (Central-Eastern, Côte D’Ivoire)

This study aims at predicting nitrate concentrations in wells capturing the M’Bahiakro phreatic groundwater from in-situ measurable physico-chemical parameters. The study was conducted using neural models based on the gradient error back-propagation learning method (BPNN). The configuration of the models was performed by a constructivist approach according to the R2 and MSE performance criteria using supervised learning. The MATLAB simulation code was run for this purpose to find the optimal model. Pearson correlation analysis (r) on the data set indicates that NO3− concentrations are significantly correlated with EC, O2, depth, Eh, and T (r >  ± 0.5) during the wet and dry season. These significant correlations revealed the contribution of these physico-chemical parameters in the process of NO3− contamination of well water and were selected as input variables for the BPNN models. The constructivist approach allowed to define of four types of BPNN models with a single hidden layer of 6 neurons and an output layer of one neuron, including BPNN1 (2-6-1), BPNN2 (3-6-1), BPNN3 and BPNN4 (4-6-1). The BPNN models of the 4-6-1 vector type developed gave the best performance during learning with MSE values around 0.01 (BPNN3: 0.068-0.466-0.378 and BPNN4: 0.221-0.013-0.001) and R2 close to 1 (BPNN3: 0.99; 0.90; 0.96 and BPNN4: 0.98; 0.99; 0.99) respectively for 70% of the training data, 15% of the test data, and 15% of the validation data during the dry and wet periods. Moreover, using the weights and biases acquired after the final learning phase, these developed 4-6-1 vector models were able to satisfactorily represent the experimentally obtained nitrate values in the studied well waters in the city of M’Bahiakro. The BPNN models (BPNN3 and BPNN4) of the 4-6-1 vector type were able to reproduce satisfactorily the experimentally obtained nitrate concentration data in the 19 wells of the M’Bahiakro municipality. Thus, through the application of these proposed models, it would be more cost-effective to monitor nitrates in the well water of M’Bahiakro in order to prevent adverse health effects on the population.