Prediction of the Dissolved Oxygen by Using Multi-Layer Perceptron and KNN Approaches: A Case Study in Coosa River, Alabama, USA

Abstract

The dissolved oxygen amount of a water body, such as a reservoir, stream, or river, is an important water quality parameter that may affect society's health directly. The daily mean dissolved oxygen of the Coosa River was investigated in this presented study. The multi-layer Perceptron (MLP) approach and k-nearest neighbor (KNN) algorithm, recently widely used for hydrological and environmental problems, was chosen for the prediction. Daily water temperature (Max, Min, and Mean), daily mean specific conductivity, daily median water pH, and discharge parameters were inputs in the MLP and KNN models. A total of 3535 daily records were implemented into the model. 2951 daily data were used as the training set, while 584 was the test set. Results were compared with each other by using statistical evaluation methods. The KNN approach was also generated by applying the same training and test sets. Based on the results, it is evident that the MLP (Multilayer Perceptron) model provided satisfactory dissolved oxygen prediction results. However, the KNN (K-Nearest Neighbors) model outperformed the MLP approach, despite having a lower correlation coefficient than the MLP.