Modeling the Turbidity Level of Asejire Reservoir Using Machine Learning Techniques

In this study, five machine learning techniques were employed to model water turbidity in the Asejire reservoir, utilizing Color, pH, Dissolved Oxygen, Alkalinity, Total Hardness, Calcium Hardness, Chloride, Iron, Silicon, Solids, Dissolved Solids, and Suspended Solids as independent variables. Our analysis reveals a robust positive correlation between turbidity and chemicals like Iron and Silica, alongside weaker positive correlations with Color, pH, dissolved oxygen, and dissolved solids. We observed a moderate negative relationship between turbidity and alkalinity, and weaker negative relationships with total hardness, calcium hardness, chloride, total solid, and total suspended solids. A strong positive correlation is also observed between silica and iron, dissolved solids and total solids, as well as total suspended solids and dissolved solids. The procedures carried out during the analysis included splitting the data into training and test datasets and standardizing the data after conducting exploratory data analysis. Five algorithms (including Extreme Gradient Boosting, Extra Tree Regression, K-nearest neighbors Regression, Random Forest, and Decision Tree Regression) were considered. Three metrics, namely mean absolute error, coefficient of determination, and mean squared error, were adopted to check for the accuracy of the algorithms. Four out of the five algorithms (Extreme Gradient Boosting, Extra Tree Regression, Random Forest, and Decision Tree Regression) performed best, accounting for over 95% of the variation in the training dataset. However, for the test dataset, Extra Tree Regression outperformed other considered algorithms since it had the highest R
²
and the least MAE and MSE. Additionally, we found that iron and silica are the most important physico-chemical properties using Extra Tree Regression.