Effect of hyperparameter tuning on classical machine learning models in detecting potholes

Potholes are an increasing and persistent challenge plaguing the timely upkeep of vital road infrastructure. Millions of money are lost each year on repairing damages and using alternate routes with longer travel times resulting from potholes. Early, accurate, and frugal means of pothole detection have a significant role in improving the quality and safety of a road transport network. In recent years machine learning has received much attention in underpinning pothole detection systems. This has resulted in a plethora of machine learning-based detection systems with little agreement on which are the best performing. This paper compares six machine learning algorithms to determine the most suitable for pothole detection when using an online dataset. Additionally, the ideal hyperparameter tuning of each machine learning algorithm is determined. The experimental results in this study demonstrate that the hyperparameter adjustment of machine learning algorithms has varying effects on pothole detection. The KNN algorithm is the best-performing machine learning algorithm with hyperparameter tuning achieving 80%, 76%, 78%, and 77% respectively for accuracy, precision, recall, and F1-Score with an average runtime of 0.11 minutes. The lowest-performing machine learning algorithm is the NB algorithm which achieved an accuracy of 73%, precision of 66%, recall of 74%, and F1-Score of 69% with an average runtime of 0.01 minutes. Overall the machine learning algorithm with hyperparameter tuning has accuracy, precision, recall, and F1-scores closely correlated as compared to machine learning algorithms without hyperparameter tuning.