Pothole prediction based on machine learning and pavement condition indicators

Potholes are dangerous defects on road surfaces, contributing to numerous crashes involving vehicles, motorcycles and bicycles. They also impose a significant economic burden on highway authorities. Currently, no effective tool predicts the number and location of potholes in a road network. In this work, an attempt to address this gap was conducted by developing novel pothole prediction tools built using two machine learning methods – random forest and K-nearest neighbour. The final prediction model requires nine pavement condition indicators, quantifiable through Surface Condition Assessment for the National Network of Roads surveys, commonly conducted in the UK. This unique approach allows for direct implementation by highway authorities. The model was trained on a large dataset of pavement condition and pothole data covering the Transport for London network. Validation results suggest that the model successfully predicted 55.5% of sections with potholes and 99.6% of sections without potholes. Although the model demonstrates limited accuracy in predicting potholes, recommendations are provided to enhance its performance. This work holds the potential to significantly aid strategic financial planning for pavement management in the UK and elsewhere.