Comparative Analysis of Parametric & Non-parametric ML Models for Pavement Deterioration Prediction

Abstract: Accurate modeling of pavement condition deterioration and the impact of maintenance treatments are essential components of modern pavement management systems (PMS). Deterioration curves provide a predictive framework for pavement performance over time, allowing municipalities to allocate resources more effectively and ensure optimal service levels. The complexity of these models stems from the interaction of a range of factors, including pavement type, environmental conditions, traffic loads, and the type of treatments applied.
In this study, we derived pavement condition index (PCI) deterioration curves and quantified treatment impacts using two complementary approaches: parametric regression, represented by a Gompertzian model, and random forest. As a traditional modeling approach, parametric regression models are widely used in PMS, mainly due to their simplicity and ability to produce interpretable results. However, they are inherently limited in their ability to capture complex, non-linear relationships and interactions between variables, particularly when data patterns deviate from assumed functional forms. To address this limitation, we also implemented a random forest model, a robust, non-parametric machine learning technique known for its ability to handle high-dimensional data and uncover non-linear interactions.
The analysis was conducted using pavement condition data from the City of San Diego, which includes three pavement types: asphalt improved, asphalt unimproved, and jointed concrete. The comparison of the two approaches yielded valuable insights. The Gompertz regression model performed adequately for general trend estimation and provided the advantage of interpretability, making it easier to explain the relationships between pavement age, condition, and treatment impacts to stakeholders. On the other hand, the random forest model outperformed the parametric model in terms of prediction accuracy, particularly for complex cases with multiple interacting factors. This advantage, however, comes at the cost of transparency and interpretability compared with the traditional parametric models.
This study contributes to the growing body of knowledge on the use of machine learning in infrastructure asset management by demonstrating the practical benefits and trade-offs of each approach. While parametric models are well-established in PMS, the use of advanced machine learning techniques offers potential for improving the precision and reliability of deterioration predictions, especially when integrated with large, diverse datasets. The combination of these approaches, leveraging the strengths of each, represents a promising direction for future pavement management practices, allowing agencies to make more informed, data-driven decisions regarding pavement preservation and rehabilitation.