Predicting Road Infrastructure Deterioration with AI: A Multi-Source Data Approach

This thesis presents a data-driven framework for predicting the deterioration of road infrastructure, with a focus on modeling the progression of alligator cracking in terms of area and severity. This research effort brings together multiple sources of data, covering but not limited to climate, traffic patterns, and visual imagery of pavement, and makes use of transfer learning to adapt models based on the U.S. Long-Term Pavement Performance (LTPP) dataset to the context of Italy, where historical data is limited. The study was completed as part of a collaboration with LOKI s.r.l. through the Asfalto Sicuro project, and aims to characterize and automate the detection and monitoring of major pavement distresses using Artificial Intelligence (AI), space-based data, and on-vehicle sensors. Traditional empirical and mechanistic-empirical models cannot capture the nonlinear and spatiotemporal behavior of deterioration; thus, the research benefits from machine-learning approaches that can learn complex deterioration trends directly from data. The objectives will focus on characterization of data, predictive model development, transfer learning, and evaluation of models across domains. Ultimately, the research examines climatic and traffic variables and their role in the development of distresses, while assessing the transfer of U.S. trained models to road segments in Italy. In summary, the research highlights the promise of AI-based models and pavement management for facilitating more timely, accurate and cost-effective maintenance planning for pavements in areas where data availability is scarce.