Deep learning for estimating pavement roughness using synthetic aperture radar data

Because of the high costs of ground-based pavement condition methods used to monitor pavement condition, transportation agencies often limit distress surveys to their major roads. As a result, the condition of local and ancillary roads remains unknown to decision-makers. This study addresses this gap by exploring the capabilities of publicly available Synthetic Aperture Radar (SAR) data to estimate pavement roughness. This paper introduces a novel framework to address the challenges of using SAR images in evaluating pavement condition. The trunk highway network in Minnesota is analyzed to develop deep learning models that predict International Roughness Index (IRI) and associated prediction intervals. This analysis found that SAR images have a strong potential in quantifying pavement condition. The deep learning models were able to predict IRI with a mean absolute error of 14.6 in./miles and provide intervals of pavement condition that capture actual IRI values with an accuracy of 81%.

Automated summary

Due to high costs, transportation agencies often limit pavement condition surveys to major roads, leaving local and ancillary roads unknown. This study explores the use of publicly available Synthetic Aperture Radar (SAR) data to estimate pavement roughness and proposes a deep learning model to predict International Roughness Index (IRI) with associated prediction intervals. The analysis on the trunk highway network in Minnesota shows that SAR images have strong potential in quantifying pavement condition, achieving an accuracy of 81% in capturing actual IRI values.