Graduate Thesis Or Dissertation


Leveraging Satellite Data and Machine Learning to Enhance Pavement Condition Assessment Public Deposited
  • This research aims at establishing the feasibility of applying satellite data and machine learning (ML) to pavement applications as a way to envision the future of infrastructure asset management. The motivation for this study stems from several knowledge gaps including the assumption of accurate observations in pavement maintenance decision-making, the exclusion of low priority roads from annual distress surveys due to cost constraints, and the lack of quantitative evidence supporting the effectiveness of machine learning algorithms in modeling pavement performance. To address these gaps, the following specific objectives were defined: (i) quantify the value of uncertain optical satellite imagery in the pavement maintenance decision-making, (ii) evaluate the performance of machine learning algorithms in predicting pavement condition, and (iii) apply deep learning to publicly available satellite data to cost-effectively estimate pavement condition. Two satellite data types (i.e., high-resolution multispectral and Synthetic Aperture Radar (SAR) imagery) and different machine learning techniques including partially observable Markov decision process (POMDP) and deep learning were investigated in the context of evaluating pavement condition and making maintenance decisions. The optimal maintenance policies simulated using POMDP models show that satellite observations result in up to 6.5% reductions in cost over the pavement life cycle. The meta-analysis of existing literature indicated that machine learning algorithms can capture on average 15.6% more variability in International Roughness Index (IRI) than traditional techniques. Artificial Neural Network model is recommended to model IRI because of its consistent performance over a significant number of studies with varying sample sizes and data sources. The framework introduced to model IRI using SAR data was found to be highly effective in addressing the challenges of removing traffic noises from pavements, suppressing speckles without comprising the road features, and eliminating the effects of terrain on SAR backscatters. The resulting deep learning model resulted in accurate IRI predictions with mean absolute errors ranging from 13.9 to 14.6 inches/mile. The associated prediction intervals were found to capture 81% of the actual IRI values within their upper and lower limits. The developed framework was packaged as a software with a graphical user interface to facilitate its implementation by transportation agencies.
Date Issued
  • 2022-07-25
Academic Affiliation
Committee Member
Degree Grantor
Commencement Year
Last Modified
  • 2022-09-14
Resource Type
Rights Statement