Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Towards a More Resilient Condition Assessment of Road Infrastructure using Machine Learning and Satellite Imagery
The rapid expansion of Egypt’s road infrastructure presents significant challenges in terms of maintenance and monitoring. These difficulties are mainly due to the lack of current data on road conditions and limited resources to gather new data. In such contexts, relying on traditional survey-based...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
AUC Knowledge Fountain
2026
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | The rapid expansion of Egypt’s road infrastructure presents significant challenges in terms of maintenance and monitoring. These difficulties are mainly due to the lack of current data on road conditions and limited resources to gather new data. In such contexts, relying on traditional survey-based approaches for assessing road conditions proves to be both costly and time consuming. As a result, there is a growing need for more efficient and accessible methods to address the maintenance burden imposed by the expanding network. The objective of this thesis is to develop a resilient, data-driven framework that predicts pavement condition using open-access satellite imagery integrated with low-cost International Roughness Index (IRI) measurements, enabling local road authorities to prioritize maintenance more efficiently.
To meet this objective, the study focuses on a 1405‑segment network in New Cairo, Egypt, covering arterial, collector, and local roads, for which IRI readings were collected via a smartphone-based application and matched with Google Earth Pro satellite images. Five convolutional neural network architectures—LeNet, EfficientNet, VGG‑16, AlexNet, and ResNet‑50—were trained and evaluated under three‑category and binary condition rating schemes across three datasets (482, 809, and 1405 segments) to analyze the influence of data volume and model choice on predictive performance.
The results show that the best three‑category model (VGG‑16) achieved a test accuracy of about 64%, whereas the best binary models (ResNet‑50 and AlexNet) reached approximately 77% accuracy, indicating greater robustness for binary classification under the available data quality. These findings suggest that open-access satellite imagery combined with smartphone collected IRI measurements can deliver performance comparable to studies using commercial high-resolution satellite Imagery more accurate IRI measurements. Moreover, the study recommends that binary models can be utilized as a practical decision-support tool while pursuing future enhancements through higher-resolution satellite imagery, expanded and more variable datasets, and more advanced hybrid deep learning architectures. |
|---|