Abstract
dc:description.abstractAccurate, comprehensive, and consistently maintained roadway feature data are essential for roadway safety analysis, operational studies, asset management, and compliance with national requirements such as the Federal Highway Administration’s (FHWA) Model Inventory of Roadway Elements (MIRE). However, many roadway elements remain inconsistently collected or entirely missing across statewide databases. Traditional field surveys and mobile mapping collections are costly and impractical for large-scale, frequent updates. At the same time, existing automated extraction studies often emphasize model training but lack a complete workflow that integrates data preprocessing, spatial post-processing, and validation at statewide scale.This dissertation develops an end-to-end framework for large-scale roadway element extraction by integrating multi-source data, deep learning models, and GIS-based spatial reasoning into a unified and scalable pipeline. Using application of Nevada, the dissertation makes four primary contributions: (1) it introduces a statewide intersection type extraction method using road-map imagery and YOLOv5 detection to produce consistent and comprehensive intersection inventories; (2) it proposes an enhanced YOLOv5 architecture with dual-channel attention and adaptive feature fusion for detecting midblock crosswalks from 2.4 million aerial images, addressing the challenges of small object size, faint markings, and complex background patterns; (3) it develops a LiDAR–imagery fusion approach for median-type classification that combines cross-sectional profiling, geometric and radiometric feature engineering, and supervised learning; and (4) it establishes a robust post-processing framework—including pixel-to-coordinate transformation, roadway-constrained filtering, orientation analysis, duplicate consolidation, and centerline-based smoothing—that ensures geographic precision and substantially reduces false positives across all three roadway-feature extraction tasks. When applied at the statewide scale, this integrated workflow produced a substantially improved roadway dataset for Nevada, identifying 79,824 intersections, 770 midblock crosswalks, and validated on 6705 sample points manually check dataset across the entire state. When compared with existing records from NDOT and OpenStreetMap, the results demonstrated the framework’s ability to correct outdated entries, remove erroneous features, close long-standing data gaps, and add previously undocumented roadway elements. The proposed median detection framework achieved an overall classification accuracy of 96.1% on an independent, manually validated dataset, demonstrating strong performance across diverse roadway conditions. These improvements highlight the framework’s practical value for modernizing statewide roadway data, including both geometric and median-related attributes, and supporting reliable, data-driven roadway asset management. Overall, this dissertation demonstrates a practical, scalable, and cost-effective solution for statewide roadway data automation. The methodology is feature-agnostic and can be extended to other roadway elements such as lane markings, TWLTL medians and turn lanes—providing a foundation for future digital-infrastructure development, roadway safety analysis, and compliance with the MIRE data requirements.
Degree
thesis:*- Level thesis:degree_level
- Doctorate Degree
- Department thesis:degree_department
- Civil and Environmental Engineering
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wang, Ziru
- Advisor dc:contributor.advisor
-
- Xu, Hao
- Committee members dc:contributor.committeemember
-
- Tian, Zong
- Park, Seri
- Kelley, Scott B
- Feng, Jia
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en_US, English
Identifiers
dc:identifier.*- Repository record dc:identifier.uri
- https://scholarwolf.unr.edu/handle/11714/11908