Civil and Environmental Engineering electronic theses and dissertations (MU)
Permanent URI for this collection
The items in this collection are the theses and dissertations written by students of the Department of Civil and Environmental Engineering. Some items may be viewed only by members of the University of Missouri System and/or University of Missouri-Columbia. Click on one of the browse buttons above for a complete listing of the works.
Browse
Recent Submissions
Item Integrating multi-task learning and large language models for advanced transportation and infrastructure applications(University of Missouri--Columbia, 2025) Owor, Neema Jakisa; Adu-Gyamfi, Yaw[EMBARGOED UNTIL 12/01/2026] Multi-task Learning (MTL) has emerged as a pivotal approach in machine learning, facilitating advancements across diverse fields by integrating multiple tasks within a single model. Unlike traditional models that handle tasks independently, MTL's joint learning approach allows interconnected tasks to leverage shared data features, enhancing generalization, accuracy, and computational efficiency--especially advantageous for edge devices with limited memory and processing capacity. By training various tasks simultaneously, MTL improves resource use and knowledge-sharing across tasks, making it ideal for complex applications requiring efficient processing. This study explores two main applications: Pavement Condition Assessment and Work Zone Safety through an Automated Audible Truck Mounted Attenuator (TMA) Alert System. For pavement condition assessments, MTL concurrently detects pavement distress types, severity, extent, and calculating the Pavement Condition Index (PCI), achieving resource efficiency and enhancing accuracy through shared knowledge. The Automated Audible Truck Mounted Attenuator (TMA) Alert System uses a MTL to simultaneously detect and classify vehicles, estimate distance zones (danger, warning, safe), and perform lane and road segmentation. The system identifies vehicles on a potential collision course with the TMA by assessing lane position, speed, acceleration, and estimated collision time, issuing alerts. This MTL approach enhances safety by providing real-time, proactive warnings for vehicles on collision course. The integration of Large Language Models (LLMs) significantly enhances a model's capabilities, adding layers of detailed human-like interpretation through advanced language processing and multi-modal analysis. LLMs leverage vast, diverse datasets and contextual understanding to support a comprehensive view of complex transportation and infrastructure scenarios. The combination of MTL and LLMs results in a powerful, adaptable system, excelling in computational efficiency, contextual insight, and precision, making it an invaluable tool for a wide range of transportation applications. The primary objective of this dissertation is to develop a multi-task learning framework and leverage Large Language Models advance infrastructure condition assessment and enhance transportation safety. The study's first objective focuses on developing a zero-shot segmentation model, PaveSAM, designed to detect pavement distresses that it was not explicitly trained on. Unlike current supervised segmentation models requiring costly pixel-level annotations, PaveSAM uses bounding box prompts to achieve effective segmentation with minimal input. This approach drastically reduces the time and labor involved in labeling pavement images, offering a new, efficient method for distress segmentation. By retraining SAM's mask decoder on just 180 images, PaveSAM shows that bounding box prompts are viable for pavement analysis, making it the first application of SAM in this domain. Additionally, this model enables the use of existing open-source distress images annotated with bounding boxes to generate segmentation masks, expanding the availability and diversity of pavement distress datasets. The second objective proposes a unified multitask model to directly predict the Pavement Condition Index (PCI) from pavement images by assessing distress type, extent, and severity simultaneously. Current PCI estimation models require multiple models, leading to increased complexity and resource demands. The proposed multitask model addresses this by using a single encoder for feature extraction and multiple decoders for specific tasks, including two detection heads, a segmentation head, and a PCI estimation head. This architecture enables efficient PCI estimation on edge devices, significantly reducing computational requirements and simplifying deployment. Tested on a benchmarked pavement distress dataset with polygonal labels, this model represents a pioneering approach by offering real-time PCI estimation with high accuracy. The study's third objective introduces an automated AI vision system to improve work zone safety by issuing real-time alerts to drivers on collision course with Truck-Mounted Attenuators (TMAs). This system autonomously detects, classifies vehicles, assessing collision risks and activating alerts to eliminate manual intervention and improve safety. Using MTL, the model integrates vehicle detection, distance zone classification, lane segmentation, and road segmentation tasks. A Generalized Efficient Layer Aggregation Network (GELAN) backbone enhances the system's performance, while an alert module triggers warnings based on speed, acceleration, and time to collision. Achieving high recall (90.5%), mAP (0.792) for vehicle detection, mIOU (0.948) for road segmentation, and accuracy rates over 80% for lane segmentation and distance classification, this system demonstrates robust performance in preventing TMA collisions. The fourth objective aims to develop a Large Language Model (LLM) framework that revolutionizes infrastructure condition assessments by providing context-sensitive, human-like interpretations of infrastructure imagery. This approach significantly surpasses traditional image detection models by combining visual analysis with natural language understanding. The proposed framework employs fine-tuned vision-language models (VLMs) to automate pavement condition evaluation through comprehensive analysis of visual inputs and generation of descriptive language outputs. The system identifies and characterizes key aspects of pavement distress including distress type, severity level, and extent while estimating both quantitative Pavement Condition Index (PCI) scores and generating detailed qualitative descriptions of the distresses. To optimize model performance, we implement a sophisticated fine-tuning strategy that incorporates four key innovations: instructional prompt diversification for robust context handling, bounding box grounding for precise spatial localization, PCI scale normalization for consistent scoring, and extended Low-Rank Adaptation (LoRA) applied across both transformer and cross-modal layers for efficient parameter optimization. The optimized model demonstrates great performance, achieving a METEOR score of 0.8989 and ROUGE-L score of 0.9023 for descriptive accuracy, alongside numerical precision with a mean absolute error (MAE) of 1.19 and root mean square error (RMSE) of 1.89 for PCI predictions. This research establishes the viability of vision-language models for scalable, interpretable, and highly accurate pavement assessment applications. Overall, the proposed models offer a resilient, cost-effective, and automated approach to road infrastructure management, significantly advancing pavement distress detection, PCI estimation, and work zone safety.Item Towards advanced sustainable technologies for asphalt pavement materials(University of Missouri--Columbia, 2025) Gettu, Nandita; Buttlar, William G.[EMBARGOED UNTIL 12/01/2026] The transportation sector remains one of the largest contributors to the greenhouse gas emissions in the United States, per economic sector, accounting for 28% of the total emissions nationwide. In particular, the asphalt industry holds major responsibility in the transportation infrastructure sector as 94% of all paved roads in the U.S. are asphalt overlays or full-depth pavements. Therefore, advancing technologies that reduce the material-level carbon emissions while ensuring structural performance is a critical step toward national sustainability goals. This dissertation addresses the need to develop a comprehensive and inclusive asphalt mixture design framework that incorporates mixture resistance to pavement distresses and associated environmental impacts. This framework is subsequently integrated into a decision-support system for efficient and feasible optimization of pavement sustainability and engineering performance. The research is conducted around four core objectives: (1) identifying mixture parameters that critically influence the Global Warming Potential (GWP) of asphalt mixtures; (2) establishing a GWP benchmarking methodology and GWP thresholds for practical implementation, (3) evaluating performance implications and the environmental effects of incorporating recycled materials such as reclaimed asphalt pavement, engineered crumb rubber and post-consumer polymers into routine asphalt mixtures; and (4) developing an adapted, LCA-integrated Balanced Mix Design (BMD) framework that enables practitioners to meet both performance criteria and environmental thresholds simultaneously. Extensive analyses of asphalt Environmental Product Declarations (EPDs) were conducted to identify the key factors influencing GWP during the material production stages. Findings revealed that the material acquisition phase is the dominant contributor to total emissions, largely driven by asphalt binder production and content. Specialty mixtures and mixtures with polymer-modified binders were also found to have increased environmental impacts. Furthermore, the study highlighted the necessity for establishing standardized GWP thresholds to objectively define what constitutes an acceptable environmental performance for asphalt mixtures. To address the limitations of existing GWP benchmarking methodologies, a new application-based benchmarking approach was developed. The proposed benchmarks provide phase-specific and mix-type or application-specific GWP thresholds that reflect real-world contractor practices, local specifications, and mixture production trends. This creates a more implementable system, supporting potential use in procurement incentives, material qualification processes, and agency-wide sustainability programs. Laboratory performance testing was conducted to assess the feasibility of incorporating recycled materials into routine and high environmental-impact asphalt mixes i.e., stone mastic asphalt and polymer-modified mixes. The mixes were developed through the Balanced Mix Design (BMD) methodology where the performance tests were applied to characterize the mixes across key pavement distresses. Specifically, rutting susceptibility was characterized via the Hamburg Wheel Tracking Test, intermediate-temperature cracking resistance was assessed using the IDEAL-CT procedure, and low-temperature fracture performance was quantified through the Disk-Shaped Compact Tension (DC(T)) test. rutting (Hamburg Wheel Tracking Test), intermediate-temperature cracking (IDEALCT test), and low-temperature cracking (Disk-shaped Compact Tension test). Moreover, laboratory investigations assessed strategies to reduce GWP through mix design optimization and material selection. The incorporation of reclaimed asphalt pavement (RAP) was identified as the most effective mitigation measure, followed the use of asphalt binders with lower GWP impacts and finally, lowering binder content through the use of recycled polymers such as post-consumer polyethylene. The outcomes of this work culminated in the development of a comprehensive decisionsupport framework that integrates LCA and BMD principles, enabling simultaneous evaluation of mechanical and environmental metrics. This framework provides a scientifically defensible tool for policymakers, engineers, and contractors to design asphalt mixtures that optimize both performance and sustainability, thereby advancing state-of-the-art pavement engineering practices.Item Calibration of Nortek Signature 1000 echosounder for bubble measurement in controlled laboratory conditions(University of Missouri--Columbia, 2025) Reasad, Mustahsin; Wang, BinbinThis thesis presents a laboratory-based calibration and validation of the Nortek Signature 1000 echosounder for measuring bubble Target Strength (TS) in freshwater environments. Calibration was performed using Tungsten Carbide (WC) spheres with known acoustic properties to determine system gain, applying theoretical models and sonar equation-based corrections for attenuation and background noise. A controlled experiment was developed to release discrete air bubbles, which were captured using a high-speed shadow imaging system. By synchronizing the optical and acoustic measurements, bubble diameters were directly compared with corresponding acoustic TS values. Results showed a consistent relationship between bubble size and acoustic backscatter intensity measured by the echosounder, with strong agreement against optical validation in the 4–6 mm diameter range. The findings demonstrate that, with proper calibration, the Nortek Signature 1000 can effectively estimate bubble TS. This approach provides a useful method for improving acoustic measurements of gas bubbles in both laboratory and field studies.Item Investigation of water quality in artificial hellbender shelters(University of Missouri--Columbia, 2025) Holzer, Jack Nicholas; Trauth, KathleenHellbender salamanders have been facing population decline due to loss of natural habitat. Concrete shelters have been developed to provide additional nesting habitat. The first concrete boot-shaped artificial hellbender nestboxes (Briggler and Ackerson 2012) could be dislodged due to drag forces from large flows. Mohammed et al. (2016) created a hydrodynamically redesigned nestbox to improve stability by reducing drag forces. Even with greater hydraulic stability, many nest boxes were still not being used. A two-part hypothesis is that poor water quality inside the box may be responsible. If circulation of water into the box is limited, then dissolved oxygen (DO) may decrease over time. To test this hypothesis, an experimental plexiglass flume was used. Five pumps provided a range of flow velocities. Two half-scale models of the hydrodynamically redesigned nestbox were constructed, one with potential reaeration holes in the upstream face and one without. To experimentally assess the concentration of DO in the nest boxes, DO sensors were placed in each half-scale model. Results of these experiments show that absent any reaeration holes, DO in the upstream most portion of a shelter is lower than in the downstream location near the entrance. Multiple holes on the upstream face increase the DO, as do placing the holes on either side of the leading edge. Even with the addition of a DO scavenger, the holes were able to reaerate the system. In summary, the addition of one or more reaeration holes in the upstream face of the shelter creates flowthrough, increasing DO, even after the addition of an oxygen scavenger.Item Early age and time-dependent behaviour of ternary blends of high-performance concrete(University of Missouri--Columbia, 2025) Earney, Timothy Patrick; Salim, Hani A.This dissertation investigates how to proportion high-performance concrete (HPC) mixtures that achieve necessary concrete properties without incurring undue early-age and time-dependent deformations. Motivated by broad adoption of low water-to-binder (w/b) concretes with supplementary cementitious materials (SCMs) such as silica fume and Class C fly ash, the work studies time dependent deformation and durability performance of concrete. Twenty-four mixes systematically vary w/b and binder composition. Autogenous shrinkage is measured immediately after casting using a new method developed for the study. Chloride penetration is studied using the Rapid Chloride Penetration Test. Results show that w/b and SCM selection jointly govern autogenous shrinkage and permeability. Lowering w/b increases autogenous shrinkage; silica fume alone tends to amplify it, while fly ash tends to reduce it. Judicious ternary combinations offset shrinkage amplification while maintaining low permeability. A field study of instrumented prestressed I-girder bridges integrates these earlyage insights at structural scale. An analysis of stresses from prestress transfer, curingstage thermal gradients, autogenous shrinkage, in-service thermal cycles, and service loads shows that their superposition indicates likely cracking near girder ends. A decision method (ELECTRE III) is shown to be effective for selecting appropriate mixtures to satisfy targeted performance goals. It was found that the method will produce recommendations that are supported by the available data so long as the data is complete enough.