14277 research outputs found
Sort by
Pothole detection using machine learning and computer vision techniques
No full textRoad infrastructure plays a crucial role in ensuring safety and efficiency in transportation systems. However, road surface defects, such as potholes and cracks, cause significant risks to vehicles and drivers while increasing maintenance costs. Traditional inspection methods are often time-consuming, hazardous, and inefficient, emphasizing the need for automated approaches to road condition monitoring. This research investigates advanced computer vision and machine learning techniques for detecting and segmenting road defects, providing a robust and efficient solution for road maintenance. The thesis begins with a comprehensive review of current vision-based methods for pothole detection, including traditional 2D image processing, 3D point cloud techniques, machine learning, and hybrid approaches. The review highlights that combining traditional and advanced machine learning methods offers superior accuracy and adaptability for road defect detection. Building on this foundation, this research evaluates the performance of state-of-the-art deep learning models, such as UNet, UNet++, DeepLabv3+, and PSPNet, for semantic segmentation using dashboard camera images. The results illustrate that UNet++ with an EfficientNet-B5 backbone outperforms other models, achieving higher accuracy and reliability in segmenting road defects. By combining in-depth analysis of existing techniques with the development of high-performing deep learning models, this thesis contributes to the design of effective systems for automated road surface condition assessment. These advancements aim to enhance roadway safety, reduce maintenance costs, and improve infrastructure management
The palaeoethnobotany of 16th-century Ferryland (CgAf-02): an entanglement of Beothuk and European migratory fishers
No full textFerryland (CgAf-02) is regarded as an opportune site for investigating the nature of interactions
between the Beothuk and European migratory fishers in Newfoundland during the 16th century due
to the identification of both Beothuk and European migratory fisher material culture within the
same contexts. Investigation of the palaeoethnobotanical record has been proposed as the most
likely means to this end because non-native grape seeds were recovered in association with hearths
consistent with Beothuk construction. While several sediment samples excavated from these
contexts underwent preliminary macrobotanical analyses over the course of almost three decades,
the question of Beothuk-European interactions at Ferryland remains unanswered. This thesis re-
evaluates the suitability of the sediment samples and their macrobotanical contents for informing
on this question by revisiting previous analyses, conducting new macrobotanical analyses,
amalgamating and standardizing data produced during previous and new analyses, and representing
that data spatially. The data is situated within a critical and theoretical discussion of archaeologies
of identity and site and feature formation processes to demonstrate that rather than seeing 16th-
century Ferryland as representative of interactions between two parties, it can be more productively
understood as a shared and entangled colonial space
Advanced deep learning methods for remote sensing data analysis: hyperspectral images and global navigation satellite system reflectometry data
No full textThis research develops state-of-the-art deep learning models for remote sensing data
analysis, focusing on the classification of hyperspectral images (HSI) and the retrieval of sea
surface wind speed and wave height from global navigation satellite system reflectometry
(GNSS-R) data.
In the task of HSI classification, a novel Multiscale Neighborhood Attention Transformer
(MSNAT) is first proposed to address the limitations of conventional Transformer-based
models in capturing multiscale spatial features. Subsequently, a Dual Frequency Transformer
Network (DFTN) is developed to rethink HSI classification from a frequency-domain
perspective. In addition, a novel and effective 3D Sharpened Cosine Similarity (3D SCS)
operation is introduced as an alternative to conventional 3D convolutions. Building on this
operation, a 3D SCS neural network (3D SCS-NN) is constructed to enhance classification
accuracy. The proposed methods are evaluated on three widely used benchmark datasets,
i.e., University of Pavia (UP), University of Houston (UH), and University of Trento (UT),
and the results illustrate their effectiveness and generalizability in HSI classification tasks.
Specifically, the MSNAT achieves overall accuracies of 93.34%, 86.26%, and 96.63% on the
UP, UH, and UT datasets, respectively, and the DFTN further improves performance, attaining
overall accuracies of 94.14%, 86.92%, and 96.72% on the same datasets. In addition,
3D SCS-NN achieves overall accuracies of 88.41%, 81.84%, and 95.56% on the UP, UH, and
UT datasets, respectively, achieving improvements of 0.93%, 1.54%, and 0.43% compared
with the vanilla 3D-CNN.
For global sea surface significant wave height (SWH) and wind speed retrieval from
GNSS-R data, a transformer-based network, WaveTransNet, is proposed to estimate global
SWH. To leverage the complementary strengths of convolutional neural networks (CNNs)
and transformer architectures, a hybrid CNN-Transformer network (CTN) is developed to
improve the accuracy of wind speed retrieval. In addition, a novel weighted mean squared
error (MSE) loss function is introduced to mitigate underestimation at high wind speeds.
Moreover, a physics-informed attention-aided CNN model (PA-CNN) is designed to incorporate
both attention mechanisms and physical constraints. All proposed methods are
validated using data from the Cyclone Global Navigation Satellite System (CYGNSS) and
the National Data Buoy Center (NDBC), and the experimental results demonstrate their
superior effectiveness in GNSS-R–based ocean remote sensing tasks. More specifically, Wave-
TransNet achieves SWH estimation with a root mean square difference (RMSD) of 0.443 m
against ERA5 data and 0.444 m against NDBC buoy observations, demonstrating consistent
accuracy across different reference data. For wind speed retrieval, CTN attains an overall
RMSD of 1.42 m/s when compared with ERA5 and 1.62 m/s against buoy measurements,
while PA-CNN with spatial-temporal smoothing achieves RMSDs of 1.38 m/s and 1.56 m/s,
respectively, outperforming existing deep learning baselines. For high wind speed estimation,
the incorporation of physical constraints reduces the RMSD for wind speeds above 15
m/s from 4.94 m/s to 4.59 m/s, and the use of weighted MSE loss further lowers the RMSD
for high wind speed retrieval from 4.952 m/s to 3.945 m/s, with an improvement of 20.3%
A study in orthogonal Latin Squares and strong starters
No full textCombinatorial designs are powerful mathematical frameworks that solve problems involving arranging elements within a set according to specific constraints. Among this field's most widely studied structures are Latin Squares1 and Room Squares2.
This research explores the intricate structures of Latin Squares, Room Squares, and orthogonal starters3 within combinatorial design. By employing techniques such as the Kronecker product, the study systematically generates higher-order Mutually Orthogonal Latin Squares4 (MOLS) and explores the concept of 3D-orthogonality, which demonstrates significant potential for advancing our understanding of complex combinatorial structures in the context of orthogonal Latin squares, which is one of the primary focuses of this research, along with the orthogonal starters5.
The research also delves into the practical applications of these combinatorial designs (Latin and Room Squares), highlighting their importance in fields such as experimental design, cryptography, and tournament scheduling. Studying starters and orthogonal starters, particularly in the context of Room squares, offers new insights into their construction and utility.
Methodologically, the research combines mathematical rigor with systematic search techniques, including hill-climbing algorithms and exhaustive search with backtracking, to address the challenges of large solution spaces. Also, identifying orthogonal starters and advancing lower bounds for specific orders, notably n = 21, 33, 35, and 39, these mark significant contributions to the field.
Overall, this research provides both theoretical advancements by exploring the 3D-orthogonality and practical solutions, such as new findings, that pave the way for further exploration in combinatorial design. The findings offer a solid foundation for developing new algorithms and strategies for generating and analyzing critical mathematical structures within this domain.Includes bibliographical references (page 66
The “Greatest Frontier Days’ Celebration Ever Attempted”: shaping civic and regional identity in the Calgary Stampede
No full textThe Calgary Stampede, presently marketed as “The Greatest Outdoor Show on Earth,”
has been the subject of scholarly attention from various disciplines ranging from veterinary
medicine to sociology and history. This study examines several aspects of the event, with the aim
of examining the ways in which the Stampede has been presented to the public, and especially
the ways in which that presentation has been adapted to fit changing cultural norms and
responses to public scrutiny. Furthermore, this study examines how the Stampede has shaped
civic and regional identities. This study looks thematically at the representation of Indigenous
peoples, masculinity, and non-human animals in the event. It further examines how the Stampede
organization has shaped the event’s marketing around history, tradition, and spectacle. This
study uses historical records examined through frameworks such as Judith Butler’s theory of
gender performativity, Guy Debord’s theory of the spectacle, and Mary Louise Pratt’s notion of
contact zones to explain the event’s continued popularity. The result is a series of thematically
organized chapters which illustrate the ways in which the seemingly static aspects of the
Stampede are adjusted to suit ever-changing cultural expectations and norms while maintaining
an air of historicity that links the event to its origins. The competing goals of rooting the event in
the province’s past and keeping it relevant to contemporary audiences necessitate a continuous
renegotiation of the attractions, marketing, and aesthetics of the Stampede, along with a
simultaneous focus on historical moments from its past.Includes bibliographical references (pages 244-260
Eigenvalue spectrum of marginally outer trapped surface stability operator in the Weyl-distorted Schwarzschild black holes
No full textMarginally outer trapped surfaces (MOTS) are closed spacelike surfaces from which
outgoing light rays neither converge nor diverge. In recent years they have been found
to be a key tool for understanding black hole geometries. In particular, the stability
operator provides information on whether the marginally outer trapped surfaces
(MOTS) bounds a trapped region. This study investigates the eigenvalue problem
associated with the stability operator for MOTS in the context of Weyl-distorted
Schwarzschild solutions. By solving the eigenvalue problem, we aim to understand
whether these solutions can always be understood as black holes.Includes bibliographical references (pages 52-56
Impacts of oil pollution on sub-Arctic and Arctic marine invertebrates: a study on blue mussels and beyond
No full textOil pollution from human activities continues to be a significant threat to marine
animals despite advancements in the safety of oil extraction and transportation. Oil
spills and their effects are highly complex and unique, and much is yet to be
understood about some of the major exposure pathways and the toxic impacts of
compositional changes of oil after it has been spilled. This dissertation aimed to
expand our knowledge regarding exposure pathways that had previously been
severely understudied or overlooked entirely.
First, I explored the potential differences in toxicity between non-biodegraded
and biodegraded oil to the blue mussel, Mytilus spp. Biodegradation by microbes is a
process that changes the composition of oil and can lead to a full mineralization of
those compounds that are accessible to the microbes. The results obtained from this
experiment paint a complex picture, and challenge the generally held belief that
biodegradation could lead to a decrease in oil toxicity.
Second, I explored the eIects of oil on blue mussels, and specifically how these
eIects change depending on the exposure pathway. Oil can get trapped in, or
adsorbed to, marine snow, which are aggregates consisting of (dead) phytoplankton
cells, fecal pellets, fibers, and minerals, and which are responsible for the downward
transport of organic matter to the ocean floor. The question posed in this study was
whether the eIects of oil would change depending on the exposure route, i.e. when it
is taken up as droplets or trapped within a natural food source, marine snow. I found
that marine snow seems to moderate the effects of oil on blue mussels, compared to
the uptake of free droplets. This was unexpected, but supported by all endpoints
measured (clearance rate, condition index, and measured DNA damage). Possibly,
the nutritional value of the aggregates, or simply the concentration of food particles
outbalance any negative eIects from the oil.
Finally, I conducted a meta-analysis of oil toxicology studies using invertebrates
from the arctic and sub-arctic regions, to explore how the vast knowledge we have
gathered regarding oil toxicity could be consolidated. Such a meta-analysis had not
been attempted before due to serious challenges in comparing methods and results
across oil toxicology studies. Even though these challenges persist and led to many
studies not being usable in a quantitative meta-analysis, I could still elucidate trends
that are worth exploring. For example, I found that ecologically relevant endpoints
tend to be relatively insensitive to oil pollution, while endpoints that might not be as
relevant for a wider population can be highly sensitive markers of oil exposure or
effects.
Together, the results from this dissertation add to our understanding of some of
the impacts of oil pollution on marine organisms, but highlight the complexities of oil
toxicology research and the need to increase comparability across studies.Includes bibliographical reference
The representation of synoptic-scale cyclone climatology in CMIP6 models over Atlantic Canada using self-organizing maps
No full textThis study compares the ability of selected CMIP6 Earth System Models (ESMs) to simulate synoptic-scale cyclone climatology over Atlantic Canada by examining their representation of pressure anomalies against that of ERA5 reanalysis data. A Self-Organizing Map (SOM) was trained using daily mean sea level pressure (MSLP) anomalies from both historical model simulations and ERA5 reanalysis data for the warm and cold seasons over a 45 year period. The ESMs are assessed based on their ability to reproduce the frequency and spatial structure of synoptic pressure anomaly patterns using both statistical and visual inspection techniques. Results indicate notable inter-model variability, with MPI-ESM1-2-HR and MRI-ESM2-0 showing relatively stable performance across both seasons, NorESM2-MM excelling in the cold season but declining in the warm season, and CMCC-CM2-HR4 exhibiting the highest biases overall. In general, models captured cold-season patterns more faithfully than warm-season ones. This research contributes an understanding on how to compare ESM biases in simulating cyclone climatology and also informs future efforts in regional climate impact assessment in the region surrounding Newfoundland
New perspectives on gas adsorption and diffusion in solid porous systems: integrating numerical modeling and machine learning
No full textThe escalating levels of atmospheric CO₂ necessitate efficient and sustainable solutions for its
capture and sequestration. This thesis investigates the theoretical and practical aspects of gas
diffusion and adsorption in KOH-treated activated carbon and Biomass Waste Derived Porous
Carbon, employing advanced mathematical modeling and machine learning techniques. The
objective is to contribute to the development of sustainable technologies to mitigate climate change
impacts. In this work, we correlate diffusivity with the uptake rate and physical properties of both
gas and solid phases, integrating these factors into the governing diffusion equations. Through
an extended Fick’s Law model, we develop a new mathematical framework that predicts gas
adsorption behaviors under diverse operational and thermodynamic conditions. This model is
validated against experimental data across various adsorbents, demonstrating excellent alignment
with real-world adsorption rates, particularly for KOH-treated activated carbons. The high
accuracy of this model underscores its robustness and reliability in predicting adsorption dynamics.
To further enhance predictive accuracy and computational efficiency, advanced machine learning
models including GBR, DNN, CNN, and DWNN are applied. Trained on extensive datasets of CO₂
adsorption characteristics, these models outperform traditional approaches and identify the critical
features influencing adsorption, such as surface area and carbon-to-pressure ratios, which are
essential for optimizing gas adsorption systems. Our findings illustrate the potential of combining
theoretical modeling with machine learning to improve the design, operation, and optimization of
gas adsorption and separation processes, such as CO2 capture and natural gas processing. This
work contributes significantly to the advancement of scalable, efficient gas separation technologies,
paving the way for sustainable solutions in climate change mitigation and resource management.Includes bibliographical references (pages 108-121
The effect of plasticity on the mechanical behavior and failuire of aluminum and steel used in ship structures
No full textThe aim of this work is to analyze the residual strains and stresses developed in materials during bending processes. Residual stresses can lead to several undesirable effects such as inducing plastic deformation affecting ship structures' load bearing capacity. Accordingly, both analytical methods from previous works and numerical finite element analysis models using Abaqus were used to investigate the post-yield behavior. In addition, extensive experimental testing was conducted on two materials, namely, steel 44w and aluminum 6061 in accordance with ASTM standards. Three-point bending was utilized to mimic cold forming process with a maximum displacement reaching double the thickness. Through thickness residual strains were captured by noncontact 3D-digital image correlation (DIC) stereo system. Unlike traditional methods, 3D-DIC provides a full-field continuous deformation history throughout the loading and unloading stages. Finally, the effect of plasticity on the mechanical behavior of both materials has been investigated by analyzing specimens extracted from the innermost and outermost layer of bent plates