VTechWorks
VTechWorks provides global access to Virginia Tech scholarship, including journal articles, books, theses, dissertations, conference papers, slide presentations, technical reports, working papers, administrative documents, videos, images, and more by faculty, students, and staff. Faculty can deposit items to VTechWorks from Elements, including journal articles covered by the University open access policy. Email vtechworks@vt.edu for help.
Communities in VTechWorks
Select a community to browse its collections.
Recent Submissions
UAV-Based Deep Learning for Multi-Crop Plant Counting, Geolocation, and Field-Scale Decision Support
Seetharaman, Ragul (Virginia Tech, 2026-07-27)
Accurate plant stand assessment is essential for early-season crop management because plant population, emergence uniformity, and spatial distribution influence crop growth, yield potential, and management decisions such as replanting. Traditional plant counting is commonly performed through manual field assessment, but manual counting is time-consuming, labor-intensive, and difficult to repeat consistently across large fields, multiple cropping systems, and different imaging conditions. Recent advances in unmanned aerial vehicle imagery and deep learning have created new opportunities for automated plant detection and counting. However, many existing workflows remain crop-specific, model-specific, or dependent on programming-based processing steps that limit practical field use.
This thesis presents a manuscript-based study on UAV-based deep learning for automated plant counting, geolocation, and field-scale visualization across multiple crops. The first manuscript evaluates a standardized object detection pipeline for plant counting across faba bean, strawberry, cotton, and tobacco. The pipeline applies image preprocessing, tiled inference, and consistent evaluation procedures to compare YOLO-style and DETR-style object detectors under varying crop morphology, plant density, background texture, and illumination conditions. The study benchmarks YOLOv12, YOLO26, RT-DETR, and RF-DETR using detection-level metrics and count-level agreement metrics. Results show that YOLOv12 achieved the strongest detection performance across the crop datasets, while count reliability varied by crop due to factors such as canopy overlap, shadows, dense rows, small targets, and background texture.
The second manuscript extends the detection workflow into a graphical user interface-based plant-counting system. The developed PlantCountGUI integrates trained object detection models with orthomosaic tiling, tile-level inference, coordinate transformation, duplicate detection removal, count summarization, visualization, and GIS-compatible export. Users can upload georeferenced UAV orthomosaics, select processing parameters, run detection, inspect plant distribution maps and density heatmaps, and export CSV and shapefile outputs. By converting image-based detections into projected field coordinates and presenting results through an accessible interface, the system reduces the gap between model development and practical deployment for agricultural researchers, extension personnel, and growers.
Together, these manuscripts contribute a standardized multi-crop evaluation framework and an applied software tool for UAV-based plant stand assessment. The thesis demonstrates that deep learning-based plant counting can support precision agriculture when detection accuracy, count reliability, geospatial processing, and usability are considered together.
2025 Insect Identification Laboratory Annual Report
Day, Eric R.; Dellinger, Theresa A.; Wooden, Patricia L. S. (2026-11-17)
This report summarizes the activity of the Insect Identification Laboratory at Virginia Tech for 2025.
A social-ecological approach for understanding the sustainability of a small-scale fishery
Hultin, Emma Ann (Virginia Tech, 2026-11-17)
Small scale and inland fisheries have been underrepresented in scientific literature, despite their contributions to global biodiversity and critical role in supporting human livelihoods and wellbeing. The assumptions that underlie conventional fisheries modeling techniques break down in small-scale fisheries, and there is no generalizable framework for integrating social and ecological data in a mechanistic fashion. In this study, I aimed to disentangle the effects of environmental variability on growth and quantify the relative effects of legal and illegal fishing mortality on a floodplain fish population. Using a co-managed small-scale fishery as a case study, I built an ecological model to test drivers of population growth for Arapaima spp. in the Brazilian Amazon. I then extended this framework to partition out legal and illegal fishing mortality using data on the frequency of rule-breaking. I further extended the model by replacing direct rule breaking data with social governance conditions in an attempt to explain variation in illegal fishing across the study region. This is the first study to explicitly and jointly model environmental variables alongside socially-regulated catch dynamics using real fisheries data. I found that maximum annual flood height was robustly and positively associated with population growth of arapaima, but was unable to detect effects of static landcover and hydrologic variables. Poaching and internal non-compliance were both identified as contributing mechanisms for illegal fishing. I was able to recapture the variation in illegal fishing across the study region using governance constructs that characterize the socio-economic context of the fishing communities. My findings have implications for setting quotas within the management system as the current approach may be underestimating the true sustainable yield if unreported mortality is not considered. This study also highlights leverage points where managers could focus intervention efforts to better support the sustainability of community-based management.
The Transportation Impact of Cyberattacks, Detection, and Mitigation in Connected and Autonomous Vehicles
McManus, Ian Patrick (Virginia Tech, 2026-11-17)
Cyberattacks in transportation have become a fast-growing field of research with the rise of Connected Vehicles (CVs) and Autonomous Vehicles (AVs) and the future promise of Connected and Autonomous Vehicles (CAVs). There are multiple categories of cyberattack detection schemes presented in research – each with distinct tradeoffs. These are often based on statistics, Artificial Intelligence (AI), and Machine Learning (ML). Active and real-time cyberattack mitigation is a comparatively less studied subject. Policy guidance and best practices to guide cybersecurity in the transportation field are also sparse. Undetected and unmitigated cyberattacks in transportation represent a real concern for safe and effective operations. This research developed and evaluated an ML-based detection scheme paired with an active mitigation scheme for false data attacks against CAVs at a fully autonomous intersection and proposed policy recommendations and lessons learned based on the findings. This was done through a Vehicles in Network Simulation (Veins) simulation which incorporated a Temporal Convolutional Network (TCN) detection scheme, detection threshold, and a real-time mitigation strategy. Two magnitudes of cyberattacks were simulated – a stealth attack and a disruptive attack. The findings were analyzed to compare baseline operations, unmitigated cyberattack operations, and mitigation operations. This research found that the implemented mitigation strategy was effective at reducing the impact of a cyberattack but did not eliminate delay entirely. The residual delay was driven largely by undetected cyberattacks. The detection mechanism was the limiting factor in the effectiveness of the mitigation strategy. Low-magnitude attacks succeeded at a higher rate than high magnitude attacks, but caused less delay compared to high magnitude attacks. Overall, the research provides three main contributions: a TCN-based ML detection model for a fully CAV intersection, a real-time mitigation strategy, and policy recommendations based on the results to bridge transportation operations and cybersecurity.
Improving Efficiency of Early-Generation Multi-Environment Testing in Soft Red Winter Wheat
Frias Ramirez, Sunilda Patricia (Virginia Tech, 2026-11-16)
Early-generation yield testing requires plant breeders to evaluate large numbers of new lines across diverse environments with limited seed, land, time, and financial resources. Genomic prediction uses phenotypic and genomic information from related individuals to predict the genetic merit of breeding lines. This dissertation investigated genomic prediction for early-generation multi-environment testing in soft red winter wheat (Triticum aestivum L.), focusing on sparse testing, the use of secondary traits to improve prediction of agronomic traits, and evaluation of the best available estimate of genetic merit against which genomic predictions can be compared.
The first study evaluated genomic prediction for sparse testing, in which breeding lines are evaluated in subsets of environments. Across four years of testing, predictive ability for unobserved line–environment combinations ranged from 0.54 to 0.90 for grain yield and from 0.49 to 0.95 for test weight when evaluated against full-data genomic predictions. Sparse testing also consistently outperformed indirect selection across the selection–response environment comparisons evaluated. The advantage was greatest when genetic correlations between environments were low; for example, when the grain-yield genetic correlation between Blacksburg and Warsaw was 0.28, reciprocal indirect-selection accuracies were only 0.20 and 0.24, compared with sparse-testing predictive abilities of 0.75 and 0.78. These results demonstrate that genomic prediction can support broader environmental testing without requiring every line to be phenotyped in every environment, particularly when performance cannot be reliably inferred from another location.
The second study evaluated UAV-derived vegetation indices and traditional field traits as secondary traits for genomic prediction of grain yield and test weight. Vegetation indices were often strongly associated with grain yield; at Warsaw in 2022, mean within-trial correlations reached 0.73 for NDVI, 0.80 for NDRE, and 0.79 for CIRE. However, their apparent predictive value depended strongly on the treatment of residual covariance. Under a model that assumed no residual covariance between traits, CIRE, NDRE, and NDVI increased mean grain-yield predictive ability by approximately 0.12, 0.12, and 0.13, respectively. When residual covariance between traits measured on the same experimental unit was modeled, these gains declined to approximately 0.01, 0.02, and 0.00. In contrast, powdery mildew retained predictive value after accounting for residual covariance, increasing mean grain-yield predictive ability from 0.43 to 0.49. These results indicate that strong phenotypic associations do not necessarily represent additional genetic information and emphasize the value of secondary traits that capture biologically relevant genetic variation.
The final study examined how genomic prediction should be evaluated when true genetic merit is unknown and must instead be estimated from available data. The same historic-only genomic predictions produced substantially different predictive abilities depending on the validation target. For grain yield, correlations ranged from 0.127 to 0.299 with year-specific BLUEs, 0.289 to 0.415 with pedigree-based ABLUP estimates, and 0.646 to 0.812 with a genomic model incorporating all phenotypic information available through the target year. Differences were particularly pronounced within pedigree families. Mean within-family predictive ability for grain yield was 0.130 with BLUEs and 0.145 with ABLUP, compared with 0.745 with the full genomic model. These results demonstrate that predictive ability is not solely a property of the prediction being evaluated; it also depends strongly on the information content and uncertainty of the estimate used to represent genetic merit.
Collectively, these studies demonstrate that improving early-generation testing efficiency is not simply a matter of collecting more data. Sparse testing can increase environmental representation by using genomic connectivity to predict incompletely tested material, while the value of additional phenotypes depends on whether they contribute genetic information beyond shared environmental variation. Similarly, evaluation of genomic prediction depends on the quality and interpretation of the validation target. Together, these findings support designing early-generation testing systems around the relevance, connectivity, and genetic information content of available data, rather than data volume alone.


