Announcements

Title: Integrating Satellite and UAV Remote Sensing for Percent Vegetation Cover Estimation in Military Training Grasslands 
Major Professor: Guangxing Wang 
Committee Members: Ruopu Li, Jonathan Remo, David Gibson, John Groninger 
Date: October 30, 2026 
Virtual Link:  Join Xiaoyu Xu's Ph.D. Defense via Teams 
Time: 2:00 p.m.
Abstract: Military training activities can produce spatially heterogeneous vegetation disturbance, creating a need for reliable and spatially explicit monitoring of vegetation condition across extensive training landscapes. Percent vegetation cover (PVC) is a useful indicator of vegetation removal and recovery, but the predictive value of different remote-sensing platforms and the benefits of integrating their information remain uncertain, particularly when sensors differ in spatial resolution and physical sensitivity. This dissertation evaluated the information value, complementarity, and appropriate analytical roles of Sentinel-2 imagery, unmanned aerial vehicle (UAV) RGB and multispectral (MS) imagery, Structure-from-Motion (SfM) photogrammetry, and UAV LiDAR for PVC estimation in military training grasslands at Fort Riley, Kansas.  

Field reference data comprised 72 independent 10 × 10 m PVC plots sampled during July and August 2020, with 36 different plots in each period. The dissertation addressed three complementary objectives. First, Sentinel-2 spectral and derived predictors were evaluated for period-specific PVC estimation, landscape-scale mapping, and modeled July-to-August PVC change. Second, UAV predictors were organized into predefined spectral, textural, spatial-configuration, and structural information domains to quantify their individual and combined predictive value and to test whether SfM- and LiDAR-derived structural information improved PVC prediction beyond optical information. Third, Sentinel-2 and UAV optical models were compared under a harmonized analytical framework using the same field observations, common 10-m analytical support, and leakage-controlled nested leave-one-out cross-validation. Cross-platform complementarity was further evaluated using early feature-level fusion, prediction-level stacking, paired bootstrap comparisons, residual analysis, and SHapley Additive exPlanations (SHAP). 

The original Sentinel-2 modeling framework demonstrated strong capability for period-specific PVC estimation. The best pooled Random Forest models achieved a cross-validated R² of 0.915 with a root mean square error (RMSE) of 4.35 percentage points in July and an R² of 0.900 with an RMSE of 4.81 percentage points in August. These models supported landscape-scale PVC mapping and the generation of modeled July-to-August PVC difference surfaces. Because the July and August field observations were collected from different plots, these surfaces were interpreted as modeled spatial differences rather than direct paired measurements of vegetation change.

The UAV analyses showed that information type was more important than simply increasing the number of available sensors or predictors. Multispectral texture was the strongest individual UAV information domain in both periods, achieving an RMSE of 7.30 percentage points and a cross-validated R² of 0.629 in July and an RMSE of 8.23 percentage points and an R² of 0.608 in August. At the sensor level, multispectral models also performed consistently well. The primary RGB + MS optical architecture achieved an RMSE of 8.28 percentage points and an R² of 0.523 in July and an RMSE of 8.09 percentage points and an R² of 0.621 in August. RGB information provided additional predictive information under some conditions, but optical fusion did not consistently outperform MS alone. SfM-derived structural metrics provided little reproducible improvement beyond the optical predictors. LiDAR showed an even clearer mismatch with the PVC response: on the common LiDAR-supported subset, the strongest July LiDAR model produced an RMSE of 11.75 percentage points and an R² of 0.084, while August LiDAR models had RMSE values greater than 14.7 percentage points and negative cross-validated R² values. These results indicate that optical spectral and spatial heterogeneity were more directly aligned with horizontally projected PVC than were vertical structural measurements.

Title: Evaluating the Impact of Self-Reflection on Agentic Large Language Model
Major Professor: Henry Hexmoor
Committee Members: Xiaopeng Jiang, Xiaolan Huang
Date: October 28, 2026
Location: Engineering A 309C Conference Room
Time: 11:00 a.m.

Title: Secure Inference & Processing using Differential Privacy and Homomorphic Encryption
Major Professor: Henry Hexmoor
Committee Members: Bidyut Gupta, Tom Imboden
Date: October 26, 2026
Time: 3:00 p.m.
Abstract: Protecting sensitive information during machine learning inference is an important challenge in privacy-preserving computing. While homomorphic encryption enables computations on encrypted data, machine learning prediction outputs may still reveal sensitive information. This research presents a privacy-aware machine learning inference framework that integrates Cheon–Kim–Kim–Song (CKKS) approximate homomorphic encryption with Gaussian Differential Privacy (DP) to address input confidentiality and output privacy.

The proposed framework focuses on three contributions: ciphertext sensitivity mapping, dynamic differential privacy composition, and unified noise calibration. First, CKKS-aware sensitivity analysis incorporates approximation-error considerations into the sensitivity calculation. Second, Rényi Differential Privacy (RDP) is used to calculate and track cumulative privacy loss across repeated inference queries. Third, a unified noise-calibration function determines the Gaussian noise required for a specified privacy budget while accounting for a supplied CKKS approximation-error bound. The framework encrypts input data, performs inference on ciphertext, and applies calibrated Gaussian noise to prediction scores after private client-side decryption.

The framework was implemented using Python, PyTorch, and TenSEAL with the Microsoft SEAL encryption library and evaluated using the MNIST handwritten digit dataset and a logistic regression classifier. Experimental results demonstrated 100% prediction agreement between plaintext and CKKS inference across 100 test images, with small numerical approximation errors. Under a fixed privacy budget of ε = 0.5, the required Gaussian noise standard deviation increased from approximately 30 to 213 as the number of queries increased from 1 to 50. The privacy–utility evaluation showed that prediction accuracy decreased from a 91% baseline to approximately 14% after Gaussian noise was applied for a single query.

These findings demonstrate the feasibility of integrating encrypted machine learning inference with Gaussian privacy mechanisms and repeated-query privacy accounting. The research provides an implementation and experimental evaluation of CKKS-aware noise calibration while highlighting the tradeoff between privacy protection, prediction accuracy, and repeated-query usage.

Keywords: Homomorphic Encryption, CKKS, Differential Privacy, Rényi Differential Privacy, Privacy-Preserving Machine Learning, Unified Noise Calibration, Secure Inference.

Title: Context-Aware Argumentation AI Agents using LLM 
Major Professor: Xiaoqing Liu 
Committee Members: Henry Hexmoor, Zhong Chen 
Date: October 26, 2026 
Location: Engineering E 101 
Time: 2:00 p.m. 
Abstract: Large Language Models (LLMs) are increasingly being used to generate responses in argumentative settings, but the quality of those responses depends in part on the context provided to the model. In structured discussions, relevant information may appear in the immediate context of a target argument, earlier along the same discussion path, or in other branches of the argument tree. This thesis examines how different strategies for selecting and representing discussion context affect the quality of LLM-generated argumentative responses in the Xargue platform. Four context-selection strategies were implemented using the same GPT-4o-mini response-generation model: Minimal-Context, Path-Based, Full-Tree, and Relevance-Filtered. The strategies differ in the amount and form of discussion context supplied to the model. The Relevance-Filtered strategy further summarizes discussion paths, generates embeddings, compares them semantically with the target argument, and retains only summaries that satisfy a relevance threshold. The evaluation was conducted using 50 human-authored target arguments from a Fall 2017 discussion, producing 200 generated responses across the four agents. The responses were assessed by one subject-matter expert using a 0–5 rubric based on relevance, degree of comprehension relative to the context, argumentative consistency, and stance rating. Input-token usage was recorded separately to compare the amount of context required by each strategy. The results show that providing more context does not necessarily lead to better argumentative responses. The Full-Tree Agent achieved the highest mean comprehension score, indicating that access to the complete discussion tree improved awareness of the broader context. However, its lower relevance, argumentative-consistency, and stance scores suggest that GPT4o-mini did not always distinguish the most useful information from less relevant branches, allowing some of the additional context to function as contextual noise. The Path-Based Agent achieved the highest mean stance rating, while the Relevance-Filtered Agent obtained the highest mean scores for relevance and argumentative consistency, as well as the highest overall mean evaluation score. It also used fewer input tokens than the Full-Tree Agent. Muchall, the findings indicate that the effectiveness of discussion context depends not only on how much information is available, but also on how effectively that information is selected, organized, and presented to the model. The results suggest that context availability and context usability are not the same, and that careful context selection can support more focused, consistent, and efficient LLM-generated argumentative responses in structured online discussions. 

Title: Deep Learning Approaches for Early Detection and Identification of Escherichia Coli and Salmonella Typhimurium in Time-stamped Microscopy Images
Major Professor: Abdur Rahman Bin Shahid                                                                                          Committee Members: Anas Alsobeh,Henry Hexmoor
Date: September 30, 2026
Virtual Link: Join Bibek Koirala's Thesis Defence via Teams
Time: 12:00 p.m.
Abstract: Conventional culture-based detection of foodborne pathogens needs from two days to over a week - far longer than raw produce spends in the modern cold chain. This thesis investigates how early, how accurately, and at what cost Escherichia coli and Salmonella enterica serovar Typhimurium can be detected and identified in microscopy images acquired during the first four hours of incubation, under plain and onion food-matrix backgrounds. Three approaches form a progression, each motivated by a limitation of the one before it. The first couples phase-contrast microscopy with a YOLOv8x detector trained and evaluated on expert-verified images acquired between 2 and 4 h, including mixed-culture fields, drawn from a collection of 4,981 images. It achieves precision 0.891, recall 0.867 and mAP@0.50 of 0.898, and documents Salmonella micro-colony growth from approximately 25 µm² at 2 h to about 1,050 µm² at 4 h. Its limitation is one of supervision: within mixed-culture fields the species label of each micro-colony rests on expert visual annotation, and because the two species are morphologically similar when colonies are only tens of pixels across, model error cannot be separated from label error. The second removes that ambiguity by training an end-to-end Mask R-CNN exclusively on separately incubated cultures, where every label follows from the incubation protocol, reaching mean intersection-over-union of 0.91–0.98 and mAP@0.50 of 0.95 at two hours. The third keeps that clean supervision, restricted to 2–4 h, and compares four segmentation architectures under a shared loss toolbox and evaluation protocol to ask whether a convolutional model can match mask-pretrained and self-supervised transformer-based ones. Two hours emerges as the earliest time point at which detection is reliable under both backgrounds: below it, onion debris is visually indistinguishable from emerging colonies and cannot be annotated reliably. The convolutional model, comparable in size and compute to the self-supervised models, is strongest overall on pixel-wise and region metrics while running roughly twenty-two times faster than the Segment Anything Model variant. The residual error concentrates at two-hour Salmonella and persists across every backbone tested, indicating a limit of the visual signal at that incubation time rather than of any one architecture.

Title: Evaluating Reasoning in Modern Large Language Models
Major Professor: Henry Hexmoor
Committee Members: Sinha Koushik, Holly Lewis
Date: October 16, 2026
Location: Engineering A, Room 309C, School of Computing Conference Room
Time: 1:30 p.m.
Abstract: Large Language Models (LLMs) achieve high accuracy on factual question answering, yet it remains unclear whether correct answers reflect correct reasoning. Standard evaluation measures only the final answer, which cannot distinguish a conclusion reached through valid inference from one reached by pattern matching or coincidence. This limitation is particularly consequential in knowledge graph tasks, where the correctness of a derivation — not merely the answer —determines whether a system will generalize. This thesis investigates the gap between answer accuracy and reasoning quality in LLM knowledge graph reasoning. Knowledge graph question answering is selected as the evaluation substrate because datasets such as LC-QuAD 2.0 provide gold SPARQL queries that constitute machine-readable ground-truth derivations, specifying not only the correct answer but the correct sequence of properties to traverse. This property enables direct comparison between a model's stated reasoning path and a reference derivation, rather than the indirect inference used in prior faithfulness research.

Title: Vauxy Universal Cosmetic Exhibition Show
Major Professor: Aaron Scott
Committee Members: Robert Lopez, Alison A. Smith, Jason Wonnell, Corey Tester
Date: October 16, 2026
Location: Pulliam Hall Design Wing, Room 59
Time: 3:00 p.m.
Abstract: The exhibition presents the culmination of my graduate research and creative practice in accessible and universal design. My thesis explores how tactile systems, ergonomic product forms, and interactive experiences can enhance accessibility. The exhibition will feature my Vauxy Universal Cosmetic Product System, along with research, digital applications, and supporting design outcomes.

Title: Quantifying Suitable Habitat for Survival and Recruitment of Age-0 Invasive Carp Along the Upper Mississippi and Illinois Rivers
Major Professor: Jonathan Remo                                                                                                          Committee Members: Justin Schoof
Date: October 15, 2026
Location: Geography Conference Room, Faner 4523
Time: 3:00 p.m.
Abstract: Native fish communities in the Upper Mississippi River (UMR), upstream of the Missouri River confluence, are being adversely affected by invasive Hypophthalmichthys nobilis (bighead carp) and H. molitrix (silver carp). These species compete with native planktivores by reducing zooplankton abundance, altering community structure and habitat availability. Survival and recruitment during the age 0 life stage depend on the availability and quality of suitable physical habitat. Preferred conditions for this life stage, commonly referred to as shallow water habitat (SWH), include depths less than 1.5 m and velocities below 0.6 m/s. These conditions are most critical from May through October, the primary spawning and first year growing season, and are typically found in floodplain and backwater areas of the Upper Mississippi and Illinois rivers. This study quantified the temporal and spatial distribution of SWH in Pool 26 of the UMR and the La Grange segment of the Illinois River using hydrodynamic model outputs paired with fish community data from the Upper Mississippi River Long Term Resource Monitoring Program (LTRM), which provides extensive age 0 invasive carp records from 2000 to 2024. The two river segments exhibited distinct SWH patterns driven by differences in geomorphology and floodplain connectivity. Pool 26, a larger and less levee-confined river segment, contained substantially more SWH, including 18.4 km² of high persistence habitat that remained available for 55% of the spawning season. The highly leveed La Grange Reach contained only 6.2 km² of high persistence SWH and lost habitat rapidly during high flows. In Pool 26, SWH was most common along main channel margins and side channels, while persistent SWH in La Grange occurred primarily in connected floodplain lakes and tributary junctions where low velocities created favorable conditions for age 0 Asian carp. Relationships between SWH and catch per unit effort (CPUE) were generally weak across both segments. The La Grange study segment showed statistically significant positive correlations between age 0 Asian carp CPUE and both minimum and average seasonal (p-value).

Title: Navigating Covid-19 Uncertainties: Postsecondary Language Instructors' Experiences
Major Professor: Heidi R. Bacon
Date: August 31, 2026
Location: Wham 112
Time: 10:00 a.m.
Abstract: The COVID-19 pandemic left a lasting impact on second and world language education. Although extensive literature has focused on student experiences, instructors' voices are often left out of the conversation. Language teaching relies heavily on relational, interactive communication, making sudden disruptions to physical classrooms especially complex. Postsecondary language instructors had mixed experiences with these changes, navigating heavy stress and emotional exhaustion while finding opportunities to build new practices and develop growth mindsets. To capture and understand how language instructors lived through these transitions before, during, and after the pandemic, I conducted this narrative inquiry through the lens of pragmatic constructivism. I interviewed seven postsecondary language instructors at a Midwestern U.S. public university, including faculty and graduate teaching assistants, and incorporated aspects of self-study through reflective journaling. Our collective experiences are revealed through these individual and shared accounts and practices. To analyze the data, I used Braun and Clarke's (2022) six-phase thematic analysis and generated three key themes: surviving the shifts and refining instruction, steering disconnection and reclaiming the classroom community, and sustaining the self and re-evaluating professional values. Despite persistent challenges associated with changes in instructional locations and modalities, the instructors pursued student-centered pedagogies and prioritized students' changing needs. Over time, these experiences reshaped their teaching efficacy and overall outlook on the profession. I share theoretical and practical implications for language instruction and higher education, pointing to the need for healthy professional boundaries and institutional support for language educators. Methodologically, I recommend extending this inquiry through longitudinal designs to track instructors' ongoing experiences or expanding the self-study by fully embracing fluid, autoethnographic approaches. I also recommend investigating how emerging technologies and automated tools continue to shape language instructors' professional identities and efficacy.

Title: Assessing Water and Sediment Control Basins (WASCoBs) on Trapping Sediment and Nutrients in Southern Illinois: Influence on Water Quality
Major Professor: Jon Schoonover
Committee Members: Karl Williard, Samuel Indorante
Date: August 24, 2026
Location: AG 209
Virtual Link: Join Durga Kandel's Thesis Defence via Teams
Time: 2:00 p.m.
Abstract: Agricultural non-point source pollution is a major cause of water quality degradation in the United States, primarily due to sediment and nutrient runoff from intensified agricultural practices. This has increased the need for effective Best Management Practices (BMPs) to mitigate soil erosion and reduce the export of sediment and nutrients to receiving waters. In-field structural BMPs, such as Water and Sediment Control Basins (WASCoBs), are considered effective and have been widely implemented by landowners to reduce soil and nutrient losses from agricultural fields. While some studies have evaluated their sediment trapping efficiency, limited research exists on their nutrient retention performance and on quantifying annual sediment deposition within basins, particularly under real-world field conditions. Over a two-year period, this study quantified sediment accumulation in thirty-two WASCoBs in Randolph, Jackson, and Williamson counties of southern IL, where six basins were intensively monitored for water quality using inflow and outflow during runoff-generating rainfall events. Runoff samples were collected from these watersheds and analyzed for total suspended solids (TSS), total phosphorus (TP), dissolved reactive phosphorus (DRP), ammonium-N, and nitrate-N. Annual sediment accumulation was measured from all 32 WASCoBs using feldspar pads and compared with UAV-derived digital surface models. An Unmanned Aerial Vehicle (UAV) was used to estimate sedimentation rates and proved to be effective for multi-year sediment estimates rather than annual estimates, as it is difficult to detect small accumulation annually, with an observed error of ±3 cm. These WASCoBs were highly effective in trapping sediment and total phosphorus, with average trapping efficiencies of 89% for TSS and 64% for TP. In contrast, dissolved nutrients exhibited substantially lower retention efficiencies, with 21% for DRP, 47% for ammonium, and 16% for nitrate, reflecting their mobility in solutions. Rainfall and slope were found to be significant drivers of sediment and nutrient transport. Along with that, field management practices played a critical role, with tilled and fallow conditions producing significantly higher sediment and nutrient loads compared to planted conditions. Feldspar pads placed within the basins showed varied accumulation of sediment across basins, ranging from 24.7 to 44,602 kg/ha annually, and trapped phosphorus ranged from 0.005 to 13.5 kg/ha, with a strong positive correlation between sediment deposition and trapped phosphorus (R² = 0.97, pp < 0.001). UAV photogrammetry was effective for estimating multiyear sediment accumulation (RMSE = 3.52cm), although surface residue and vegetation contributed to overestimation of sediment deposition in some basins. These results suggest that WASCoBs are highly effective in trapping sediment and particulate nutrients from agricultural fields to receiving waterways. However, to enhance WASCoB performance for overall nutrient retention, including water-soluble nutrients, WASCoBs should be combined with improved land management practices, such as reduced tillage and cover cropping. Additionally, consistent monitoring and maintenance can further improve their long-term performance. Therefore, the continued and integrated implementation of WASCoBs has strong potential to support sustainable agricultural production while achieving water quality and soil conservation goals.