Master of Science (M.S.), University of Dayton, 2017, Renewable and Clean Energy

The purpose of this study is to generate a regression model tying the vehicular emissions in Cochabamba, Bolivia to input factors including the current state of the public fleet, city population, weather, and GDP. The finished model and the process to generate it can act as a tool to project future emissions in the city, accounting for the aforementioned input factors. It can also be used to estimate the drop in city pollution levels in a scenario where the public transportation fleet is partially replaced by non-emitting, electric vehicles. The main pollutant focused on in this study is particulate matter (PM10), but data also exists for ozone (O3), nitrogen dioxide (NO2), and sulfur dioxide (SO2). The model generation process explained in the study could be applied to these pollutants as well. The regression model is generated using the open source software, R. Its final form utilizes a random forest regression model, but neural net, gradient boosting, and support vector machine models were also explored.

Committee: Robert Brecha Ph.D. (Committee Chair); Andrew Chiasson Ph.D. (Committee Member); Malcolm Daniels Ph.D. (Committee Member) Subjects: Engineering; Environmental Engineering; Mechanical Engineering

MS, Kent State University, 2024, College of Arts and Sciences / Department of Geography

Wildfire and drought are key drivers of shrubland expansion in southwestern US landscapes. Stand-replacing fires in dry conifer forests induce shrub-dominated stages, and changing climatic patterns may cause a long-term shift from coniferous forests to deciduous shrublands. This study assessed recent changes in deciduous fractional shrub cover (DFSC) in the eastern Jemez Mountains from 2019-2023 using topographic and Sentinel-2 satellite data in a random forest model. Sentinel-2 provides multispectral bands at 10 and 20 meters, including three 20 meter red edge bands, which are highly sensitive to variation in vegetation. There is no consensus in the literature on whether upscaling imagery to 20 meters or downscaling to 10 meters is more advantageous. Therefore, an additional goal of this study was to evaluate the impact of spatial scale on DFSC model performance. Two random forest models were built, a 10 and 20 meter model. The 20 meter model outperformed the 10 meter model, achieving an R-squared value of 0.82 and an RMSE of 7.85, compared to the 10 meter model (0.76 and 9.99, respectively). The 20 meter model, built from 2020 satellite imagery, was projected to the other years of the study, by replacing the spectral variables with satellite imagery from the respective year, resulting in yearly predictions of DFSC from 2019-2023. DFSC decreased from 2019-2022, coinciding with severe drought and a 2022 fire, followed by a significant increase in 2023, particularly within the 2022 fire footprint. Overall trends showed a general increase in DFSC despite high interannual variability, with elevation being a key topographic variable influencing these trends. This study revealed yearly vegetation dynamics in a semi-arid system and provided a close look at post-fire regeneration patterns in deciduous resprouting shrubs. Understanding this complex system is crucial for informing management strategies as the landscape continues to shift from conifer forest to shrubland du (open full item for complete abstract)

Committee: Scott Sheridan (Advisor); Christie Bahlai (Committee Member); Timothy Assal (Advisor) Subjects: Ecology; Geography

PhD, University of Cincinnati, 2023, Engineering and Applied Science: Mechanical Engineering

The aim of this dissertation is to present the development of new approaches to air traffic management for vehicles operating within the national airspace. This dissertation examines air traffic management techniques for both manned commercial aircraft along with small unmanned aerial systems (sUAS). This dissertation starts with the development of a supervised learning algorithm to generate estimated time of arrival for commercial aircraft and covers a unique conflict prediction algorithm for sUAS which is based on near future path predictions. Both contributions assist in air traffic management by minimizing required human intervention. The dissertation culminates with the development of a novel predictive artificial potential field (PAPF) navigation system for sUAS. The PAPF relies upon Interacting Multiple Model (IMM) vehicle predictions and is designed to eliminate the need for human intervention as sUAS operate within the airspace by preemptively rerouting vehicles in real time based on conflicts that will likely happen in the near future. The PAPF system is run completely on-board the sUAS and does not require additional ground based resources. The first element of the dissertation covers the estimated time of arrival (ETA) predictor for commercial aircraft. The ETA predictor focused on aircraft landing at the Dallas Fort Worth airport and used historical flight and meteorological data to train a random forest regressor. The trained algorithm could be implemented to make use of real time data and estimate the remaining flight time once aircraft have entered within a 200 mile radius circle of the airport. The data used is readily available from airport facilities and the aircraft themselves via Automatic Dependent Surveillance-Broadcast (ADS-B). This work was supported by NASA Office of STEM Engagement and used historical flight data for aircraft landing at Dallas Fort Worth airport. The second element discussed in the dissertation consists of a (open full item for complete abstract)

Committee: Manish Kumar Ph.D. (Committee Chair); Michael Alexander-Ramos Ph.D. (Committee Member); David Thompson Ph.D. (Committee Member); Tejas Puranik Ph.D. (Committee Member); Rajnikant Sharma Ph.D. (Committee Member); Zachariah Fuchs Ph.D. (Committee Member) Subjects: Mechanical Engineering

Doctor of Philosophy (Ph.D.), Bowling Green State University, 2023, Media and Communication

This dissertation investigates the effectiveness of counterspeech strategies employed on Twitter in response to anti-Asian hate during the COVID-19 pandemic. This research delves into the communicative strategies, emotional tones, and geospatial distribution of counterspeech, specifically focusing on its effectiveness in the United States. A supervised machine learning was employed to classify counterspeech tweets and counterspeech strategies based on empirical typology. By analyzing 106,388 tweets associated with the hashtag #StopAsianHate collected from November 2021 to May 2022, this research provides insights into the varied effectiveness of counterspeech strategies. The analysis revealed that though counterspeakers were using more negative tones in counterspeech tweets, the tweets with visual media and positive emotional tone received more engagement on Twitter through retweets and favorites compared to those with a negative or neutral tone. This study also breaks new ground by recognizing that higher level of racial diversity does not facilitate higher level of counterspeech against hate speech and hate crime. Additionally, this study highlights the varying degrees of participation in counterspeech across different ethnic groups within Asian American community and underscores the importance of tailored strategies in addressing hate speech. Recognizing this distinction proved essential in crafting evidence-based guidance for community and individual interventions while fostering support from allies of diverse racial backgrounds.

Committee: Louisa Ha Ph.D. (Committee Chair); Lisa Hanasono Ph.D. (Committee Member); Yanqin Lu Ph.D. (Committee Member); William Sawaya Ph.D. (Other) Subjects: Artificial Intelligence; Asian American Studies; Communication; Mass Media; South Asian Studies

Master of Science (MS), Bowling Green State University, 2020, Applied Statistics (Math)

Breast cancer is the second leading cause of cancer death in women, and chronic kidney disease (CKD) is a silent illness, affecting people long before they even know it. Early detection is important in order to intervene and prevent the disease from reaching an incurable stage. Although some resources, such as medical imaging, are available, it is still imperative to look for accurate, reliable, and cost-effective models to detect the disease symptoms in the early stages. The goals of this thesis are to find the smallest combination of features that display strong predictive accuracy and to develop cost-effective models that can accurately forecast a disease outcome. To achieve this, stepwise variable selection is implemented to find the most appropriate set of features to build the models. Logistic regression, decision tree, bagged trees, random forest, and neural networks are used to detect breast cancer and CKD cases. Each of the five models are constructed and evaluated on a 70-30% and 50-50% train-test split. Altogether, 20 models are developed for breast cancer and CKD classification in this thesis. A model comparison is performed for each data set based on accuracy and misclassification error. The results show that neural networks performed better in breast cancer diagnostic procedures. For CKD classification, bagged trees and random forest were able to achieve greater accuracy. Between the two train-test splits, the 70-30% split was more advantageous to breast cancer classification. For CKD models, the split had its own advantages depending on the classifier.

Committee: Junfeng Shang Ph.D. (Advisor); John Chen Ph.D. (Committee Member); Wei Ning Ph.D. (Committee Member) Subjects: Statistics

Master of Science, The Ohio State University, 2020, Public Health

The main purpose of research for this thesis is to compare the use of logistic and random forest classification models in machine learning to predict the outcome of adolescent tobacco use. The logistic classification model is one in which the conditional probability of the binary outcome is assumed to be equal to a linear combination of asset of independent variables, transformed by the logistic function. The random forest classification model is one in which the independent variables are used in creating many decision trees which are used to predict the binary outcome of interest. The data source is Buckeye Teen Health Study (BTHS), a survey among adolescent males to examine tobacco use behaviors. The goal of BTHS was to examine factors associated with cigarettes and smokeless tobacco product usage among urban and rural male adolescents in Ohio. Participants who answered questions at baseline and 12-month follow-up were 11- to 16- year old boys (N=1046) with 625 from the urban county and 421 from the nine rural Appalachian counties. The classification models focused on cigarette, e-cigarette, any tobacco and past 30-day tobacco usage at 12 months as the outcomes of interest. The dataset was split into two random groups with a 70:30 ratio for training and validation purposes to assess the classification accuracy of each model. The predictive capabilities of the models were assessed using ROC curves, overall classification error rates, specificity and sensitivity measurements. Overall, the logistic models performed slightly better than the random forest counterparts, but both models had high classification accuracy in determining adolescents who did not display the outcomes of interest. The random forest models and logistic models displayed high specificity measures for all outcomes which shows that these classification models are promising techniques for determining adolescents who will not initiate tobacco use.

Committee: Bo Lu Dr. (Advisor); Amy Ferketich Dr. (Committee Member) Subjects: Biostatistics; Public Health

Doctor of Philosophy (PhD), Ohio University, 2019, Experimental Psychology (Arts and Sciences)

First introduced by Brunswik (1952), judgment analysis, the statistical modeling of human judgment, has greatly contributed to psychology's understanding of human judgment. Judgment analysis has the potential to more broadly impact psychological and behavioral research yet complexities of its individual-level approach to modeling lead to challenges in its application. One way to increase the accessibility of judgment analysis would be to employ variable selection methods that decrease the predictor pool to allow for research designs that are representative, but do not require large numbers of observations from single individuals. The following research tests three modeling approaches: stepwise regression, all subsets method, and random forest method, to compare each methods approach to variable selection. Study One, which applied each of these methods to empirical judgment analysis data, indicated that the random forest method has lower goodness-of-fit than the all-subsets method, but greater generalizability than both the all subsets and stepwise regression methods. Study Two found that in two out of four cases, random forest was better than the all-subsets method and in three cases better than the stepwise regression approach in including relevant predictors in the final model. Study Three found that the random forest method was less susceptible to multicollinearity, more likely than either other method to exclude irrelevant predictors even when they were correlated with relevant ones. Overall the random forest approach shows great promise and its use may facilitate broader applications of judgment analysis.

Committee: Claudia González-Vallejo (Advisor); Bruce Carlson (Committee Member); Jeffrey Vancouver (Committee Member); Lonnie Welch (Committee Member); Yuchun Zhou (Committee Member) Subjects: Psychology; Quantitative Psychology

PhD, University of Cincinnati, 2016, Medicine: Biostatistics (Environmental Health)

Particulate matter (PM) has long been known to have a negative effect on public health. Epidemiological studies associating air pollution and other sources of PM often rely on land use modeling for exposure assessment. This approach relies on the association of characteristics of the surrounding land with PM concentrations. Land use regression (LUR) is the most commonly implemented land use model and has several drawbacks, including model instability due to correlated predictors and an inability to capture non-linear relationships and complex interactions. Here, I utilize the machine learning random forest model within a land use framework to generate a novel land use random forest (LURF) model. Using ambient air sampling data from the Cincinnati Childhood Allergy and Air Pollution (CCAAPS) study, I developed LURF and LUR models for eleven elemental components of particulate matter, including Al, Cu, Fe, K, Mn, Ni, Pb, S, Si, V, Zn. We show that LURF models utilized a higher number and more diverse selection of land use predictors than the LUR models. Furthermore, the LURF models were more accurate and precise predictors of all elemental PM concentrations, except for Fe, Mn, and Ni. To extend the usability of the LURF models, I utilized the recent application of the infinitesimal jackknife (IJ) to the random forest model in order to estimate the prediction variance. The IJ theorems were originally verified under the assumptions of traditional random forest framework, namely using CART trees and bootstrap resampling. Alternatives to the traditional random forest, such as subsampling instead of bootstrap resampling and conditional inference trees instead of CART trees have been shown to increase the accuracy of the random forest algorithm and eliminate its variable selection bias. Here, I conduct simulation experiments to show that the IJ performs well when using these random forest variations. Specifically, using the conditional inference tree instead of the C (open full item for complete abstract)

Committee: Patrick Ryan Ph.D. (Committee Chair); Roman A. Jandarov PH.D. (Committee Member); Marepalli Rao Ph.D. (Committee Member) Subjects: Biostatistics