33760 research outputs found
Sort by
Deriving Vegetation Variables from Satellite Observations using a Data-driven Approach
No full textThese slides were presented on 20 August 2025 by Alan Wang during the virtual session of the 2025 CISESS Summer Internship Presentation workshop held online via Zoom and hosted by the Earth System Science Interdisciplinary Center (ESSIC) located at 5825 University Research Court, Suite 4001, University of Maryland in College Park, Maryland.This presentation was shared at the virtual conference of the 2025 CISESS Summer Internship Program held on 20 August 2025. The slides were presented by Alan Wang, an undergraduate student at the University of Maryland, under the mentorship of Heshun Wang (CISESS/UMD). Building on previous research in processing Earth Observation data, the presentation detailed the performance of the XGBoost, Cubist, and random forest regression models in deriving in-situ measurements of vegetation cover fraction (fCover) from satellite observations. Ground measurements of fCover from 43 National Ecological Observatory Network (NEON) sites provided labels for training, validation, and testing, which were then upscaled to align with the high spatial resolution land surface reflectance data provided by the Visible Infrared Imaging Radiometer Suite (VIIRS) daily surface reflectance (VNP09GA) product. When evaluated against unseen data, the random forest regression model demonstrated the best agreement (R-squared = 0.912, MAE = 0.043), followed by the XGBoost regressor (R-squared = 0.910, MAE = 0.043) and lastly the Cubist model (R-squared = 0.904, MAE = 0.047). Applying the random forest model on the 2023 VIIRS data for the East Coast produced estimates consistent with the expected annual phenological cycle.This study was supported by NOAA grant NA24NESX432C0001 (Cooperative Institute for Satellite Earth System Studies - CISESS) at the University of Maryland/ESSIC
HOW INTERNATIONAL ORGANIZATIONS COOPERATE FOR CONFLICT MANAGEMENT AND ITS OUTCOMES
No full textThis dissertation examines the extent to which international organizations (IOs) cooperate in conflict zones and how such cooperation influences civil conflict outcomes and prospects for peace. Although IOs often share the overarching goal of reducing violence, their levels of cooperation in civil conflicts vary. I argue that conflict complexity and public support are two dimensions that IOs consider when debating the costs and benefits of cooperating for conflict management. Building on these variations, I examine how cooperation among IOs affects conflict outcomes and the durability of peace. Much of the literature on conflict outcomes focuses on the ways increased amounts of resources help reduce violence. However, I suggest that with cooperation between organizations, the resources can be distributed more effectively, reduce redundancies, improve coordination between mediation and peacekeeping, and show the organizations’ resolve to solve the conflict. Additionally, the literature has thoroughly examined how international actors promote peace by addressing commitment and information problems. Through cooperation, IOs can amplify their leverage, reinforce agreements, and raise the costs of renewed violence. I argue that cooperation increases both the likelihood and longevity of peace by multiplying resources, strengthening signals of resolve, and forming institutional partnerships that can be mobilized for future conflict management. To test these claims, I developed and employed a novel cooperation score that captures IO cooperation annually across Africa from 1991 to 2018. This dissertation contributes to the literature by conceptualizing cooperation as a scale of activity rather than a dichotomous phenomenon, offering a more nuanced understanding of IO behavior. Ultimately, it highlights IO cooperation as a critical mechanism for reducing violence and fostering sustainable peace
RACIAL STIGMA AND SUBSTANCE USE OUTCOMES AMONG AFRICAN AMERICAN EMERGING ADULTS: A REPLICATION AND EXTENSION
No full textThe present dissertation employs cognitive neuroscience methods (EEG/ERP analyses) to replicate and extend a novel, cue-based paradigm developed to investigate underlying neural mechanisms involved in the immediate context of a racial stigma event. Racial discrimination is a chronic psychosocial stressor for African Americans associated with a broad range of deleterious physical and mental health outcomes, including increased rates of substance use, especially in emerging adults (Gibbons & Stock, 2017; Williams et al., 2019; Brondolo et al. 2009; Pascoe & Smart Richmann, 2009). Several studies have outlined this relationship between racial discrimination and risky behavior, but there are few lab-controlled, neuroscientific experimental paradigms that have examined this link. Discrimination has been shown to impair self-regulatory strategies and cognitive control processes; thus, we sought to leverage cognitive neuroscience approaches to assess the immediate response to a racial stigma event, and if these responses lead to changes in emotional processing, executive function, substance cue reactivity, and reward processing. The development of the primary racial-stigma, cue-focused task served as the basis for a current NIDA K08 DA053441-01A1; PI: Risco; Mentor: Bernat. The task builds on cue reactivity and substance use literature to assess stigma-related change in brain systems relevant to risk-taking vulnerability. We sought to replicate previous findings with this paradigm, as well as extend the assessment to substance cue reactivity. Succinctly, we sought to measure changes in affective and regulatory responding both during and after exposure to racial stigma cues, aligned with previous work (Risco, Butler et al., 2020, 2023, 2025), and newly assess the impact of these modulations on drug cue reactivity. In the replication, with a larger sample, and extension, we sought to assess neural underpinnings of the link between discrimination and substance use. Results revealed two separable groups with unique sensitivities to racial stigma cues, but after exposure, a similar responding across groups. In the extension to substance cues, results were inconclusive most likely due to characteristics of the sample (i.e. well-functioning, non-using). Overall, results address a dearth in health disparities research and providing a foundation for future research studying the inextricable link between racial discrimination and risk behavior
THERAPISTS’ RESPONSE TO PROSPECTIVE CLIENTS: THE ROLE OF PERCEIVED GENDERED, RELIGIOUS IDENTITY ON THERAPISTS’ ELECTRONIC RESPONSE TO NEW CLIENTS
No full textMuslim American adults experience significant discrimination (approx. 60%) in personal and institutional settings in part due to disproportionately negative media portrayals that racialize Muslim Americans in gendered ways (Joshi, 2006; Karim & Eid, 2012; Mogahed & Ikramullah, 2022; Shaheen, 2003; Zaal, 2012). Despite this, no comprehensive data exists on the extent to which Muslim Americans can access mental health care (Outadi & Bedi, 2024). This is concerning in light of recent findings that biases based on religion, gender, socioeconomic class, and sexual orientation influence the way mental health clinicians respond to individuals seeking counseling or psychotherapy (Kugelmass, 2016, 2019; Moscovitz et al., 2023; Outadi & Bedi, 2024; Shin et al., 2016, 2021). The current study investigated whether prospective clients arelikely to experience differential treatment when seeking mental health services based on their gendered, religious identities. A sample of clinicians (N=980) received help-seeking emails from one of four fictitious clients (Muslim woman, Muslim man, Christian woman, Christian man). Findings provide potential evidence for gendered, religious bias as well as overall gender bias among mental health care providers across the United States. Results demonstrated that across all primary measures (responsiveness, receptiveness, time to respond), one of the two Muslim American confederates received significantly less access to mental healthcare services compared to a Christian American confederate, with the Christian woman emerging as the most likely to be favored and the Muslim man emerging as the least likely to be favored. Furthermore, women confederates, regardless of religion, received more access to services than men. Recommendations for future clinical practice, training, and research are provided
MODELING THE IMPACT OF PARTICULATE MATTER ON RESPIRATORY HOSPITAL ADMISSIONS IN ALASKA DURING THE WILDFIRE SEASON
No full textWildfire-derived smoke poses a global hazard to residents of urban and rural areas alike. Studies have documented the adverse health effects associated with exposure to wildfire-derived smoke, which include respiratory and cardiovascular morbidity and mortality. The State of Alaska is a natural case study to understand smoke exposure assessment in rural, sparsely populated areas, where most pollution in late spring and early fall can be attributed to wildfires. This dissertation addresses critical methodological gaps in wildfire-derived PM2.5 exposure assessment. There are three distinct studies in this dissertation. Following an introductory chapter, chapter two is a stock take of existing surface monitoring on the country and global scales as it can be applied to monitoring wildfire-derived smoke. Ground-based air quality monitors are the operating standard for assessing air quality for regulatory purposes, but there are significant gaps in the coverage of monitors, especially in sparsely populated areas of the world that warrant approaches that use satellite data and other modeling approaches. Chapter 3 presents a computationally light modeling approach that derives PM2.5 at 1 km resolution using satellite-derived fire metrics with cross-validated R2 values over 0.80. Chapter 4 demonstrates how selecting exposure datasets in common epidemiological studies is a fundamentally important decision. Sensitivity analyses to test the robustness of results are standard practice, but changes are made within the same single exposure dataset. This research underscores the importance of interdisciplinary approaches bridging geospatial science, public health, and epidemiology. Future work should incorporate social science perspectives and develop open-source tools to translate air quality models into actionable resources for wildfire preparedness and response. As Alaska faces growing wildfire risk, improved exposure estimation methods are critical for protecting vulnerable populations
SYSTEM LEVEL DESIGN FOR EMERGING SECURITY AND 3D IC TECHNOLOGIES
No full textElectronic design at the system level enables greater flexibility and qualityimpact over design algorithms that work with more detailed design abstractions. In
this dissertation, new system level methods are applied to two areas in integrated
circuit security and design that are typically addressed at the circuit or physical
design stages: logic locking and three-dimensional integrated circuits (3D ICs).
Stripped functionality logic locking (SFLL) is one of the leading logic lock-ing methods in the current literature. While SFLL and its variations have good
mathematical guarantees against resilience to various attacks, at its core SFLL is
a circuit level technique and hence a naive implementation does not depend on
system level details. Additionally, implementations of SFLL can come with high
overhead, in particular the restore units needed to fully implement SFLL. We show
that extending SFLL beyond circuit level boundaries and into the system design
level through resource sharing and reuse significantly reduces SFLL implementation
overhead. In particular, we examine two system design approaches that enable
sharing at the system level to lower overall locking implementation overhead, as
well as a clock gating method that relies on strategic application of SFLL at the
datapath architecture level in order to reduce total system power.
3D ICs present a new dimension in the design and packaging of ICs and promiseincreased density, shorter wirelength, and improved performance. However, the
complex interplay between placement and PPA increases substantially, particularly
in regards to inter-tier TSVs needed for inter-tier connectivity, complicating the
design process. In this dissertation, we describe three co-design approaches to 3D
global placement and datapath architecture synthesis via HLS for timing, dynamic
power, and TSV area optimization. We show that leveraging the greater flexibility
afforded by the system-level perspective of HLS in the loop with feedback from
global placement along with providing global placement with design details from
HLS significantly enhances overall design quality. Our results show that our co-
design approaches significantly reduces total negative slack (TNS), dynamic power
consumption, and total TSV usage over conventional methodS
The Victim-Offender Overlap Contextualized: Unpacking Heterogeneity by Victimization Degree, Dimension, and Offending Type Among System-Involved Youth
No full textVictimized youth face elevated risks of enduring adverse life outcomes that impede successful transition into adulthood, including a higher propensity for criminal involvement than non-victimized youth. Extant studies have since revealed a significant overlap among juvenile victim and offending populations (“victim-offenders”): youth, especially system-involved youth, encounter a range of victimization experiences, participate in a variety of crime, and their experiences of victimization vary in their capacity to induce a criminal response. So far, however, criminologists have devoted limited attention to examining such variations in the victimization-offending link that influence the prevalence and magnitude of the overlap. In fact, most victim-offender research either depends on general measures of victimization and offending or disproportionately focuses on violence, which neglects the heterogeneity in victimization histories and offending behaviors among victimized youths. To better understand the intra-group variability of the youth victim-offender group, a deeper look into how different victimization events motivate distinct types of offending is needed.
Using data from the Pathways to Desistance Study, this thesis describes and dissects differences in violent victimization experiences as a means of understanding variation in future criminal behavior among a high-risk, serious offending youth sample. The current research draws on general strain theory to assess whether differentiating degree of victimization and dimensions of victimization (i.e., frequency, variety, and recency) significantly affect predictions of participation in later violent and property crime. Findings indicate that differences in violent victimization experiences may shape meaningful differences in subsequent offending behaviors, but in nuanced ways. While the degree of victimization may matter, as there is notable variation in the magnitude of the effects of direct and vicarious victimization on both offending outcomes, such discrepancies do not appear to be statistically significant for this sample. Further, while specifying the strain dimensions of victimization may matter, as recency and variety dimensions distinguish degrees of vicarious victimization, they appear only to be salient considerations for less proximal exposure to violence (as opposed to direct experiences). Adding to the contributions of prior victim-offender literature, the results of this project highlight the importance of paying attention to sources of heterogeneity in the trauma histories of juvenile victim-offenders and bring to the forefront a discussion of the factors that may influence the historically robust association between victimization and crime
Deep Learning-Enabled Intelligent Goal-Oriented and Semantic Communication for 6G Networks
No full textIn the emerging sixth generation (6G) wireless networks, two major technological directions are expected to play pivotal roles: semantic communication and federated learning. Although independent in their approaches, both aim to make networks more intelligent, efficient, and adaptive to user and application demands. Semantic communication fundamentally shifts the focus from transmitting raw data to conveying only information that carries meaning and is directly relevant to a given task. This shift enables more efficient use of spectrum and reduces unnecessary transmission overhead, which is critical for supporting services such as augmented reality, holographic telepresence, and real-time control systems. In parallel, federated learning enables distributed edge devices to collaboratively train powerful machine learning models without sharing their raw data, thereby ensuring privacy, reducing communication burdens, and supporting large-scale, decentralized intelligence. Together, these advancements address key 6G challenges such as ultra-low latency, massive connectivity, and the need for real-time, context-aware decision-making at the edge. By integrating semantic understanding and distributed learning capabilities into the communication fabric, 6G networks will evolve from simple data carriers into intelligent infrastructures that empower a new generation of applications and services. This thesis addresses the design, analysis, and integration of advanced techniques that enable intelligent and efficient communication across multiple emerging 6G applications, including distributed learning, real-time immersive media, and multi-modal AI systems.
A primary contribution of this thesis lies in the development of semantic communication frameworks, where data is encoded and transmitted in a manner that preserves semantically important content while discarding redundant or irrelevant information. We propose an adaptive semantic encoder that utilizes attention mechanisms from vision-language models to dynamically identify and prioritize semantically salient regions in images and videos. By allocating communication resources according to semantic importance, our framework significantly reduces bandwidth consumption and latency, enabling critical 6G use cases such as holographic telepresence, remote surgery, and augmented/virtual reality with enhanced reliability and efficiency.
Parallel to semantic encoding, this thesis investigates federated learning (FL) as a foundational pillar for distributed intelligence in 6G. In FL, multiple devices collaboratively train a shared model without sharing their raw local data, offering strong privacy guarantees. However, practical challenges arise due to severe data heterogeneity across devices and limited wireless resources. To address these, we develop FedGradNorm, a dynamic gradient normalization scheme that harmonizes learning speeds across heterogeneous tasks, ensuring fair and efficient convergence in personalized federated multi-task learning (PF-MTL) settings. Furthermore, we propose a novel graph-based decentralized learning approach that adaptively adjusts communication topologies by automatically learning inter-client task correlations, thereby minimizing negative interference and enhancing convergence performance. These contributions enable FL frameworks that are communication-efficient, robust to statistical heterogeneity, and tailored for highly dynamic and bandwidth-limited 6G environments.
In addition to semantic encoding and federated learning, this thesis addresses the challenge of adaptive video streaming over wireless networks, a critical service for 6G that supports ultra-high-definition and interactive media applications. Traditional rate control algorithms lack the ability to dynamically adapt to rapidly changing network conditions and user experience requirements. We design a machine learning-based rate control algorithm that employs deep learning to jointly optimize video bitrate selection by considering real-time channel states, buffer dynamics, and semantic content complexity. Our approach achieves superior perceptual quality and stable playback, providing a seamless user experience even under highly variable wireless channel conditions.
Lastly, to ensure reliable and trustworthy information delivery in complex 6G multi-modal systems, we introduce RAG-Check, a novel framework for evaluating and mitigating hallucinations in multi-modal retrieval-augmented generation (RAG) systems. As 6G networks are envisioned to support advanced AI services that integrate visual, textual, and contextual data, ensuring the factual correctness and semantic alignment of generated responses becomes critical. RAG-Check rigorously analyzes the consistency between retrieved image evidence and generated textual outputs, providing quantitative and interpretable hallucination scores. This framework enhances the robustness and credibility of large-scale vision-language models in real-world deployments.
By cohesively integrating these contributions, this thesis offers a comprehensive design blueprint for goal-oriented and semantic communication systems in 6G networks. The thesis bridges the gaps between intelligent content encoding, collaborative distributed learning, adaptive multimedia delivery, and reliable multi-modal AI, collectively advancing the vision of 6G networks that do not merely transmit data but understand, reason, and act on information to effectively meet diverse user goals and application requirements, thereby laying the foundations for future intelligent communication systems that are efficient, scalable and also contextually aware and user-centric, enabling the next era of ubiquitous and intelligent connectivity
MINING AND TESTING ON HIERARCHICAL STOCHASTIC BLOCK MODELS
No full textThe rise in complexity of network data in neuroscience, social networks, and protein-protein interaction networks has been accompanied by efforts to model and understand
these data on different scales. A key multiscale network modeling technique posits a hier-
archical structure in the network. One such example of hierarchical modeling is the hierar-
chical stochastic blockmodel, which seeks to model complex networks as being composed
of community structures repeated across the network. Incorporating repeated structure al-
lows for parameter tying across communities, reducing the model complexity compared to
the traditional block model. In this work, we describe a model that naturally expresses net-
works as a hierarchy of sub-networks with a set of motifs repeating across it and formally
define the subgraph nomination framework with an emphasis on the notion of a user-in-
the-loop in the subgraph nomination pipeline
Welcome to 1856 Project Podcast Season One
No full textWelcome to the 1856 project podcast presented by the 1856 project, the University of Maryland's chapter of Universities Studying Slavery (USS). USS is a collaborative group of universities that work together to address the impact of human bondage within their respective institutions. Our podcast explores the history of African Americans in Maryland and on campus, examines the University's ties to American slavery, and presents our findings to those who can take action toward repair.https://soundcloud.com/1856projectpod/sets/1856-project-podcast-seaso