UM-CIP Program
Developing a Scalable Bioacoustic Monitoring Pipeline for Rare Tropical Endemics (Michael Ocasio)
Tropical biodiversity hotspots face a critical data gap; many endemic species remain "data deficient" because their low encounter rates and remote habitats make traditional field surveys logistically and financially prohibitive at scale. This project aims to bridge this gap by developing and validating a high-performance computing (HPC) bioacoustic monitoring pipeline, using the endemic Puerto Rican Oriole (Icterus portoricensis) as a model system. Building upon three years of intensive site-specific field data and a preliminary BirdNET-based classifier, this research will leverage UMBC’s HPC cluster to scale inferences to an island-wide level.
The proposed workflow integrates feature embedding extraction via the BirdNET foundational model with FAISS-indexed vector similarity searches and Slurm-aided unsupervised clustering to analyze millions of minutes of passive acoustic data. These “Big Data” approaches will enable the identification of rare vocalizations across heterogeneous soundscapes and allow for mathematical corrections of environmental noise (e.g., wind, sea, and insect communities). The ultimate goal is to produce a validated, open-source pipeline that translates raw audio into accurate species identification and population estimates. This framework will provide a sustainable, scalable solution for monitoring rare tropical birds and will be disseminated through training initiatives across the Caribbean to empower local conservation efforts.
DS & AI Program
Anthromes Analysis and Shared Landscapes: Geographic Data Analysis and Mapping (Renee Shalom)
The goal of this work is to reproduce and revise geographic data analyses used in the Anthromes and Shared Landscapes projects, including the Anthromes revision that is being planned for publication. This includes understanding map projections and discrete global grid systems, reproducing and updating published Anthromes figures, and exploring how high- resolution geographic data can be processed and visualized. The work will help develop skills in GIS, spatial data processing, global mapping, research reproducibility, and the use of computational tools and AI-assisted workflows. It will also support professional development through collaboration, documentation, and preparation of research outputs.
Understanding Students’ Learning Journey in Personal Informatics Project: Implications for Design (Emma Shay)
This project analyzes field data collected from the Personal Informatics project in the introductory data science course IS 296 to inform the human-centered design of an agentic AI system to support personal informatics for a broader audience. The data from Spring 2025 and Fall 2025, which includes students’ planning documents, end-of-project reflections (both qualitative and quantitative), students’ weekly reflection submissions, and creative artefacts, are used to conduct a qualitative analysis, contributing to developing a coding scheme and coding result.
Stream 1: Automated Spreadsheet-to-Web Catalog Pipeline for iHARP Literature, Deepfake Models & Curriculum Portals; Stream 2: Reproducible AI Workflows for the iHARP Morpholayer/ISLAY ML Challenge (Nebiyeleul Yifru)
Stream 1 – Automated Spreadsheet-to-Web Catalog Pipeline (iHARP Catalog): Build an automated pipeline that parses structured spreadsheet data (Google Sheets published as CSV/JSON and multi-sheet Excel workbooks) and generates an interactive, card-based web catalog that updates whenever the underlying data source is modified. The portal organizes research and educational resources on audio/visual deepfake detection (Papers with Data, Detection Models, Generation Models) and Linguistics & AI Syllabi into an accessible, searchable interface.
Stream 2 – iHARP Morpholayer / ISLAY Reproducible-Workflow ML Challenge: Design and support a multi-step reproducible AI workflow competition focused on tracing internal ice layers in polar radargrams, addressing preprocessing, layer segmentation, spatial crossover validation, out-of-distribution generalization, and evaluation metrics.
Sea Level Rise (Sharddha Baral)
This project focuses on the development of data products and visualizations to communicate the impacts and patterns of sea level rise in Maryland. Using Python and data analysis techniques, the project supports the exploration and presentation of sea level rise data through accessible visualizations that can help illustrate changes and potential impacts across the state.
iHARP Polar Science Chatbot: A Grounded Retrieval-Augmented Generation System for Polar Science Literature (Molly Balkan)
This project explores the development of a Retrieval-Augmented Generation (RAG) chatbot designed to make iHARP’s polar science publications more accessible and support information discovery across a growing body of specialized scientific literature. A central focus is improving the reliability of AI-generated responses by investigating and mitigating hallucinations in RAG systems. The project examines common sources of retrieval and generation errors, including retrieval misses, fragmented document chunks, cross-document contamination, and unsupported model-generated information. It explores several approaches to improving response accuracy, including keyword-overlap grounding, knowledge graph integration, and hybrid retrieval combining dense embeddings with keyword-based search. A gold-standard evaluation benchmark supports systematic assessment of these approaches, while automated publication collection and document processing support the continued expansion of the knowledge base.