Sara Davies
MIT Department: Urban Studies and Planning
Faculty Mentor: Prof. Fabio Duarte
Research Supervisor: Martina Mazzarello, Diego Morra, Titus Venverloo
Undergraduate Institution: Rice University
Website:
Biography
Sara Caetano Davies is a rising junior at Rice University, studying Computer Science andCognitive Sciences. Originally from Lisbon, Portugal, she now lives in Boston. This coming year, she will study at the University of Cambridge as part of the Abraham-Broad Fellowship. Her academic interests center on human-computer interaction. Sara began research in highschool at the Laboratory for Laser Energetics in computational plasma physics, then worked on understanding educational inequities at the Levine Center to End Hate. Last summer, she developed a large-scale study of advertising’s impact on consumer welfare at Carnegie Mellon University. At Rice, she works in the Computational Wellbeing Group researching how stress impacts spousal Alzheimer’s caregivers. At MIT’s Senseable City Lab, she works on SidewalkAI and Sensing Garden, projects using computer vision to improve urban accessibility and biodiversity monitoring. She plans to continue pursuing human-centered research. Outside academics, she enjoys sailing, traveling, and reading.
Abstract
Towards Real-Time Urban Biodiversity Monitoring
Sara Davies1, Titus Venverloo2, and Fabio Duarte2
1Department of Computer Science, Rice University
2Department of Urban Studies and Planning, Massachusetts Institute of Technology
Insect biodiversity is globally declining, posing ecological and economic threats. Although methods to monitor and aid in preservation of insect species are being developed, their adoption has been slow and limited to specific use cases. To address these challenges, we develop a real-time, large-scale insect species identification model built on YOLOv8 architecture, a real-time object detection system. The model is trained using images sourced from the Global Biodiversity Information Facility (GBIF), which contains over 16 million instances. However, this dataset includes image quality issues such as non-adult life stages, multiple insects in a single frame, miscentered images, and mislabeled specimens, all of which hinder model performance. To address this, we explore filtering techniques using GBIF's lifeStage metadata tag to retain only adult specimens, combined with color-based filtering to remove low-quality images. These steps aim to reduce dataset noise and enhance detection reliability. By improving the accuracy and scalability of insect identification, this work represents a step toward automated biodiversity monitoring to support conservation efforts.