True Absence or Detection Failure: Generative Augmentation and Expert-Calibrated Filtering for Long-Tailed Bumblebee Classification
Name
sun-msun14-mcp-dusp-2026-thesis.pdf
Size
7.77 MB
Format
Adobe PDF
Checksum (MD5)
ba5225f94604125e43f7a4d802dbca45
Author(s)
Sun, Mingyang
Advisor(s)
Duarte, Fabio
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Urban biodiversity monitoring increasingly relies on automated image classification, yet rare species remain difficult to detect under long-tailed data distributions. We study 16 Massachusetts Bombus species with an approximately 60:1 class imbalance, where rare taxa are often confused with morphologically similar common species, obscuring true absence
from detection failure. Recent advances in multimodal generative image models make it possible to synthesize visually plausible rare-class examples from text prompts and reference images, motivating their use for long-tailed biodiversity classification. We evaluate a pipeline
combining morphology-guided generation, LLM-based quality filtering, and expert-calibrated BioCLIP filtering. Results show that synthetic data is not automatically useful: copy-paste provides the clearest statistically supported gain, while generated augmentation can reduce rare-tier performance. Overall, we argue that useful augmentation depends not on image volume alone, but on expanding visual diversity without corrupting species-diagnostic morphology.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Massachusetts Institute of Technology. Department of Urban Studies and Planning
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
