Identifying Local Emission Sources via an Integrated Computer Vision and Vision–Language Models (VLMs) Approach

JINTAO GU, Hongpufan Huang, Boyuan Niu, K. Max Zhang, Cornell University

     Abstract Number: 528
     Working Group: Advancing Aerosol Science through Data Analysis Tools

Abstract
The current National Emissions Inventory (NEI) estimates the number of wood stove appliances in each U.S. County by applying a fraction, derived from a survey of 2,984 responses across 21 states, to the number of occupied homes. However, the limited sample size and the methodology’s inability to capture sub-county spatial variation hinder accurate estimation of woodsmoke-related PM emissions, which is critical for addressing neighborhood-level exposure equity issues and guiding policy changes. To overcome this limitation, we propose identifying wood stove units through visible residential rooftop features. Freestanding wood stove units often have distinctive, tall metal chimneys that comply with specific codes, as verified by our field survey in upstate New York. Our framework consists of two components: (1) data sources, including data collection methods, and (2) chimney identification techniques. In a pilot study, we evaluated the potential of vehicle-based videos, drone footage, and street view imagery, and assessed their adaptability to both urban and rural environments. For the chimney identification task, we propose combining Vision–Language Models (VLMs) with object detection computer vision (CV) techniques such as YOLOv11. Many challenges in aerosol source identification, including wood stove chimney detection, involve ambiguous appearance criteria and highly heterogeneous data sources, which complicate traditional object detection methods. VLMs, which can leverage contextual understanding and accommodate less rigid object definitions, demonstrate strong potential for accurate detection in such cases. We recommend broader adoption of VLMs in similar PM source estimation applications where context-driven identification is necessary.