A benchmark dataset for binary segmentation and quantification of dust emissions from unsealed roads

Area of Science:

• Environmental Science
• Computer Vision
• Machine Learning

Background:

• Generating reference data for machine learning models is challenging due to dynamic environmental conditions, especially for dust emissions.
• Accurate quantification of vehicle-induced dust is crucial for environmental monitoring and impact assessment.

Purpose of the Study:

• To develop a novel vision dataset for advancing semantic segmentation techniques.
• To enable the identification and quantification of vehicle-induced dust clouds from images.

Main Methods:

• Conducted field experiments on 10 unsealed road segments with diverse surface materials and climatic conditions.
• Captured images of dust clouds using a DSLR camera while a utility vehicle traversed the roads at varying speeds.
• Utilized a research-grade dust monitor to measure traffic-related dust emissions.
• Manually annotated approximately 7,000 refined images to create dust segmentation masks.

Main Results:

• A dataset of ~7,000 manually annotated images was generated from an initial ~210,000 photographs.
• The dataset captures vehicle-induced road dust under varied environmental and road conditions.
• Baseline performance evaluation of the U-Net architecture was conducted on a subset of ~900 images.

Conclusions:

• The developed dataset provides a valuable resource for training and evaluating machine learning models for dust emission analysis.
• This work advances the capability of semantic segmentation in identifying and quantifying airborne particulate matter from traffic.
• The dataset facilitates further research into mitigating environmental impacts of road dust.