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

Abstract

The generation of reference data for machine learning models is challenging for dust emissions due to perpetually dynamic environmental conditions. We generated a new vision dataset with the goal of advancing semantic segmentation to identify and quantify vehicle-induced dust clouds from images. We conducted field experiments on 10 unsealed road segments with different types of road surface materials in varying climatic conditions to capture vehicle-induced road dust. A direct single-lens reflex (DSLR) camera was used to capture the dust clouds generated due to a utility vehicle travelling at different speeds. A research-grade dust monitor was used to measure the dust emissions due to traffic. A total of ~210,000 images were photographed and refined to obtain ~7,000 images. These images were manually annotated to generate masks for dust segmentation. The baseline performance of a truncated sample of ~900 images from the dataset is evaluated for U-Net architecture.