Adapting Public Annotated Data Sets and Low-Quality Dash Cameras for Spatiotemporal Estimation of Traffic-Related Air Pollution: A Transfer-Learning Approach

Abstract

This study investigated the utilization of images collected from low-quality dash cameras on passenger vehicles for the estimation of traffic-related air pollution (TRAP). We conducted mobile monitoring along Taiwan Avenue, Taichung, Taiwan, and collected pollution concentration data including carbon dioxide (CO2), nitrogen oxides (NOx), black carbon (BC), and particle number (PN). Dash cameras record images that reveal the environment through which the vehicle passes. Image semantic information such as the proportion of sky, buildings, traffic, and vegetation can be extracted through deep learning models. Training of deep learning models requires the pixel-level labeling of each image, which is labor intensive. We propose the use of publicly available data sets for the training of the deep learning model. Transfer learning was utilized to customize the model for locally collected, unlabeled, low-quality dash camera images. TRAP was estimated with a hybrid model consisting the land-use regression (LUR) and image semantic information. With a five-fold cross-validation, the hybrid model with transfer learning resulted in improved R2 values for CO2 (R2=0.81), NOx (R2=0.64), PN (R2=0.65), and BC (R2=0.87). Public labeled data sets and transfer learning may be helpful when labeled data are difficult to acquire in the local region. This work demonstrates the adaptation of image semantic information, extracted from videos captured from vehicle dash cameras, into a LUR model to improve pollutant estimation.