Cost-Effective Mapping of Hyperlocal Air Pollution Using Large-Scale Mobile Monitoring and Land-Use Machine Learning

Abstract

The advent of large-scale mobile monitoring using fast-response instruments has enabled hyperlocal mapping (≤100 m) of traffic-related air pollution (TRAP), with important implications for air quality management. However, most related studies have been confined within small areas due to the high cost and labor intensity. This study pioneers a cost-effective TRAP mapping method by incorporating large-scale mobile monitoring and land-use machine learning (LUML). Here, over 4.6 million 1 Hz high-frequency measurements (∼1300 h) were collected on a part of major roadways in the Chinese megacity of Shenzhen. Unmeasured locations were estimated by LUML models to reduce measurement costs and labor intensity. Various ML algorithms and varying spatial aggregation segment lengths were incorporated to optimize the model performance. Hyperlocal maps of NO, NO₂, and PM_2.5 were predicted across the entire road network covering over 1700 km². Based on our results, LU-RF (random forest) for mapping NO and NO₂ and LU-GBM (Gradient Boosting Machine) for mapping PM_2.5, demonstrated superior performance. Deep learning models, in contrast, did not yield comparable results. Finer partitioning of road segments (≤100 m) improved NO prediction performance, but worsened that for NO₂ and PM_2.5. By deployment of optimal ML algorithms and segment lengths, the TRAP mapping accuracy increased by 20–80% compared to conventional land-use regression models. This study provides a promising and cost-effective approach to hyperlocal air pollution mapping and management in cities worldwide.