Altitude-resolved prediction of roadside air pollution using UAV measurements and machine learning.

Kho, CJ; Seo, S; Hwang, H; Kim, SY; Jung, MH; Kang, HJ; Lee, JY · 2026 · PubMed Central

DOI: 10.1038/s41598-026-44153-9

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of characterizing and predicting the three-dimensional spatial distribution of roadside air pollution, which is often limited by fixed ground-based monitoring. The authors aimed to integrate Unmanned Aerial Vehicle (UAV) measurements with machine learning to estimate pollutant concentrations at various altitudes and distances from a major highway in South Korea. This approach seeks to overcome the spatial constraints of traditional monitoring and the operational limitations of UAVs, such as short flight endurance, by using predictive modeling to interpolate data. The research was conducted along the Gyeongbu Expressway, utilizing a DJI Matrice 300 RTK UAV equipped with an OA200M sensor to measure CO, NO₂, O₃, and PM2.5, alongside temperature and humidity. Measurements were performed during summer (August 2024) and winter (February 2025) campaigns. The UAV conducted vertical profiling from 10 to 60 meters altitude and horizontal profiling up to 60 meters from the roadside. Ground-level data were collected simultaneously, supplemented by traffic volume/speed data, meteorological records, and HYSPLIT backward trajectory analysis to assess long-range transport influences. To predict the ratio of UAV-to-ground pollutant concentrations, the authors trained four machine learning models: Random Forest, LightGBM, XGBoost, and CatBoost. These models utilized inputs including ground concentrations, traffic metrics, weather conditions, spatial coordinates, and seasonal indicators. The results revealed distinct spatiotemporal patterns in pollutant distribution. Concentrations of NO₂ and O₃ decreased with increasing altitude and distance from the road, while CO and PM2.5 showed more uniform vertical distributions. NO₂ levels peaked at 7 a.m. due to morning traffic congestion, whereas O₃ was lowest at that time due to reduced photochemical activity. Seasonally, O₃ and PM2.5 concentrations were higher in summer, with HYSPLIT analysis confirming that summer PM2.5 increases were influenced by long-range transport from eastern China. In contrast, winter NO₂ levels were higher due to atmospheric stagnation. Among the machine learning models, CatBoost demonstrated the highest predictive performance, achieving R² values between 0.65 and 0.95. SHAP analysis indicated that ground-level concentrations and meteorological factors were primary drivers for NO₂ and PM2.5 predictions, while O₃ predictions relied heavily on meteorological and seasonal features. CO predictions were notably influenced by traffic speed, which showed a negative association with the UAV/Ground ratio. The significance of this work lies in the development of a novel framework that combines UAV observations with machine learning to provide high-resolution, three-dimensional estimates of roadside air quality. By successfully predicting vertical pollutant distributions using ground-based data, the study offers a scalable method to expand the spatial applicability of UAV measurements. This approach enhances the understanding of traffic-related emission dispersion and provides a robust tool for assessing air quality in areas where direct UAV measurement is operationally difficult or restricted.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• exposure measurement