Abstract

Simultaneous prediction of multiple air pollutants is essential for quantifying human co-exposure and evaluating the health impacts of pollutant mixtures. However, spatial and temporal gaps in geostationary satellite observations, chemical transport models, and ground-based monitoring networks hinder accurate hourly assessments of multi-pollutant dynamics. Here, we present Deep Learning for Multiple Air Pollutant analysis (DeepMAP), a deep learning framework that simultaneously predicts six major air pollutants─PM10, PM2.5, O3, NO2, CO, and SO2─at hourly resolution. DeepMAP demonstrated robust performance across multiple pollutants and generalized well to unseen regions. The framework accurately captured dynamic high-concentration co-pollution episodes during March 2021, with normalized RMSE values below 0.36 for all pollutants. DeepMAP revealed that PM10-PM2.5 co-exceedance was the most frequent across East Asia (91 days/year), followed by PM10-PM2.5-NO2 (42), PM2.5-O3 (18), and PM10-PM2.5-O3 (12). Hotspots for PM10-PM2.5-NO2–O3 co-exceedance were identified over the North China Plain, East China, and South Korea, where the regional annual totals reached 24, 19, and 15 days, respectively. A novel co-exposure index further identified three distinct hotspot regions where the contribution of NO2 was approximately twice that observed elsewhere. Our findings provide a high-resolution, data-driven framework for characterizing multi-pollutant co-exposure and identifying regional priorities for air quality management and public health protection.

A research team, led by Professor Jungho Im from the Department of Civil, Urban, Earth, and Environmental Engineering at UNIST, has introduced DeepMAP, a cutting-edge artificial intelligence model that accurately estimates hourly levels of six key air pollutants across East Asia. The technology, developed from 2021 to 2023, provides new insights into the widespread and simultaneous exceedance of air quality standards, with significant implications for health policies and environmental management. 

DeepMAP integrates diverse data sources—including geostationary satellite imagery, atmospheric chemical transport models, meteorological data, and ground observations—to produce real-time, high-resolution maps of PM10, PM2.5, O₃, NO₂, SO₂, and CO. Operating at a 10 km spatial resolution and providing hourly predictions, the model captures dynamic pollution patterns and hotspots, revealing that Korea experiences about 15 days per year with four pollutants exceeding WHO safety thresholds simultaneously. Unlike traditional methods that estimate pollutants individually, DeepMAP’s multi-task learning approach models interactions among pollutants, significantly enhancing estimation accuracy. 

“By accurately capturing the complex interplay of multiple pollutants in real time, our model offers a powerful tool for assessing exposure risks and guiding effective policy decisions,” said Professor Jungho Im, lead researcher. “This represents a major step toward more realistic and comprehensive air quality management.” 

The study underscores the health risks posed by combined pollutant exposure, which can worsen respiratory and cardiovascular diseases. Traditional monitoring methods often fall short in providing detailed, regional, and real-time data. DeepMAP’s capabilities open new avenues for environmental monitoring, public health research, and proactive policymaking, especially during pollution episodes driven by seasonal phenomena like dust storms and high-pressure systems. 

The findings of this research have been published in Environmental Science & Technology on March 20, 2026. The study has been supported by the National Institute of Environmental Research (NIER) under the Ministry of Environment (ME), and by the National Research Foundation of Korea (NRF), funded by the Ministry of Science and ICT (MSIT).

Journal Reference

Eunjin Kang, Sihun Jung, Jungho Im, et al., "Quantifying Multi-pollutant Co-exposure via Deep Learning-Based Simultaneous Prediction Using Geostationary Satellite Data," Environ. Sci. Technol., (2026).