Background Air pollution modeling generally suffers from the limited and uneven distribution of fixed air quality monitoring (AQM) stations. Low-cost sensors have emerged to fill these monitoring gaps; however, few studies have integrated low-cost sensor data with satellite-based measurements. Objective We aimed to develop a daily 250-m resolution fine particulate matter (PM2.5) estimation model in Taiwan by incorporating low-cost sensor data with aerosol optical depth (AOD), meteorological variables, and land-use data. Methods PM2.5 measurements from 16,795 sensors and 74 AQMs collected between 2019 and 2021 were used. Multi-Angle Implementation of Atmospheric Correction (MAIAC) satellite AOD data were obtained. A highly efficient LightGBM model was applied to correct bias in low-cost sensor measurements, accounting for distance to AQMs and meteorological variables. Subsequently, a second LightGBM model was employed to develop the PM2.5 estimation model. Model performance was evaluated using the overall, temporal, spatial 10-fold, and rolling-origin cross-validation (CV). SHapley Additive exPlanations (SHAP) summary plots and accumulative local effect plots were used to interpret feature contributions Results The overall, temporal, spatial 10-fold, and rolling-origin CV coefficient of determination (R2; root mean square error [RMSE]) were 0.98 (1.23 μg/m3), 0.74 (4.62 μg/m3), 0.93 (2.48 μg/m3), and 0.61 (5.77 μg/m3), respectively. The ten most important predictors were: day of the year, u- and v-wind components, relative humidity, year, total precipitation, surface pressure, boundary layer height, AOD, and temperature. Higher PM2.5 estimates were observed in the western region compared to the east, with the southwestern region being the most polluted. Significance This efficient model provides PM2.5 estimates that can be used to identify hotspots, abnormal events, and potential emission sources. The ultra-high-resolution estimates can also inform design of healthier walking routes for general public. Furthermore, the model enables the definition of short- and long-term exposure metrics and supports back-extrapolate historical PM2.5 exposures in epidemiological studies. Impact A highly efficient LightGBM model integrating low-cost sensor data with satellite-based measurement was developed and validated. The model demonstrated excellent performance, with a 10-fold CV R2 (RMSE) of 0.98 (1.23 μg/m3). By integrating Internet of Things (IoT), this approach mitigates the monitoring gaps, facilitates the identification of pollution hotspots, abnormal events, and sources, and enhances exposure assessment in epidemiological research.

Development of a high-resolution PM_2.5 model in Taiwan using integrated low-cost air quality sensors and satellite observations

期刊论文
发表日期：

2026年7月22日

查看全文

Chau-Ren Jung,
Wen-Hsuan Chuang,
Wei-Ting Chen,
Wei-Hsuan Lin,
Re-Yang Lee,
Bing-Fang Hwang

摘要

Background

Air pollution modeling generally suffers from the limited and uneven distribution of fixed air quality monitoring (AQM) stations. Low-cost sensors have emerged to fill these monitoring gaps; however, few studies have integrated low-cost sensor data with satellite-based measurements.

Objective

We aimed to develop a daily 250-m resolution fine particulate matter (PM_2.5) estimation model in Taiwan by incorporating low-cost sensor data with aerosol optical depth (AOD), meteorological variables, and land-use data.

Methods

PM_2.5 measurements from 16,795 sensors and 74 AQMs collected between 2019 and 2021 were used. Multi-Angle Implementation of Atmospheric Correction (MAIAC) satellite AOD data were obtained. A highly efficient LightGBM model was applied to correct bias in low-cost sensor measurements, accounting for distance to AQMs and meteorological variables. Subsequently, a second LightGBM model was employed to develop the PM_2.5 estimation model. Model performance was evaluated using the overall, temporal, spatial 10-fold, and rolling-origin cross-validation (CV). SHapley Additive exPlanations (SHAP) summary plots and accumulative local effect plots were used to interpret feature contributions

Results

The overall, temporal, spatial 10-fold, and rolling-origin CV coefficient of determination (R²; root mean square error [RMSE]) were 0.98 (1.23 μg/m³), 0.74 (4.62 μg/m³), 0.93 (2.48 μg/m³), and 0.61 (5.77 μg/m³), respectively. The ten most important predictors were: day of the year, u- and v-wind components, relative humidity, year, total precipitation, surface pressure, boundary layer height, AOD, and temperature. Higher PM_2.5 estimates were observed in the western region compared to the east, with the southwestern region being the most polluted.

Significance

This efficient model provides PM_2.5 estimates that can be used to identify hotspots, abnormal events, and potential emission sources. The ultra-high-resolution estimates can also inform design of healthier walking routes for general public. Furthermore, the model enables the definition of short- and long-term exposure metrics and supports back-extrapolate historical PM_2.5 exposures in epidemiological studies.

Impact

A highly efficient LightGBM model integrating low-cost sensor data with satellite-based measurement was developed and validated. The model demonstrated excellent performance, with a 10-fold CV R² (RMSE) of 0.98 (1.23 μg/m³). By integrating Internet of Things (IoT), this approach mitigates the monitoring gaps, facilitates the identification of pollution hotspots, abnormal events, and sources, and enhances exposure assessment in epidemiological research.

Development of a high-resolution PM2.5 model in Taiwan using integrated low-cost air quality sensors and satellite observations

摘要

Development of a high-resolution PM_2.5 model in Taiwan using integrated low-cost air quality sensors and satellite observations