.png){kind=icon}
Return
Developing a Machine Learning Approach for Estimating Urban NO2 Concentrations Using Multi-Source Remote Sensing Data
.png){kind=icon}
OA
AI
DOI：10.3390/smartcities9100166.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate estimation of air pollutants, particularly nitrogen dioxide (NO2), is essential for assessing air quality and supporting informed policy decisions. This study investigates the application of machine learning techniques to estimate ground-level NO2 concentrations in the city of Zagreb, employing two ensemble machine learning methods: Random Forest (RF) and Extreme Gradient Boosting (XGB). The resulting models were developed using a multivariable modelling approach that integrated diverse data sources, including ground-based measurements from national air quality monitoring stations, publicly available satellite data collected through remote sensing (i.e., TROPOMI on Sentinel-5P and MSI on Sentinel-2), and meteorological and climate variables from the ERA5 reanalysis dataset and Copernicus Climate Change Service (C3S). These variables capture spatial, temporal, and atmospheric dynamics relevant to NO2 distribution across the urban area. Performance metrics indicate strong predictive capabilities, with the RF model achieving R2 = 0.74, RMSE = 9.86 µg/m3, and MAE = 6.89 µg/m3, while the XGB model yielded slightly better results (R2 = 0.78; RMSE = 8.76 µg/m3; MAE = 6.09 µg/m3). Analysis using SHAP values revealed that the tropospheric vertical column of NO2, elevation, and boundary layer height are the variables with greatest importance in both models. High-resolution maps of estimated ground-level NO2 concentrations were generated for the area of interest to gain insight into the spatial visualization of pollutant distribution across the city. Furthermore, a detailed seasonality analysis was performed, which indicated elevated NO2 levels during autumn and winter, with lower concentrations observed in spring and summer. These seasonal variations are likely driven by changes in meteorological conditions, emission sources, and atmospheric dispersion characteristics. This study has shown that high-quality models for local mapping the pollutant NO2 can be developed using data from air quality monitoring stations and publicly available data from the Copernicus programme.
Keywords:
air pollution
ground-level NO<sub>2</sub>
machine learning
SHAP analysis
TROPOMI
Journal
IF:
5.5
Papers:
954
Citations:
3.0K
Organization
Cited Papers
.png){kind=icon}
ShareSave
ShareSaveShareSaveShareSaveShareSaveShareSaveShareSave
View more