Integrating Sentinel-5P Satellite Observations and Machine Learning for Estimating Ground-Level NO 2 Concentrations: A Case Study of Urmia, Iran

Abstract The process of monitoring nitrogen dioxide (NO 2 ) levels in areas that lack sufficient ground-based measurement stations presents significant difficulties for effective air quality control. The TROPOMI sensor on Sentinel-5P satellite provides a comprehensive spatial monitoring system but requires ground-based data and meteorological information to correctly analyze tropospheric column measurements and surface-level concentration data. The researchers assessed how effectively satellite data and ground-based information and meteorological data work together within a machine learning system to predict surface NO 2 levels in Urmia city from 2019 to 2024. The researchers used Google Earth Engine to extract Sentinel-5P satellite products which included NRTI and OFFL and they combined this information with data from the Urmia Air Quality Monitoring Station and the Urmia Airport Synoptic Station. The machine learning models which included Decision Tree Random Forest Extra Trees Gradient Boosting XGBoost CatBoost and AdaBoost were trained using an 80% training and 20% testing split after the researchers completed data normalization through incomplete and outlier record removal and Pearson correlation coefficient (|R| ≥ 0.10) variable selection. The researchers used R 2 , RMSE, and MAE metrics to evaluate performance. The tested models showed Gradient Boosting as the most effective model because it produced R 2 = 0.9961, MAE = 0.0212 and RMSE = 0.0578 results for NRTI data and R 2 = 0.9959, MAE = 0.0290 and RMSE = 0.0634 results for OFFL data. The researchers used scatter plots and residuals and error distributions to establish that their study results had no systematic bias and no overfitting problems. The results demonstrate that Sentinel-5P data combined with ground-based measurements and meteorological data through Boosting algorithms, especially Gradient Boosting, enables accurate surface-level NO 2 estimation in cities that lack ground monitoring stations. The development of region-specific models requires local data in Urmia because the city has distinct climatic conditions and pollutant emission patterns.