Search

While low-cost sensors (LCSs) and satellite retrievals are valuable supplements to regulatory air quality monitoring stations (AQMs), measurements from LCSs and satellite retrievals suffer from considerable bias and uncertainty. Here, we proposed a machine learning-based approach named the Fusion-Imputation-Gradient-Boosting-Machine (FI-GBM) model which fused the NO 2 measurements from AQM, LCS, and the TROPOspheric Monitoring Instrument (TROPOMI) for mapping hourly ground-level NO 2 at 1 km resolution. Based on the machine-learned relationships among AQM, LCS, TROPOMI measurements, and environmental covariates, the LCS and TROPOMI data were assimilated into AQM data. We selected Tangshan, an industrial city in North China, for the demonstration. The FI-GBM model showed high predictive performance in the sample-based cross-validation (R 2 = 0.89). The R 2 values of the cell-, area-, and month-based cross-validations were 0.67, 0.59, and 0.64, respectively. Fusing LCS and TROPOMI data improved the predictive performance compared to the benchmark models using neither or only one of them. The FI-GBM model showed decent utilization of the strengths of TROPOMI and LCS in regional and local-scale monitoring, respectively. It is noteworthy that the FI-GBM model could automatically filter noisy samples from LCS data, which was critical for discriminating between true and false-positive pollution hotspots. This study provides a data-noise-reduction approach for fusing multisource measurements in order to identify pollution hotspots and trace pollutant sources, thereby promoting cleaner production. [Display omitted] • NO 2 data from standard sites, satellite retrievals, and low-cost sensors were fused. • Hourly NO 2 was mapped at 1-km resolution for an industrial city in North China. • Satellite retrievals mitigated overall estimation bias. • The new algorithm can automatically screen noise in low-cost sensor data. • Low-cost sensor data helped identify pollution hotspots in industrial areas. [ABSTRACT FROM AUTHOR]