Deep learning approach for reconstructing three-dimensional distribution of NO2 on an urban scale

The emission, transmission, and secondary generation of atmospheric pollutants occur not only in proximity to the ground but also at elevated altitudes. Vertical distribution plays a pivotal role in understanding the intricate mechanisms that govern atmospheric pollutants. Although ground-based remote sensing offers valuable insights into vertical pollutant profiles, it is limited to obtaining vertical profiles at specific locations, which cannot capture the spatially continuous distribution of vertical profiles at the city scale. Thus, these methods cannot fully support the understanding of high-altitude transport and vertical exchange of atmospheric pollutants. This study proposes a multimodal intermediate fusion (MIF) architecture based on a residual network. Using inputs such as satellite data and meteorological parameters tied to profiles retrieved by multi-axis differential optical absorption spectroscope (MAX-DOAS), we reconstruct the daytime high spatial resolution (1 × 1 × 0.1 km) and temporal resolution (15 min) NO 2 spatiotemporal distribution. The mean correlation between MIF reconstructions and ground-based in situ observations from the China National Environmental Monitoring Center reaches 0.813, and the correlation with TROPOMI tropospheric NO 2 data reaches 0.750–0.796 in the different distance groups. The MIF performance significantly exceeds that of the WRF-Chem model in terms of resolution and accuracy. The full-coverage high-resolution 3D NO 2 distribution allows us to analyze the distribution of NO 2 and estimate its transport, revealing an annual net output of 0.780 Gg of NO 2 from Hefei's urban areas. The vertical transport at high altitudes in urban areas is 28.2 % higher than that in the suburban regions. Notably, the annual high-altitude NO 2 transport from 8:00 to 16:00 local time constitutes approximately 23.7 % of the annual NO X emissions in the Hefei urban area. During pollution events, the MIF reconstructions can identify two typical processes: 1) transport from elevated altitudes to ground level, and 2) external transport from the urban surface to high altitudes and then transport to suburban areas from elevated altitudes, ultimately reaching the suburban surface. The 3D NO 2 distribution reconstructed by MIF may help understand the impact of high-altitude transport and the vertical exchange of pollutants on atmospheric pollution and can serve as a scientific basis for air pollution control policies. • Deep learning is integrated with hyperspectral remote sensing techniques. • A high-resolution 3D distribution of NO 2 on an urban scale is reconstructed. • MAX-DOAS profiles, satellite data and WRF-Chem results are fused. • Two typical transport processes are identified.