Abstract
TEMPO (Tropospheric Emissions: Monitoring of Pollution), a geostationary satellite instrument, enables continuous daytime monitoring of NO2 over North America. Here, we develop a novel deep learning (DL) framework to extract hourly (daylight) ~2×2 km2 surface NO2 concentrations from TEMPO tropospheric column density of NO2 (TrpNO2) for August 2, 2023, to December 31, 2024. First, we impute low-quality and missing pixels in TEMPO NO2 columns, improving performance against Pandora from a correlation coefficient (R) of 0.40 and a percentage mean absolute bias (MABP) of 101% for low-quality pixels to R=0.69 and MABP=34%. This gap-filling step is critical, as it recovers approximately 32% of pixels in our TEMPO domain. Second, we fuse gap-free TrpNO2 with in-situ NO2 observations to generate preliminary surface NO2 grids (”TEMPO-Fused NO2”). Third, a residual neural network (ResNet) incorporates TEMPO-Fused NO2 and other variables to generate the final surface NO2 product over the TEMPO domain. Unlike point-based NO2 monitoring stations, our product provides gap-free surface NO2 maps across TEMPO domain, substantially enhancing NO2 monitoring capabilities. Validation against in-situ stations shows a very high accuracy of this DL framework at surface NO2 estimation (R=0.98, mean absolute bias (MAB)=0.8 ppb, and MABP=12%). The model also generalizes well spatially (R=0.95; MABP=21%) and temporally (R=0.98; MABP=13%) and remains accurate within imputed regions (R=0.97; MABP=17%) and rural environments (R=0.93; MABP=25%). The DL framework accurately captures surface NO2 concentrations during high-NO2 episodes (R=0.94; MABP=12%). We also provide hourly grids of NO2 estimation epistemic uncertainty to aid users in identifying pixels that require cautious interpretation.

This publication has 39 references indexed in Scilit: