COMPREHENSIVE ANALYSIS OF BIAS IN TEMPO NO₂ COLUMN DENSITIES & GENERATING DEEP-LEARNING-BASED TEMPO PRODUCTS
Tropospheric Emissions: Monitoring of Pollution (TEMPO) is the first geostationary satellite instrument to monitor air pollutants across North America. This study uses Pandora observations to analyze the bias in TEMPO Level-3 total column density of NO2 (TOTNO2) from August 2023 to December 2024. TEMPO achieves high accuracy at 5% cloud-filtering threshold: correlation coefficient (R) of 0.86, index of agreement (IOA) of 0.91, mean absolute bias (MAB) of 1.423×1015 molecules/cm2, and a percentage MAB (MABP) of 23.1%, corresponding to a 12% underestimation. Accuracy decreases when pixels with greater cloud-cover are included. Solar zenith angle (SZA) of 10–20° yields the highest accuracy (R: 0.87, MABP: 22.7%), whereas SZAs of 70–80° yield the lowest (R: 0.71, MABP: 35.2%). Consequently, early-morning or near-sunset observations are less reliable than midday. This discrepancy could stem from inaccurate simulation of diurnal variations in the boundary-layer height in the a-priori, and from larger uncertainties in radiative transfer at high SZAs. TEMPO overestimates TOTNO2 at low NO2 levels and underestimates at high levels, with maximum biases of +16% (low) and -31% (high), respectively. Station-to-station performance varies considerably, with R ranging from 0.29 to 0.84 and MABP from 14.9% to 49.3%. Stations situated at higher altitudes relative to the ground show reduced agreement with TEMPO, as Pandora cannot detect NO2 below the instrument’s altitude, whereas TEMPO retrieves the full column. Validation of TEMPO TOTNO₂ at TROPOMI overpass time indicates that TEMPO’s performance relative to Pandora (IOA: 0.93, MABP: 22.3%) closely matches that of TROPOMI (IOA: 0.92, MABP: 20.1%). Insights from this bias analysis were incorporated into deep learning (DL) models to impute missing TEMPO TOTNO₂ pixels. The imputed data achieve similar accuracy to the original measurements. These gap-filled TEMPO TOTNO₂ products were subsequently used in an advanced DL model to generate hourly (daylight) surface NO₂ maps with high accuracy (R = 0.96, IOA = 0.98, MABP = 16%). Spatial cross-validation further confirms the strong generalization capability of the DL model, demonstrating high performance even in regions where the model was not trained.