Harmonised satellite, meteorological and ground-observation database for 128 cities, with surface NO₂ estimates for 100 cities

Air quality estimation failure in scarcest monitoring networks: a climate-dependent limit on satellite-to-surface NO₂ inversion Satellite spectrometers measure tropospheric NO₂ columns, whereas exposure assessment requires surface concentrations. Machine-learning inversion is widely used to bridge this gap and often reports high predictive skill; however, validation is typically conducted within the monitoring network used to train the model. The extent to which these models transfer to the data-scarce regions where they are most needed remains unquantified. We assembled a harmonised database covering 128 cities from 2019 to 2025, combining Sentinel-5P TROPOMI NO₂ columns, MERRA-2 meteorology, Global Human Settlement Layer urban morphology, and surface observations from four national monitoring networks in the United States, China, India, and the Arabian Gulf. We fitted gradient-boosted trees using a validation protocol specified before modelling. Under leave-one-city-out validation across 35 temperate and subtropical cities, the inversion achieved an R² of 0.570 ± 0.004, outperforming climatology (0.437) and persistence (0.347). The same model and data achieved an R² of 0.840 under temporal hold-out, indicating that the gap between reported and deployable skill reflects validation design rather than model choice. When applied to five hot, arid cities withheld from all model development, the inversion performed poorly (R² = −0.83; bias = −7.4 ppb). Across all 54 cities with ground observations, the within-city correlation between column and surface concentrations declined systematically with mean air temperature (r = −0.707, p = 2.3 × 10⁻⁹), suggesting that expected performance can be predicted from climate. A climate-gated, two-stage model substantially reduced the arid-region bias to −0.23 ppb but recovered little variance (R² = 0.15); performance also did not improve as arid training cities were added. These results indicate that transferability is constrained primarily by column-to-surface physics, rather than model capacity or training-set size. We provide surface NO₂ estimates for 100 cities, each accompanied by two independent confidence measures that diverge in the cities of greatest interest. These are the derived datasets for this study.