A gapless 0.05° daytime-hourly tropospheric NO ₂ dataset (2019–2024) over key Asian hotspots reconstructed using physics-aware deep learning

High-frequency observations of tropospheric nitrogen dioxide (NO ₂ ) support regional air quality assessment and analysis of its photochemical evolution. Although the Geostationary Environment Monitoring Spectrometer (GEMS) and other new-generation geostationary satellites provide hourly observations during daytime, their spatial coverage is often severely compromised by cloud obstruction and inherent retrieval limitations. Furthermore, their short observation history hinders long-term, cross-regional environmental assessments. To address this, this study presents a gapless, 0.05° daytime-hourly tropospheric NO ₂ vertical column density (VCD) dataset for representative hotspots in Asia spanning 2019 to 2024, reconstructed using a physics-aware deep learning framework. The framework integrates partial convolutions to handle irregularly missing satellite observations and proposes a Physics-aware Normalization (PhysNorm) module. Specifically, the model inputs comprise the 0.05° TROPOMI tropospheric NO ₂ VCD and its validity mask, together with 0.25° predictors including ERA5 meteorological variables, EAC4 total-column NO ₂ , land cover, elevation, and the cosine of the solar zenith angle. The supervised training target is the hourly GEMS tropospheric NO ₂ VCD from 2023. PhysNorm dynamically modulates 0.05° high-resolution feature maps using 0.25° low-resolution physical features, thereby ensuring rigorous physical continuity while filling data gaps. Validation results show that the model performs exceptionally on the test set ( R ² = 0.889) and maintains high generalization stability in an independent validation on 2024 data, which was not used for training. Cross-validation against an independent polar-orbiting satellite instrument (GOME-2C) confirms the reliability of the dataset in reconstructing pollution hotspots and its applicability to periods without GEMS observations (2019–2022). Comparison with independent ground-based Pandora observations showed an overall correlation of R =0.644 for 2019–2024. Using this reconstructed dataset, this study finely delineates the periodic diurnal characteristics of NO ₂ in typical Asian regions and characterizes inter-annual variations in NO ₂ concentration patterns over the past six years. This newly generated dataset provides a spatiotemporally continuous daytime record, overcoming the limitations of cloud cover and short observational histories, thereby facilitating long-term, high-resolution air quality assessments and epidemiological studies. This dataset is available at https://doi.org/10.5281/zenodo.20427767 (Gao et al., 2026).