Global daily joint reconstruction of TROPOMI tropospheric No 2 columns and averaging kernels with temporally aware deep learning

Achieving spatiotemporal consistency in comparisons between satellite tropospheric nitrogen dioxide (NO2) retrievals and numerical models or profile-resolving observations depends on the availability of both seamless NO2 vertical columns and their corresponding averaging kernels (AKs). However, cloud cover, aerosols, and solar-zenith-angle constraints frequently invalidate TROPOMI retrievals. Because existing reconstruction studies typically provide NO2 columns in isolation, AK-weighted analyses remain restricted to the original, fragmented observation footprints. Here we develop a temporally aware deep learning framework that jointly reconstructs TROPOMI tropospheric NO2 columns and 34-layer AKs over global land at 0.05° daily resolution from May 2018 to December 2024, with the resulting gap-filled field representing a clear-view, retrieval-consistent column and providing a lower-bound representation under persistent cloud. Because static-feature ML baselines do not explicitly represent photochemical-lifetime, boundary-layer, and transport processes, we introduce a CNN-Temporal-Attention architecture that explicitly encodes 24-hour pre-overpass meteorological dependencies. For AK, a retrieval-specific operator that lacks Vertical Column Density (VCD)-equivalent spatial priors, we exploit inter-layer continuity in vertical sensitivity through a layer-preserving 1D-CNN. Both models achieve ΔR2 ≤ 0.02 across random, temporal, and spatial-block cross-validation, with the AK model reaching a pooled R2 of 0.96; attribution patterns are consistent with the seasonal variation of the NOx photochemical lifetime, and the learned temporal weighting is consistent with cross-day boundary-layer memory, a bimodal boundary-layer cycling, and a ~2-hour upwind transport-related temporal association. Independent validation against 141 Pandora and 29 MAX-DOAS stations supports daily-to-annual accuracy, while AK-smoothed MAX-DOAS and aircraft spiral-profile comparisons support reconstructed AK physical consistency. Reconstructed fields preserve emission-control trends, COVID-19 responses, and biomass-burning signals. Co-delivering VCDs, layer-resolved AKs, and per-pixel uncertainties enables rigorous AK-weighted analysis beyond the original TROPOMI footprint.