Seamless Global Mapping of HCHO from Chinese Satellite via Spectral Retrieval and Neural Operator

Accurate global mapping of atmospheric formaldehyde (HCHO), a key proxy for reactive volatile organic compounds, remains limited by cloud cover, orbital gaps, and the sensitivity limitations of current satellite instruments. Observations from China’s GF-5B satellite provide valuable morning-time coverage, but suffer from low spectral quality and signal-to-noise ratio, yielding noisy and spatially fragmented data. Here, we introduce a framework that combines physical spectral retrieval with neural operator learning to reconstruct daily global HCHO vertical column densities directly from these sparse and artifact-contaminated observations. The model operates on a spherical domain and incorporates meteorological flow fields and geographic priors to capture transport dynamics, spatial gradients, and seasonal variability. Notably, it learns to suppress striping noise inherent in the sensor, enabling high-resolution, globally continuous reconstructions from discontinuous retrievals. Unlike conventional approaches, our method avoids synthetic labels and chemical transport models, making the reconstructions observation-driven. Validation against ground-based measurements (multiaxis differential optical absorption spectroscopy and Pandora) shows a marked improvement in correlation (from 0.56 to 0.75), surpassing TROPOspheric Monitoring Instrument in several regions. The model captures dynamic enhancements of HCHO associated with wildfires, biogenic emissions, and industrial plumes, highlighting its sensitivity to short-term atmospheric perturbations. Our results establish a scalable, observation-driven approach for extracting chemically meaningful information from noisy satellite records, offering a pathway toward autonomous, high-frequency trace gas monitoring from multisensor Earth observation systems.