EGU25-12599, updated on 15 Mar 2025
https://doi.org/10.5194/egusphere-egu25-12599
EGU General Assembly 2025
© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.
A Deep Learning Method for Model-Measurement Fusion of Atmospheric Concentrations with Physical Constraints
Jia Xing1, Bok H Baek2, Siwei Li3, Chi-Tsan Wang2, Ge Song3, Siqi Ma2, Daniel Tong2, and Joshua Fu1
Jia Xing et al.
  • 1Department of Civil and Environmental Engineering, the University of Tennessee, Knoxville, TN 37996, USA
  • 2Center for Spatial Information Science and Systems, George Mason University, Fairfax, VA 22030, USA
  • 3School of Remote Sensing and Information Engineering, Wuhan University, Hubei, 430000, China

Accurate and efficient retrieval of atmospheric chemical concentrations across space and time is crucial for weather prediction and health assessments. However, existing model-measurement fusion methods suffer from limitations due to imbalanced samples from ground measurements or less effective assimilation of satellite data along with numerical modeling. To address these limitations, this study introduces a novel Deep-learning Measurement-Model Fusion method (DeepMMF) constrained by physical and chemical laws inferred from numerical chemical transport models (CTM). This method is applied to NO₂ species over the Continental United States (CONUS) domain for the years 2019 and 2020. By pre-training with abundant CTM simulations, fine-tuning with satellite and ground measurements, and employing a novel optimization strategy for selecting weighting loss and prior emissions, the retrieved spatiotemporally continuous surface NO₂ concentrations present consistent values and daily variations with observations (NMB reduced from -0.3 to -0.1 compared to original CTM simulation). Importantly, the corresponding emissions have been simultaneously adjusted, showing good agreement with changes reported in the national emission inventory (NEI) between 2019 and 2020. Interpretation analysis suggests that the DeepMMF model effectively identifies the importance of satellite data at the regional level and ground measurements at the city level, which is scientifically sound. It exhibits consistent prediction of ground measurements while successfully avoiding the sample imbalance problem that leads to overestimation (up to +100%) of downwind/rural concentrations compared to other existing methods. These results demonstrate the great potential of DeepMMF in data assimilation and retrieval studies for other pollutants and regions, to better support weather forecasting and heatlh studies.

How to cite: Xing, J., Baek, B. H., Li, S., Wang, C.-T., Song, G., Ma, S., Tong, D., and Fu, J.: A Deep Learning Method for Model-Measurement Fusion of Atmospheric Concentrations with Physical Constraints, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12599, https://doi.org/10.5194/egusphere-egu25-12599, 2025.