• Citer

Advancing Machine Learning for Estimating Atmospheric NO2 Fields at High Resolution: through Uncertainty Quantification, Physics-Guided Learning, and Pretraining with Fine-Tuning

par Sun, Wenfu
Président du jury Coheur, Pierre
Promoteur Clarisse, Lieven
Co-Promoteur Tack, Frederik
Publication Non publié, 2026-09-10

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Estimating atmospheric nitrogen dioxide (NO2) at high spatial and temporal resolution is important for characterizing air pollution patterns, assessing exposure and ecosystem impacts, and supporting air-quality management. This task remains challenging because NO2 varies strongly across space and time. The rapid growth of Earth-system observations, reanalysis products, emission-related predictors, and atmospheric simulations provides the basis for high-resolution NO2 estimation. Machine learning (ML) offers a flexible framework for integrating these data sources and addressing non-linear predictor-target relationships. By the early 2020s, the first ML workflows combining satellite observations, meteorological reanalyses, and surface predictors had begun producing daily, kilometer-scale surface NO2 estimates over large regions.ML-based NO2 estimation is now moving toward finer spatial resolution, hourly temporal resolution, and three-dimensional representation. These extensions raise two central challenges: quantifying the uncertainty of the estimate and improving generalization under data-sparse conditions. This thesis addresses these challenges by focusing on three methodological frontiers: uncertainty quantification, physics-guided learning, and pretraining with fine-tuning. These directions are integrated into the development of ML frameworks demonstrated over Western Europe. BEnCQE produces daily 1 km surface NO2 estimates with calibrated prediction intervals that quantify estimation uncertainty. DACNO2 reconstructs daily three-dimensional NO2 fields at 2 km horizontal resolution across eight vertical layers from the surface to 5 km using a physics-guided, multi-constraint pretrain-finetune strategy. PHASE-AQ extends this strategy with advection-informed encoding and an observation-informed correction to reconstruct hourly surface NO2 at 2 km resolution.The uncertainty information from BEnCQE identifies potential exceedances of the WHO 2021 daily NO2 guideline that are not evident from the point estimate alone, and indicates spatial variation in estimate reliability. DACNO2 reproduces the three-dimensional NO2 field with finer spatial detail than the Copernicus Atmosphere Monitoring Service (CAMS) European regional reanalysis and improves agreement with surface station observations relative to that baseline. When used as a priori profiles in TROPOMI retrievals, the DACNO2 product strengthens the spatial gradients between low- and high-NO2 regions, with little change in the average retrieved tropospheric NO2 column. Relative to CAMS, PHASE-AQ improves hourly agreement with independent surface observations, reduces urban underestimation, especially during peak hours, and corrects rural overestimation. Wind diagnostics and ablation experiments show plausible transport behavior beyond monitored locations. The reconstructed fields resolve fine-scale spatiotemporal variability, particularly in urban and suburban areas, and show how these fine-scale patterns change across hours, seasons, and cities.The methods developed in this thesis can be extended to other atmospheric species and adapted for forecasting and inverse emission estimation. They also provide a basis for transfer learning in data-poor regions with limited surface monitoring. Together with broader progress in Earth-system ML, these methods support the development of foundation models for atmospheric composition.