Laboratoire Interuniversitaire des Systèmes Atmosphériques

Funding and Collaboration:

This PhD thesis is eligible for co-funding by CNES (French Space Agency), subject to the candidate’s selection. The successful candidate will secure a 3-year doctoral contract through a co-funding mechanism involving CNES and a partner (e.g., doctoral school, ANR project, or other funding sources). The research will be conducted in a collaborative environment between two leading laboratories:

• LISA (Laboratoire Interuniversitaire des Systèmes Atmosphériques), located in Créteil (Paris region), specializing in atmospheric science and remote sensing.
• LIPADE (Laboratoire d’Informatique Paris Descartes), located in Paris, with expertise in machine learning and data science.

The candidate will benefit from a multidisciplinary research environment, with access to cutting-edge satellite datasets (e.g., Sentinel-5P, IASI) and high-performance computing resources.

Context

20th-century industrialization and urbanization have drastically increased anthropogenic pollutant emissions, positioning air pollution as the foremost environmental health risk according to the World Health Organization, with approximately 3 million annual deaths attributed to outdoor air pollution alone. Addressing this air quality issue has prompted global efforts to implement emission regulations. However, accurately quantifying anthropogenic pollutant emissions at high spatial and temporal resolution over extended periods is essential for understanding their impact on atmospheric chemistry and climate. Such data is also vital for evaluating the effectiveness of emission regulations, refining mitigation strategies, and enhancing short-term air quality forecasting.

Current emission inventories, primarily based on self-reported and nationally aggregated data, are plagued by significant uncertainties. Additionally, the time required to collect and update this data limits their relevance in air quality modeling and forecasting. Recent advancements in satellite remote sensing, particularly with instruments like Sentinel-5 Precursor (S5P) and the newly launched Sentinel-4 and Sentinel-5, offer a new interesting opportunity to improve emission estimates. These satellites provide unprecedented spatial and temporal coverage, enabling the correction of emission inventories on a near-daily to hourly basis when combined with data assimilation techniques. However, managing and exploiting the vast amounts of high-resolution data they produce remains a significant challenge.

Objectives

The primary objective of this PhD project is to explore deep learning approaches to develop a system capable of estimating pollutant emissions with high spatial and temporal resolution, while also enabling near-real-time processing. The PhD student will focus on creating a multi-task deep learning framework designed to infer total emissions at pixel-level resolution for key pollutants and to distinguish the contributions of various emission sectors, such as traffic, residential, and industrial sources.

The project will adapt spatial regression architectures like UNet or UperNet and explore pre-trained Earth Observation models such as DOFA or TerraMind. A key challenge and innovation of this PhD will be to develop a system that can separate emissions from different sources using space-based observations of atmospheric concentrations. Training data will come from the CHIMERE chemistry-transport model, using simulations with perturbed sectoral emissions.

The focus will be on nitrogen oxides (NOx), which are key pollutants for air quality as precursors to ozone and particulate matter, and emitted by various sectors. The study will utilize high-resolution NO₂ observations from Sentinel-5P since 2018, supplemented by other pollutants (SO₂, HCHO, CO) and auxiliary data such as land cover types. This information will help improve the performance of the deep learning model. The system will be applied across Europe, focusing on periods of significant emission disruptions, such as the COVID-19 pandemic and the war in Ukraine. These events provide unique opportunities to test and validate the approach due to their strong, detectable signals from space.

Finally, another exciting challenge of the PhD will be to test the system using data from the geostationary Sentinel-4 satellite, aiming to achieve much higher temporal resolution in emission monitoring.

Candidate Profile:

• Master's Degree
• Background: Interdisciplinary (ML + atmospheric science) or strong skills in one field with motivation to learn the other.
• Technical: Proficiency in Python/Linux; experience with large datasets is a plus.
• Mindset: Interest in interdisciplinary research and real-world impact.

Interested candidates should submit the following documents to gaelle.dufour@lisa.ipsl.fr and sylvain.lobry@u-paris.fr :

1. A detailed curriculum vitae (CV).
2. A cover letter outlining their motivation, research interests, and relevant experience.
3. Academic transcripts (Master’s degree).
4. Contact details for two academic or professional references.