The important industrialization and urbanization of the society, started in the 20th century, has led to an unprecedented increase in anthropogenic emissions of pollutants. As a result, air pollution represents the biggest environmental risk to health according to the World Health Organization, with around 3 million deaths per year attributable solely to ambient (outdoor) air pollution. To tackle this air quality (AQ) issue, efforts have been made worldwide to progressively apply emission regulations.

Knowing precisely anthropogenic pollutant emissions at high spatial and temporal resolution over sufficiently long period (several years/decade) is essential to understand and quantify the implication of changing emissions on atmospheric chemistry and climate, both in terms of trends and abrupt changes (e.g. pandemic, economic or geopolitical crises). It is also crucial to evaluate the effectiveness of emission regulations, improve medium and long-term mitigation strategies at country level, and improve short-term forecasting and adaptation capabilities to AQ issues. Existing emission inventories are usually built from officially reported emissions based on self-declarations by emitters, and statistical information aggregated at the national and annual levels. This leads to trust issues, non-negligible uncertainties (from 20-30% to 300% depending on pollutants) and the time needed to collect all the necessary data limits the fast update of the inventories being used in AQ simulation and forecasting systems.

Improvements in satellite remote sensing of pollutants e.g. with the current Sentinel 5 Precursor (S5P), and the newly launched Sentinels 4 and 5, appear to be a game changer to better estimate pollutant emissions. Indeed, the unprecedented spatial and temporal coverage and resolution of these satellite observations, combined with data assimilation techniques, paves the way to provide corrected emissions of the inventories on a daily-to-hourly basis. One of the main challenges is to manage and exploit the large amount of data at high resolution provided by these satellites. The main objective of the PhD project is to explore deep learning approaches to build a system able to better estimate pollutant emissions at high spatial and temporal resolution with a near-real-time processing capability.

The PhD student will be responsible for developing a multi-task deep-learning framework able to infer (i) total emission at the pixel resolution for pollutants of interest and (ii) disentangle the contribution of each emission sector (traffic, residential, industry, etc). To this end, we propose to explore several deep learning-based approaches. As the task can be seen as a spatial regression, we propose to develop an architecture based on segmentation approaches, such as UNet or UperNet. We will also explore the use of pre-trained, EO-oriented foundation models, such as DOFA or ESA’s TerraMind. This exploration will leverage current developments of such a framework to predict total NOx emissions from NO2 satellite images. The challenge and novelty of the PhD is to build a system able to separate information from the different sources using space-based observations of atmospheric concentrations. The relationship between emissions and atmospheric concentrations will be learned from the chemistry-transport model CHIMERE including meteorological information. An ensemble of simulations with perturbed sectoral emissions will be generated and used as training set. 

The primary target in terms of pollutant emissions is the nitrogen oxides (NOx), which are involved in the major air quality issues, being precursors of both O3 and PM and emitted by different sectors. The project will take the advantages of the high performances of the Sentinel 5P atmospheric imager, which provides NO2 observations at few kilometers’ resolution since 2018. Additional information such as other pollutants observations (SO2, HCHO, CO) sharing some of the NO2 sources could be integrated as well as auxiliary information such as land cover type to better constrain the separation of the sources by the deep learning model and improve its performances. The domain of application of the system will be Europe focusing on time periods when large disruptions of the emissions are observed such as during the COVID pandemic or more recently the war in Ukraine.   Indeed, these exceptional conditions create strong detectable signals from space and provide a unique opportunity to test and validate this new approach. 

Finally, one of the other exciting challenges of the PhD will be to test the system on images from the geostationary Sentinel 4 satellite to reach a much higher temporal resolution in the emission monitoring. 

=================

For more Information about the topics and the co-financial partner (found by the lab!); contact Directeur de thèse - gaelle.dufour@lisa.ipsl.fr

Then, prepare a resume, a recent transcript and a reference letter from your M2 supervisor/ engineering school director and you will be ready to apply online  before March 13th, 2026 Midnight Paris time!