Aspot image

Learning in the extreme


Get to know Eulalie Boucher, winner of a 2022 EUMETSAT Early Career Scientist Award

Aspot image
Aspot image

Eulalie Boucher, a winner of the 2022 EUMETSAT Early Career Scientist award, develops new machine learning methods to maximise the potential of data provided by EUMETSAT’s Metop polar-orbiting satellites.

Last Updated

01 November 2023

Published on

15 September 2023

To make sense of the vast quantities of remote sensing data provided by meteorological satellite missions, specialists increasingly turn to statistical and machine learning methods.

However, the need for accuracy and consistency must be balanced against stark differences in the resolution and formats of different datasets. These trade-offs can impact the prediction of extreme events, such as floods, heatwaves, and droughts.

"Artificial intelligence and machine learning techniques play an increasingly important role in the retrieval and processing of Earth observations," says Eulalie Boucher, a PhD student at the Université Paris Sciences et Lettres, France.

“However, compromises must be made to combine data efficiently and effectively from many sources and locations across the Earth’s land, oceans, atmosphere, and cryosphere.

“In numerical weather prediction, for instance, some machine learning methods can favour the prediction of 'average' conditions over 'extreme', leading to the under-prediction or occasional over-anticipation of extreme events.

“This presents challenges because the integration of machine learning into forecasting must meet very high criteria in terms of stability and accuracy.”

Eulalie Boucher
Eulalie Boucher studied maths and computer science before an internship at the Paris Observatory convinced her of the value of applying her multidisciplinary skills to meteorology. She is now a PhD student at the Université Paris Sciences et Lettres, France. As a winner of the 2022 EUMETSAT Early Career Scientist award, Boucher is a guest of honour at this year’s EUMETSAT Meteorological Satellite Conference in Malmö, Sweden. “I am very much looking forward to conference, where I have the opportunity to present our research to the world, learn from leading experts, and to visit Sweden for the first time," she added.
Photo courtesy of Eulalie Boucher

Deep learning approaches

Boucher and colleagues apply new methods, using multiple layers of machine learning to localise observations and attempt to reduce errors.

"Deep learning applies current machine learning approaches such as neural networks at multiple layers, which allows the network to learn complex features and representations of the data,” Boucher explains.

“This is having a very positive impact on a wide range of research fields. To maximise its potential in meteorology, for example, it’s important to understand and account for potential errors caused by compromises.”

One area Boucher’s team is working on is data retrieval and the processing of temperature measurements provided by the Infrared Atmospheric Sounding Interferometer (IASI) instrument aboard EUMETSAT’s Metop polar-orbiting satellites.

“By applying deep learning methods to IASI datasets, we have been able to improve the accuracy of extreme temperature predictions, which I hope can soon be expanded to datasets provided by other instruments.

“Eventually this work could benefit applications such as numerical weather prediction models and climate studies – it gives me great satisfaction to be working on projects that could have such direct benefits for societies around the world.”


Adam Gristwood