Working group 3: Pierre Lepetit – Estimation of visibility and snow height on webcam images with learning to rank approach

L’Atelier interne « SCAI & AI4Climate » réunit les chercheurs, ingénieurs, doctorants, post doctorants concernés par les thématiques liées à conception et l’utilisation de nouvelles méthodes d’Intelligence Artificielle pour l’étude de l’environnement, allant du modèle à l’observation. Les premières réunions seront consacrées aux travaux des doctorants. L’exposé sera suivi d’une discussion avec les participants sur l’approche et les perspectives possibles du travail. 

16 Mars à 10h
sur le campus de Jussieu,
Salle de réunion SCAI
Batiment Esclangon 1er étage

Participer à la réunion Zoom
(voir information de connexion ci-dessous)

  • The image-based estimation of meteorological parameters provides clear benefits for surface weather observation. When a local event arises, as a dense fog or a snow settling, webcams and CCTV cameras are sources of valuable information. These images actually inform about the class of weather (sunny, rainy, foggy, snowy, etc). They also enable to gauge quantitative parameters as the horizontal visibility (the farest you can see), the snow height, the precipitation rate, etc, with a variable precision.
  • Recently, the weather classification task has been successfully addressed by deep learning approaches. However, the quantitative estimation faces a strong difficulty: the existing data sets that contain both images and precise weather measurements are rare and involve only few different outdoor scenes. It is virtually impossible for an expert to assign image-wise quantitative labels, but it is possible to compare two images from the same webcam and therefore assign pairwise labels. An “uncomparable” label being assigned to couples for which the expert is not able to distinguish the two images with respect to the parameter.
  • This analysis gives the starting point of the workshop. The discussion will deal with the methods of labeling, learning to rank and calibration that may help to yield such comparisons and to predict ordinal or quantitative estimations of visibility and snow height. The way uncomparable pairs could lead to predict an image-wise uncertainty will also be addressed.

Participer à la réunion Zoom

ID de réunion : 982 7831 9724

Trouvez votre numéro local :

Machine learning and natural hazards – Sophie Giffard-Roisin

Link for the slides

The seminar is on February 10th 14:00 and will be held remotely.

Link to the zoom session:

Sophie Giffard-Roisin presentation is entitled:

« Machine learning and natural hazards »

The goal of this talk is to show how we can use the strength of artificial intelligence to help making diagnosis and finding concrete and local solutions to natural hazards. Tropical cyclones, avalanches, earthquakes or landslides affects often vulnerable areas and populations, where the understanding of the phenomena and better risk assessment and predictions can make a substantial impact. The data available to monitor these natural phenomena has considerably increased in the recent years. For example, SAR (synthetic aperture radar) imaging data, provided by the Sentinel 1 satellites, is now freely available up to every 6 days in a majority of regions, even remote areas. Yet, artificial intelligence (AI) and machine learning (ML) have only scarcely been used in these domains. But these techniques have already showed their impact in many scientific fields having similar data structures (large volume of data, presence of noise, complex physical phenomena) such as medical imaging (detection/segmentation of pathologies), crop yield (prediction), security (recognition). We will see in this talk, with concrete examples, how to design machine learning models for specific tasks with real imaging or temporal data inputs. Concretely, starting mainly from convolutional neural networks, what are the key aspects to consider and what are pitfalls to avoid?

Short bio:
Sophie Giffard-Roisin is a researcher hired by IRD (French National Institute for Sustainable Development) and based at ISTerre, Grenoble (UGA, France). Her work focuses on machine learning applications for natural hazards, especially using remote sensing and time series data. She did her PhD at Inria, Nice (France) under the supervision of Nicholas Ayache on machine learning and modelling for medical image analysis. Then she did a post-doc in CU Boulder, Colorado (USA) in Claire Monteleoni’s team where she worked on climate and meteorological applications of machine learning. She moved to ISTerre, the Earth Science Laboratory of Grenoble Université (UGA, France), for a permanent position in 2019 where she now focuses on machine learning for natural hazards in geosciences.

2nd Working Group: Learning dynamics from partial and noisy observation with the help of Data Assimilation – Arthur Filoche

11 December, 10 o’clock, Salle de réunion SCAI, Jussieu.
Batiment Esclangon 1er étage

Participer à la réunion Zoom

Geosciences have long-standing experience in modeling, forecasting, or estimating complex dynamical systems like the atmosphere or the ocean. Most of these models came from physical laws and are described by PDE. Usually, sparse and noisy observations of such systems are available. The first need to produce a forecast is to estimate initial conditions. This is usually done via Data Assimilation (DA), a set of methods that optimally combines a dynamical model and observations, focusing on system state estimation. In variational formalism, it’s a PDE-constrained optimization problem that requires adjoint modeling to calculate gradients. This field is very close to Machine Learning (ML) in the sense that both learn from data.

ML algorithms have demonstrated impressive results of spatiotemporal forecasting, but to do so it needs dense data which is rarely the case in earth sciences. Also, tools provided by the deep learning community based on automatic differentiation are particularly suitable for variational DA, avoiding explicit adjoint modeling.

What motivates this discussion is that physics-based model is often
incomplete and machine learning can provide a learnable class of model
while data assimilation can provide dense data.

The a similar talk can be found here, and an early conference paper can be found here.

Power-efficient deep learning algorithms – Sébastien Loustau

Link for the slides

Next seminar is on October 14th October (14:30) in « Campus Pierre & Marie Curie » of Sorbonne University. It will take place in SCAI seminar room, building « Esclangon », 1st floor

Si vous souhaitez assister en personne à ce séminaire:

Sébastien présentera ses travaux à la salle de séminaire de SCAI (plan d’accès:
Merci de vous inscrire sur ce lien :
Nous vous conseillons néanmoins d’apporter avec vous votre ordinateur portable afin d’être connecté en même temps sur la salle zoom (voir ci-dessous)

Si vous souhaitez assister à distance: 

Voici le lien zoom:
Vous pourrez également poser des questions sur le chat qui seront retransmises dans la salle.

Sebastien Loustau presentation is entitled:

« Power-efficient deep learning algorithms»

In this talk, I will present both theoretical and practical aspect of how designing power-efficient deep learning algorithms. After a non-exhaustive survey of different contributions about the machine learning perspective (training low bit-width networks), the hardware counterpart (CNNs accelerators) and the relationship with Auto-ML and the NAS procedure, I will present a theoretically based approach to add the power efficiency constraint into the optimization procedure of training deep nets. This work in progress bridges optimal transport and information theory with online learning.

Short bio:
Sébastien is a researcher in mathematical statistics and Machine Learning. He has studied the theoretical aspect of both statistical and online learning. His research interests include online learning, unsupervised learning, adaptive algorithms and minimax theory. He also founded LumenAI 5 years ago.

Journal club meetings

Next meetings

[18-02-2020] Taking climate model evaluation to the next level (Eyring et al. 2019)
Discussion led by: Pierre Le Bras

Past meetings


  • [17-09-2020]  Process-based climate model development harnessing machine learning: I. a calibration tool for parameterization improvement (Couvreux et al. 2020)
    Discussion led by: Redouane Lguensat
  • [16-07-2020] A deep learning approach to detecting volcano deformation from satellite imagery using synthetic datasets (Anantrasirichaia et al. 2019)
    Discussion led by: Sophie Giffard-Roisin
  • [18-06-2020] Up to two billion times acceleration of scientific simulations with deep neural architecture search (Kasim et al. 2020) Papers:
    Discussion led by: Julie Deshayes
  • [16-04-2020] Universal Differential Equations for Scientific Machine Learning (Rackauckas et al. 2020)
    Discussion led by: Redouane Lguensat
  • [19-03-2020] WeatherBench: A benchmark dataset for data-driven weather forecasting by S. Rasp, P.D. Dueben, S. Scher, J.A. Weyn, S. Mouatadid and N. Thuerey (2020)
    Discussion led by: Julien Le Sommer
  • [27-02-2020] Deep learning to infer eddy heat fluxes from sea surface height patterns of mesoscale turbulence (George et al. 2020)
    Discussion led by: Julie Deshayes

Tenure track in Statistical learning at Ecole Polytechnique

Ecole polytechnique  is opening a tenure track position on statistical learning and artificial intelligence for energy/climate. The description is attached. See also

This is a position between the Applied Math departement  and the Meteorology  departement.We are primarily interested in applicants whose research in statistical learning and artificial intelligence shall contribute to address societal challenges in energy, sustainability and climate change (e.g. statistical learning for energy efficiency, for load curve prediction, for pricing mechanisms, for smart grid control, for load curve disaggregation, etc.). The Ecole Polytechnique offers an exceptional environment, with an adapted teaching load, as well as the support (scientific, administrative and budgetary). 
To  apply, follow the link
The position  number is 71.

Inferring causation from time series with perspectives in Earth system sciences – Jakob Runge

Link to the slides

The seminar is on December 4th at 10:00 14:00 and will be held remotely.

Link to the zoom session:

Jakob Runge presentation is entitled:

« Inferring causation from time series with perspectives in Earth system sciences »


The heart of the scientific enterprise is a rational effort to understand the causes behind the phenomena we observe. In disciplines dealing with complex dynamical systems, such as the Earth system, replicated real experiments are rarely feasible. However, a rapidly increasing amount of observational and simulated data opens up the use of novel data-driven causal inference methods beyond the commonly adopted correlation techniques. In this talk, I will present a recent Perspective Paper in Nature Communications giving an overview of causal inference methods and identify key tasks and major challenges where causal methods have the potential to advance the state-of-the-art in Earth system sciences. Several methods will be illustrated by `success’ examples where causal inference methods have already led to novel insights and I will close with an outlook of this relatively new and exciting field. I will also present the causal inference benchmark platform that aims to assess the performance of causal inference methods and to help practitioners choose the right method for a particular problem.

Runge, J., S. Bathiany, E. Bollt, G. Camps-Valls, D. Coumou, E. Deyle, C. Glymour, M. Kretschmer, M. D. Mahecha, J. Muñoz-Marı́, E. H. van Nes, J. Peters, R. Quax, M. Reichstein, M. Scheffer, B. Schölkopf, P. Spirtes, G. Sugihara, J. Sun, K. Zhang, and J. Zscheischler (2019). Inferring causation from time series in earth system sciences. Nature Communications 10 (1), 2553.

Short bio:

Jakob Runge heads the Climate Informatics working group at the German Aerospace Center’s Institute of Data Science since 2017. The group combines innovative data science methods from different fields (graphical models, causal inference, nonlinear dynamics, deep learning) and closely works with experts in the climate sciences. Jakob studied physics at Humboldt University Berlin and obtained his PhD at the Potsdam Institute for Climate Impact Research in 2014. For his studies, he was funded by the German National Foundation (Studienstiftung) and his thesis was awarded the Carl-Ramsauer prize by the Berlin Physical Society. In 2014 he won a $200.000 Fellowship Award in Studying Complex Systems by the James S. McDonnell Foundation and joined the Grantham Institute, Imperial College, from 2016 to 2017. On he provides Tigramite, a time series analysis python module for causal inference. For more details, see:

A direct approach to detection and attribution of climate change – Eniko Szekely – 24/01/2020

Lien pour les slides

Le prochain séminaire aura lieu le 24 Janvier à 14h30 au campus Pierre et Marie Curie de Sorbonne Université dans la salle 105 du LIP6 couloir 25-26 au 1er étage.

La présentation de Eniko Szekely est intitulée:

« A direct approach to detection and attribution of climate change »


In this talk I will present a novel statistical learning approach for detection and attribution (D&A) of climate change. Traditional optimal D&A studies try to directly model the observations from model simulations, but practically this is challenging due to high-dimensionality. Here, we propose a supervised approach where we predict a given metric or external forcing directly from the high-dimensional spatial pattern of climate variables, and use the predicted metric as a test statistic for D&A. The first part of the talk will focus on daily detection and show that we can now detect climate change from global weather for any single day since spring 2012. The second part of the talk will focus on attribution of climate change. For attribution, we want the prediction of the external forcing, e.g., anthropogenic forcing, to work well even under changes in the distribution of other external forcings, e.g., solar or volcanic forcings. Therefore we formulate the optimization problem from a distributional robustness perspective, and use anchor regression to ensure good predictions even under such distributional changes.

Notice Biographie:

Eniko is a senior data scientist at the Swiss Data Science Center, EPFL & ETH Zurich, working on machine learning for climate science. Previously, she was a postdoctoral researcher at the Courant Institute of Mathematical Sciences, New York University, and she obtained her PhD in Computer Science from the University of Geneva, Switzerland. Broadly she is interested in machine learning for high-dimensional data and nonlinear phenomena arising from dynamical systems. More recently she has been working on using machine learning and statistical learning approaches for climate science, and has been involved in the organization of the Climate Informatics workshop since 2015.