Objectives and program

Scientific objectives of GeoLearning

The objective of the Geolearning chair is to develop methods in geostatistics, extreme event theory and machine learning for data analysis in the service of climate transition.

This chair brings together two leading research teams that have established a long-standing collaboration, in particular through the supervision of doctoral theses, joint scientific projects, the coordination of a scientific network and teaching.

The Geostatistics team of the Geosciences Center (CG) of Mines Paris
(Geoscience Center (CG), Mines Paris)

The Biostatistics and Spatial Processes research unit Biostatistics and Spatial Processes (BioSP)
du MathNum Department à INRAE

1/ Geostatistical methods for spatio-temporal data

In this first axis, it is a question of developing geostatistical models and methods with a double objective: on the one hand to tend towards a geostatistics informed by physics by developing models inspired by the physics of the phenomenon studied, and on the other hand develop efficient algorithms to process very large spatial data sets. We will rely on the SPDE approach which has already been the subject of joint work by the two teams: the theses of Ricardo Carrizo-Vergara and Lucia Clarotto et the associated publication (Clarotto et al., 2023) et the publication Pereira et al., 2022.

 

2/ Methods for extreme events

Extreme events, and the potential catastrophic impacts they can bring, are a major concern, as has been dramatically illustrated in recent years. Thus, in France and neighboring countries alone, we can cite the heat wave in the summer of 2019, the floods in the summer of 2021 (in Belgium and Germany), the drought and multiple heat waves during the spring and winter. summer in 2022. The objective of the second axis is to propose approaches that combine geostatistics and extreme value theory for the probabilistic prediction, interpolation and simulation of episodes of spatial and spatio-temporal extreme events.

 

3/ Hybridize Machine Learning with geostatistics and extreme event theory

In this third axis, the objective is to bridge the gap between Machine Learning (ML) and geostatistics in order to combine these two approaches to obtain the best of both worlds: ML is very effective in predictive mode when there are nonlinear relationships between many covariates and the variable of interest, while geostatistics is effective in making spatial predictions and providing accurate quantification of uncertainties The development of new approaches in ML for data with spatiotemporal dependencies is a research angle little explored to date.

These three axes are not independent, but closely linked. For example, the analysis of spatio-temporal extreme events (axis 2) will benefit from advances on the methods developed for spatio-temporal data (axis 1), the SPDE approach may be useful for the simulation of spatial extreme events. New machine learning algorithms for space- and time-dependent data (axis 3) will be useful for establishing data-driven relationships between climate model outputs and extreme events (axis 2).

More generally, within the framework of this chair, we wish to decompartmentalise the old categories which tended to separate [tr1] geostatistics, statistics, statistical learning and computer science to found Geolearning.

The program of the chair

A research program proposed by the Chair and discussed with the patron partners during the various committees.

  • 6 to 8 theses, post-doctorates.
  • A high level academic production.
  • Methods for analysis, prediction and simulation disseminated in the form of free software.
  • Transfer of knowledge with the possibility of rapid valorization of research results.
  • A presentation of the results and regular exchanges with the scientific and professional communities.
  • Involvement in the training of engineering students from Mines Paris in connection with the option Geostatistics and Applied Probability with the possibility of educational projects and internships.