Accueil > Actualités > Séminaires > Séminaire de Grégoire Mariethoz au METIS


Titre : Data expansion: using analogues to improve remote sensing and climate datasets
Nom du conférencier : Grégoire Mariethoz
Son affiliation : IAMG lecturer 2018 (International Association for Mathematical Geosciences), éditeur en chef de Computer and Geosciences, Professeur à l'Université de Lausanne, Suisse, Institute of Earth Surface Dynamics (IDYST), responsable de l'équipe Geostatistical Algorithms & Image Analysis (GAIA lab).
Laboratoire organisateur : METIS
Date et heure : 23-02-2018 13h00
Lieu : Campus de Jussieu, salle Darcy (tour 46-56 3e étage)
Résumé :

Statistics are often used to perform data reduction: that is, summarizing a large amount of  information into a model containing a limited number of parameters. In geoscience, there is often a  need to do the opposite as data reduction: based on limited information about a studied process,  one wants to stochastically generate more data, which ideally should be statistically indistinguishable from the truth. Numerous applications include interpolation, downscaling or gap-filling of spatio-temporal datasets. This process of generating new data can be called “data expansion”, a termed first coined by André Journel in 1989 in the context of relatively simple geostatistical models.
The last years have seen ever-increasing sensing capabilities and improved numerical models that are able to reproduce complex physics. However, it appears that all such data have intrinsic limitations: any data acquisition procedure, no matter how sophisticated, is limited by sensor constraints (e.g., coverage, resolution, frequency), and numerical models are challenged for predicting the state of the environment under changing climate conditions. Addressing these limitations calls for data expansion. Since the time when André Journel came up with the notion of data expansion, improved models have been developed that are tailored to extract information from the new types of data that are available today.
This talk will provide a survey of such models and algorithms, with an emphasis on those that extract training information from analogues. While these methods have often been used for the quantification of uncertainty in 3D subsurface models, there is a wide array of domains where it can be useful for 1D and 2D data, such as Earth observation images or climate time series. Some recent geoscience applications will be shown, particularly focusing on satellite-based earth observation and hydrology, for weather generation and for the completion of missing data.