Analysis of observations most often requires physical models. The determination of the model physical parameters is done by comparing the model outputs to observations through inversions. When one tries to account for realistic physical processes, costly models are often required. The inversion of observations is difficult as it requires hundreds of simulations to estimate the parameters, and thousands to millions of simulations to estimate the uncertainties.
Surrogates or metamodels are statistical models based on the results of costly experiences (complex physical models, inversions, laboratory experiments). They are to date the main approach to by-pass the costly and lengthy numerical computations required to obtain simulation results.
During this half day course, we will introduce an important class of surrogate models called Gaussian stochastic Processes (PG, also know as Kriging). PG not only approach the observations, but they also provide a description of the associated uncertainties.
The course will start by introducing Gaussian Processes. Then, we will explain how they can be used for inversions and uncertainties propagation. The class will be illustrated by practical examples in R.
Course : The course is available on the following link https://hal.archives-ouvertes.fr/cel-01618068
Speakers :
Nicolas Durrande, Associate professor at the School of Mines in St Etienne and CNRS LIMOS, specialized in Gaussian processes.
Rodolphe Le Riche, CNRS researcher, School of Mines in St Etienne and CNRS LIMOS, specialized in the optimization of Gaussian processes.
Very high resolution optical satellites, because of their stereographic capacity, have facilitated the acquisition of multi-looking images of the earth. Surface digital elevation models and clouds of points computed from these images can provide a valuable input to geomorphological, tectonic, volcanological, hydrological and glacier studies. The more and more frequent access to time series allows the computation of horizontal displacement fields using image cross-correlation techniques. However, theses processes require an expertise, dedicated computer resources and/or costly softwares. In order to allow the Earth science community to easily and quickly process very high resolution multi-looking optical image in order to quantify ground deformation, fully automated processing chains have been developed. Implemented tools are based on photogrametic open-source programs and are designed to allow for the distributed computation on high performances infrastructures (clouds) or personal computers.
The objective of this class is to present the processing chains through practical examples; case studies on a variety of geological objects (earthquake, volcano, landslide, glacier) will be presented.
Speakers :
Jean-Philippe Malet, CNRS researcher at EOST – Strasbourg University, specialized in geomorphology, space geodesy and landslides
David Michéa, Research Engineer at ICube Laboratory at EOST - Strasbourg University, specialized in high performance computation, space geodesy and landslides
André Stumpf, post-doctoral researcher at EOST – Strasbourg University, specialized in geomorphology, space geodesy and landslides
Program :
(1) Introduction to the specificities of radar interferometry with Sentinel 1 data (TOPS Wide Swath mode), one hour course.
(2) Practical course on Sentinel InSAR processing using tools developed in the framework of the data group in solid Earth (Form@ter): use of the web service for the on-demand computation of a Sentinel 1 interferogram (ETALAB project); computation of a short sentinel 1 time series on a cluster using NSBAS (2 times 2 hours courses).
Course : The course is available on the following link Doinetal_TP_InSAR_MDIS2017.pdf
Speakers :
Marie Pierre Doin et Cécile Lasserre, CNRS Researcher at ISTerre, University of Grenoble.
Raphaël Grandin, Associate professor at IPGP, Jussieu University, Paris.
Erwan Pathier, Associate professor at ISTerre.
Online user: 2 | RSS Feed |