Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : ISTerre Grenoble / Lamont Columbia University
Durée : 24
Contact : sophie.giffard@univ-grenoble-alpes.fr
Date limite de publication : 2025-01-31

Contexte :
This post-doc is part of the ANR EDAM: Earth Deformation from Automatic Mapping. Most of geological natural hazards such as earthquakes and landslides remain, up to now, unpredictable although the processes governing them are today well apprehended thanks to advances in geophysics. In South America, while the volcanoes and the subduction megathrust are well identified, the majority of crustal faults (i.e. associated to continental crust deformation and not to subduction megathrusts) and landslides are not precisely mapped or not even known. The inventory task is largely incomplete as it is currently done manually in the field or by visual analysis of remote sensing data by geological surveys. It is an impossible task regarding the km-scale ongoing deformations all over South America. In particular, there is a need to locate the active slow ground movements. Thanks to Synthetic Aperture Radar (SAR) imaging by several SAR missions (such as the Sentinel-1 constellation), it is now possible to extract ground deformations at the centimeter accuracy since 2015. The recently processed Interferometric SAR (InSAR) time series (from the
complex pipeline by Thollard et al. 2021) covering in particular the Andes for the period 2014-21 provide unprecedented coverage at the continental scale. Yet, the automation of deformation detection, in particular at a large scale, has only been scarcely tackled.

Recent machine learning techniques have revolutionized the fields of natural language processing and then image processing by constructing foundation models able to be trained using unlabelled data itself
(usually a large set) to generate supervisory signals. The learned extractor generates relevant embeddings, which can then be used for various downstream tasks having only a small number of labelled samples. This technique has been very recently successfully applied for remote sensing data: the same geospatial foundation model can be the backbone of different applications, such as crop classification and flood
mapping (Jakubik et al. 2023). Yet, they do not give satisfactory results for earth deformation applications, because: 1) the acquisitions (i.e. image bands) to rely on are largely different from most of other earth
observation tasks, usually made to analyze land cover (using only optical and hyperspectral bands) while we are interested in structure and shape (where topography is key) and displacement (from InSAR); 2)
the extracted relevant features, i.e. embeddings, are largely unrelated to earth deformation. Indeed, earth deformation is usually a low signal mostly independent to land surface appearance.

Sujet :
The goal of this post-doc is to develop a foundation model able to capture active deformation features from InSAR. In particular, we want to combine both optical-related (RGB, topography) and radar-related (wrapped interferograms and temporal coherence map). This self-supervised model will then be used directly for two downstream tasks, the identification of faults (earthquake ruptures and creeping faults) and the detection of landslide movements (slow-moving landslides and landslides ruptures). We already have a recent database of slow-moving landslides visible in InSAR over South Peru, and the InSAR as well as DEM over the region has already been processed at ISTerre. Finally, the post-doc will interact (with a visit to Peru) with Peruvian partners in order to interpret the results and transfer the knowledge and the codes.

Profil du candidat :
We look for a motivated post-doctoral fellow, who did his/her PhD in either deep learning image processing, remote sensing, or geoscience. The post-doc will be located at ISTerre, Grenoble, within a team of many interdisciplinary students and post-docs. On top of the stimulating remote sensing environment within ISTerre, he/she will work closely with S. Giffard-Roisin, who will be on leave at Columbia University for the beginning of the contract. A few-month visit to NY is thus planned, depending on the dates and contract duration.

Formation et compétences requises :
This interdisciplinary work necessitates the knowledge of either A) deep learning and in particular recent image processing methods with cluster computing skills, and/or B) remote sensing and in particular radar interferometry. Geoscience background would be a plus, or at least a curiosity towards geoscience applications. Of course, good communication skills are fundamental, as well as a good English knowledge. Spanish would be a plus but not mandatory.
The start date is between March 2025 and November 2025.
Send me a CV with an email summarizing few lines of context of why you want to apply as soon as possible.

Adresse d’emploi :
ISTerre, Université Grenoble Alpes