Machine learning segmentation of fault scarps from optical satellite images

5 to 6 months, starting date between Feb and May 2024

Background : Most of geological natural hazards such as earthquakes remain, still up to now, unpredictable. Nonetheless, the processes governing them are today well apprehended thanks to advances in geophysics and in particular rupture physics. However, sustainable planning of territory development, hazard prevention and management still require precise mappings as inputs to physical and statistical models. While the subduction megathrusts are well identified, the majority of small to medium crustal faults are not precisely mapped or even known (see Parina earthquake in 2016 in Peru). In this way, one of the geohazard challenge, in particular in South America, consists in documenting these surface information over large territories. We can make profit of the rich and various geoscientific acquisitions, among which remote sensing, in order to map the passed surface ruptures. Nonetheless, the current mapping is limited by the amount of information that can be extracted, most of the time still manually and thus expert-dependent, at a large scale both in space (hundreds of squared kilometers) and time (from thousands of years to months). The localization and complete inventory of larger capable faults over a full landscape has never been automatized ; mainly because : 1) fault markers are not easily visible, and are hard to disentangle from vegetation, valleys, roads, or rivers. 2) few ground truth data exist 3) the optical information alone might not be always sufficient.

Objectives of the internship : The objective of this internship is to identify active fault scarps that highlight the seismic activity and represent scars of past seismic surface ruptures. We aim at identifying active fault scarps highlighting the seismic activity, at work in particular in the Andean highlands highlands and Atacama region, and representing scars of past surface ruptures. Indeed, Holocene faults have left geomorphological makers in the landscape. We will develop a deep learning model able to segment active faults from a landscape, using remote sensing imaging. In particular, the model will learn from optical satellite images (Pléiades) to identify fault scarps. The first pilot study area will be the central Andes of Southern Peru, where the landscape is free from vegetation and a first exhaustive manual inventory was already performed. For this, we will develop a deep learning U-net type model, and we will study the possibility to derive self-supervised techniques to cope the small number of labeled data.

Pre-requisites : A computer science and/or remote sensing processing degree is expected. A strong interest towards the applicative aspects of methodological developments in AI and a curiosity towards important processes in Earth sciences will be appreciated. The work environment is interdisciplinary, so communication and facilitation skills are required. As the project involves collaborating with Peruvian partners, a correct English (or Spanish) is needed.

Environnement : The intern will be based in ISTerre, the important Earth science laboratory of Grenoble, . In ISTerre, he/she will be supervised by Sophie Giffard-Roisin (expert in AI using remote sensing for natrual hazards) and Laurence Audin (tectonic geomorphologist specialized in active deformation). The intern will also collaborate with Lorena Rosell-Guevara, research engineer at Ingemmet, Lima, Peru.

Contact : send an email as soon as possible to sophie.giffard univ-grenoble-alpes.fr, with your CV and a brief explanation on why you want to apply and your interest in the topic.