Région Sud “Emploi Jeunes Doctorants” 2023-2026

Co-funding by private company ACRI-ST

Summary

This multidisciplinary project will develop new computer vision and deep learning techniques, combined to physics models, to jointly interpret multimodal and multispectral observations of Martian terrain. Mars observation missions have acquired large quantities of images to study its atmosphere and surface. These images are very diverse: they cover various spectral domains, and they have very different spatial and spectral resolutions. They are also complementary, and their joint interpretation is fundamental to characterize the planet. However, this joint analysis is traditionally limited to a simple superposition of observations for visual inspection, without fully exploiting the complementary information they contain. In addition, the number and complexity of observations are increasing dramatically. This calls for the development of new automatic and intelligent analysis tools.

The project aims at 3 main goals:

1. Align and merge images of different types

Merging images requires aligning them finely. Major challenges are their different resolutions and the scarcity of easily identifiable control points in the images. To resolve these problems, we propose to rely on the 3D information of a geometric model. In a previous study, we designed a method for 3D modelling from misaligned point clouds of different resolutions. This method is now being adapted to the alignment of Digital Elevation Models (point cloud-like) that are obtained from Martian imagery. The associated images can then be merged to create an image with very high spatial and spectral resolution.

1. Characterization of land properties

The merged images and the geometric model will contain rich information on the nature and composition of the terrain. The latest deep learning methods will be adapted to identify areas with distinct properties. We will simultaneously consider the geometric properties on a large and small scale, the mineralogical composition whose signatures we will learn to recognize on the different spectra, and the micro-texture of the materials which we will learn to recognize on images of different illuminations.

1. Visualization and interpretation

Planetary scientists need to explore the relationships between different terrains or surface structures identified by their morphology, roughness, composition, etc., to understand geological and climatic processes. These different entities, extracted in the previous steps, will be combined by artificial intelligence algorithms to answer scientific questions of interest for academic (geological faults) and industrial partners (old traces of liquid, faults, and flows of lava on Mars, lunar lava flows, monitoring of the Earth’s coastline).

This interdisciplinary work will bring together experts in data science and deep learning, and in Martian imagery and geology.

Partner: Institut de Planétologie et d’Astrophysique de Grenoble (IPAG)

Funding body: Région Sud, co-funding by private company ACRI-ST

Role in the project: Principal Investigator

Preliminary study: PhD co-supervision: Analysis of Martian terrains’ topography from multispectral orbital images, Kévin Kheng, 2018-2022