5. From Pixels to Predictions: Exploring Machine Learning Techniques for Martian Terrain Characterization

ESA supervisor: Elena Favaro
Collaborator(s): Elliot Sefton-Nash

Site: ESTEC

The surface of Mars exhibits extensive evidence of spatially and temporally diverse exogenic processes. These range from centimetre-scale ripples that have been observed to change on the order of minutes to hours, to vast fluvial valley networks carved billions of years ago under a markedly different atmospheric and climatic regime. These features provide evidence for a dynamic planet, the history of which has dramatically diverged from that of our own.

While rovers and landers have provided invaluable insight into site-specific aeolian, fluvial, and geochemical changes to the landscape, our understand of Mars has largely been derived from high-resolution remotely captured images of the surface. These have been acquired by a constellation of orbital platforms. In particular, High Resolution Imaging Science Experiment (HiRISE) images (up to 25 cm per pixel) are critical in characterizing potential landing sites for future robotic and human-led missions. By examining such images within Geographic Information Systems, we can learn about the processes which shape the Martian landscape. 

A challenge with all this data is that it takes substantial resources, both in terms of human-led investigation and computational power, to characterize and classify diverse terrains. 

In this internship, we propose to build a lightweight, site agnostic terrain characterization algorithm, powered by deep learning, that can be used on HiRISE images across Mars. This algorithm will provide a useful ‘first-pass’ analysis. It will take a HiRISE image of interest and use semantic segmentation to provide an enriched classification product which will aid researchers in quickly and confidently understanding the nature of the terrain, without having to carrying out time consuming manual analysis. The project will seek to test the effectiveness of Deep Learning algorithms using already-available data.

Project duration: 6 months.

Desirable expertise or programming language:

  • Strong command of Python, and experience with TensorFlow, PyTorch, or similar.
  • Experience of semantic segmentation and machine vision is desirable.
  • An understanding of Mars geology and geomorphology is desirable but not required.
  • Additionally, familiarity with Mars image datasets, raster data handling, projection systems, and image georeferencing. 

To apply for this project please fill in an online application form through the following link.

To see the full list of Internships available at ESA please go to our website for ESA Career Opportunities.