English
 
User Manual Privacy Policy Disclaimer Contact us
  Advanced SearchBrowse

Item

ITEM ACTIONSEXPORT

Released

Journal Article

Deep Learning-driven Detection and Mapping of Rockfalls on Mars

MPS-Authors
/persons/resource/persons221946

Bickel,  Valentin Tertius
Department Planets and Comets, Max Planck Institute for Solar System Research, Max Planck Society;

/persons/resource/persons104064

Mall,  Urs
Department Planets and Comets, Max Planck Institute for Solar System Research, Max Planck Society;

External Ressource
No external resources are shared
Fulltext (public)
There are no public fulltexts stored in PuRe
Supplementary Material (public)
There is no public supplementary material available
Citation

Bickel, V. T., Conway, S. J., Tesson, P.-A., Manconi, A., Loew, S., & Mall, U. (2020). Deep Learning-driven Detection and Mapping of Rockfalls on Mars. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. doi:10.1109/JSTARS.2020.2991588.


Cite as: http://hdl.handle.net/21.11116/0000-0006-87BF-B
Abstract
The analysis of rockfall distribution and magnitude is a useful tool to study the past and current endogenic and exogenic activity of Mars. At the same time, tracks left by rockfalls provide insights into the mechanical properties of the martian surface. While a wealth of high-resolution space-borne image data are available, manual mapping of displaced boulders with tracks is inefficient and slow, resulting in 1) a small total number of mapped features, 2) inadequate statistics, and 3) a sub-optimal utilization of the available big data. This study implements a deep learning-driven approach to automatically detect and map martian boulders with tracks in High Resolution Imaging Science Experiment (HiRISE) imagery. Six off-the-shelf neural networks have been trained either on martian or lunar rockfall data, or a combination of both, and are able to achieve a maximum recall of up to 0.78 and a maximum precision of up to 1.0, with a mean average precision of 0.71. The fusion of training data from different planets and sensors results in an increased detection precision, highlighting the value of domain generalization and multi-domain learning. Average processing time per HiRISE image is ~45 s using an NVIDIA Titan Xp, which is more than one order of magnitude faster than a human operator. The developed deep learning-driven infrastructure can be deployed to map martian rockfalls on a global scale and within a realistic timeframe.