Planetary Mapping Using Deep Learning: A Method to Evaluate Feature Identification Confidence Applied to Habitats in...

Area of Science:

• Planetary Science
• Astrobiology
• Artificial Intelligence

Background:

• Mars exploration goals are shifting towards specific biosignature detection and human exploration.
• Translating terrestrial analog research to identify rover-explorable targets on Mars is a key challenge.
• Integrating ground-based analysis with orbital data for Mars rover missions needs improved strategies.

Purpose of the Study:

• To develop and present an end-to-end methodology for mapping geomorphologic units and quantifying feature identification confidence using Deep Learning.
• To assess the identification confidence of rover-explorable habitats in a Mars-analog environment.
• To determine the necessary spatial resolution for confident habitat identification on Mars.

Main Methods:

• Utilized small Unmanned Aerial System (sUAS) technology for high-resolution imagery acquisition.
• Applied state-of-the-art fully convolutional neural networks for pixel-wise classification.
• Quantified feature identification confidence across various spatial resolutions in the Salar de Pajonales analog.

Main Results:

• Developed a Deep Learning methodology to map geomorphologic units and assess identification confidence.
• Found that spatial resolutions twice as good as current Mars orbital data are necessary for 85% confidence in habitat identification.
• Demonstrated the approach's utility in comparing habitat identifiability across different Mars-analog environments.

Conclusions:

• The presented methodology can refine Mars exploration strategies from regional habitability to specific habitat targeting.
• The approach can be adapted to map Martian surface features like domes and ridges to study their origin and astrobiological potential.
• Higher spatial resolution data is crucial for confident identification of potential habitats for future Mars missions.