Mapping “Brain Coral” Regions On Mars Using Deep Learning

Mapping “Brain Coral” Regions On Mars Using Deep Learning

An outline of the inputs for every step in our processing pipeline. a) A area within the HiRISE picture of ESP 018707 2205 is proven on the native decision (0.3m/px) with a blue field highlighting the window measurement for our classifier community. b) A 256 x 256 pixel window is used as enter for our spatial classifier algorithm nonetheless the window is at 1/16 the unique decision (1.2m/px). A grid of 8×8 squares is displayed exhibiting how that picture will get tiled and preprocessed utilizing components of a JPEG encoder which entails the discrete cosine remodel (d). c) A single 8×8 tile flattened right into a 1d array which is used for the DCT remodel (d). e) A block of Fourier coefficients is rearranged into a knowledge dice and used as enter for our Fourier classifier. Lowering the picture measurement and finally the channel measurement after the primary conv. layer considerably shortens the community’s processing time in comparison with the spatial picture enter. — astro-ph.EP

One of many foremost targets of the Mars Exploration Program is to seek for proof of previous or present life on the planet. To realize this, Mars exploration has been specializing in areas which will have liquid or frozen water.

A set of vital areas might have seen cycles of ice thawing within the comparatively current previous in response to periodic adjustments within the obliquity of Mars. On this work, we use convolutional neural networks to detect floor areas containing “Mind Coral” terrain, a landform on Mars whose similarity in morphology and scale to sorted stone circles on Earth means that it could have shaped as a consequence of freeze/thaw cycles.

We use massive photos (~100-1000 megapixels) from the Mars Reconnaissance Orbiter to seek for these landforms at resolutions shut to a couple tens of centimeters per pixel (~25–50 cm). Over 52,000 photos (~28 TB) had been searched (~5% of the Martian floor) the place we discovered detections in over 200 photos. To expedite the processing we leverage a classifier community (previous to segmentation) within the Fourier area that may make the most of JPEG compression by leveraging blocks of coefficients from a discrete cosine remodel in lieu of decoding all the picture on the full spatial decision.

The hybrid pipeline method maintains ~93% accuracy whereas slicing down on ~95% of the entire processing time in comparison with operating the segmentation community on the full decision on each picture. The well timed processing of huge information units helps inform mission operations, geologic surveys to prioritize candidate touchdown websites, keep away from hazardous areas, or map the spatial extent of sure terrain. The segmentation masks and supply code can be found on Github for the group to discover and construct upon.

Kyle A. Pearson, Eldar Noe, Daniel Zhao, Alphan Altinok, Alex Morgan

Feedback: Submitted for publication, looking for feedback from the group. Code out there: this https URL
Topics: Earth and Planetary Astrophysics (astro-ph.EP); Instrumentation and Strategies for Astrophysics (astro-ph.IM); Machine Studying (cs.LG); Picture and Video Processing (eess.IV)
Cite as: arXiv:2311.12292 [astro-ph.EP] (or arXiv:2311.12292v1 [astro-ph.EP] for this model)
Submission historical past
From: Kyle Pearson
[v1] Tue, 21 Nov 2023 02:24:52 UTC (13,166 KB)


Leave a Reply

Your email address will not be published. Required fields are marked *