Crater counting

Project Leader: Prof. Gretchen Benedix, Curtin University

Summary

To understand planetary formation and evolution, and the history of our solar system, we need to look further than our own unique planet, and see how it fits into the ‘bigger picture’.  We already know quite a lot about Mars, and have satellite imagery of the entire planet.  Planetary scientist Professor Gretchen Benedix at Curtin’s Space Science and Technology Centre is working out how old different features on the Martian surface are just by looking at them.  But her method needs Pawsey’s newest GPU clusters to count the millions of overlapping crater impacts that reveal Mars’ history.

 
core hours allocated
 
core hours allocated
 
craters found bigger than 25m in 24 hrs
Partner Institution: Curtin University System: Topaz, Magnus Areas of science: Artificial Intelligence and Image Processing, Extraterrestrial Geology Applications used: CTX, HiRISE. In the rewrite applications used; Imagemagick, Graphicsmagick, Python libraries, Docker, Singularity, Slurm, Linux, Bash

The Challenge

The craters on a planet’s surface tell its history.  The more craters, the older the surface since a volcanic or climactic event wiped it ‘clean’.  Older craters bear the scars of newer impacts on top of them.  Crater counting is the principal tool used by astrogeologists to determine the surface ages of planets, our Moon and even large asteroids throughout the solar system.  Until now, the technique has relied on scientists painstakingly identifying and counting craters by hand.  The current database for Mars contains 385,000 identified craters with diameters of 1 km or larger.  But it took at least six years to construct, before it was published in 2012.

“There is so much more information available now,” enthuses Professor Benedix.  “Using images sent by the Mars Reconnaissance Orbiter we have 5 m per pixel resolution, so we can see 25 m craters.  But here’s the catch – when asteroids collide in space, they break into a couple of really big pieces, a few more medium sized pieces, and lots of tiny pieces.  This means there is a size frequency distribution for craters with smaller craters numbering exponentially higher than larger craters.  The smaller craters we can see now amount to tens of millions.”

Counting the smallest craters would let us create age maps with much higher accuracy and spatial resolution, and let us derive the surface ages for much smaller and younger features of the Martian surface.

The Solution

“We had to automate crater counting,” says Professor Benedix.  Combining the increasing surface detail provided by the Mars Orbiter with recent developments in object detection technology and machine learning, Benedix’ team turned to specialists at the Curtin Institute of Computation (CIC) to help build a convolutional neural network to identify circles in an image.  The algorithm was then put into a pipeline to use Pawsey GPUs to analyse the images from Mars.

Machine learning algorithms need a database of examples to learn from, so the algorithm was trained to recognise craters using images of 7,048 manually-identified craters from the Mars database.  Once trained, when the algorithm analysed the remainder of the images making up the Mars crater database, it generated results comparable with the manual dataset, giving 91% positive identification for craters 1 km and larger.

 

 

Watch Gretchen on decoding the surface age of Mars:

There is significant variability – around 85% – between different manual crater data sets, arising from how subjectively people interpret an image, and their variations in attention and focus on repetitive tasks.  Professor Benedix’s algorithm identified craters just as well as a human observer, but more objectively, reproducibly, and very much faster.

With the algorithm ready to identify even smaller craters from the 5 m per pixel high resolution imagery from the Mars Orbiter, Benedix requested Astronomy Data and Computing Services (ADACS) support to get a Pawsey developer embedded in the research project for six months.

“That made such a difference,” says Professor Benedix.  “Through improving the algorithm pipeline, the library training data set, and getting access to the newest GPU clusters at Pawsey, we went from being able to count craters in 100 m per pixel images fairly quickly to being able to analyse the global Mars dataset at 5 m per pixel, identifying 94 million craters bigger than 25 m in 24 hours.  We’re now analysing image sets of specific areas of Mars at a resolution of 30 cm per pixel.  We can see craters the size of a car.”

Outcome

Professor Benedix’s team has created the largest Mars crater database in the world.

They’ve located and dated 20 of the youngest craters on the surface of Mars, and are now determining if any are related to the Martian meteorites that have been found here on Earth.  They’re working out when the last volcanoes were active and dating channel structures to see when water last flowed on the surface of Mars, along with any potential for life.  It’s information that may well help determine where the next NASA Mars Mission will explore, and all adds to our understanding of planetary formation and history.

Applications for the algorithm also extend well beyond Mars.  “We’ve already applied it to the surface of the Moon and Mercury,” says Professor Benedix.  “We just needed to create a data training library specific to those different image sets and retrain the algorithm a bit.”

It can even be used on Earth – the algorithm and pipeline at Pawsey, using the right library of training images, could equally see applications ranging from identifying cancerous spots on skin or unusual cells in pathology samples to recognising pipeline damage in images from remote underwater vehicles monitoring gas pipelines.  Radar and sonar images could also be screened to automatically identify features of interest – you just need to teach the algorithm what to look for.

It can even be used on Earth – the algorithm and pipeline at Pawsey, using the right library of training images, could equally see applications ranging from identifying cancerous spots on skin or unusual cells in pathology samples to recognising pipeline damage in images from remote underwater vehicles monitoring gas pipelines.  Radar and sonar images could also be screened to automatically identify features of interest – you just need to teach the algorithm what to look for.
Prof. Gretchen Benedix, Curtin University,
Project Leader.