Researchers in North America recently published a paper outlining how deep learning has helped create a new technique to measure the size and location of crater impacts on the moon. The process has traditionally been done by visual inspection of images, thus limiting the scope. However, when using neural networks, the researchers were able to increase the efficiency and accuracy of the search.
The deep learning model was able to identify twice as many craters as traditional manual counting, including 6,000 previously unidentified moon craters. Crater counting allows scientists to better understand the history of the solar system.
“We have demonstrated the successful performance of a convolutional neural network in recognizing Lunar craters from digital elevation map images,” the researchers from the University of Toronto, Penn State University, and Arizona State University said. “Our results suggest that deep learning will be a useful tool for rapidly and automatically extracting craters on various Solar System bodies,” they added.
By using an NVIDIA Tesla P100 GPU and the cuDNN-accelerated TensorFlow deep learning framework, the researchers first trained the convolutional neural network on a dataset covering two-thirds of the moon. They then tested the neural network on the remaining third of the moon. Afterward, the Moon-trained neural network successfully detected craters on Mercury, a planet with a completely distinct surface than the moon.
“Tens of thousands of unidentified small craters are on the moon, and it’s unrealistic for humans to efficiently characterize them all by eye,” said Ari Silbur, one of the researchers, speaking to the University of Toronto News. “There’s real potential for machines to help identify these small craters and reveal undiscovered clues about the formation of our solar system,” he added.
The research is currently under review in the journal Icarus.