Necessity may be the mother of invention.
But sometimes it’s just all the mosquitos.
Leila Donn, a doctoral student at the University of Texas at Austin studying environmental geoscience, wasn’t necessarily looking for a computer model to help her find the location of ancient Mayan caves last year. Mostly, she just was hot and tired and the work was going slowly.
“In summer the of 2018 I went to the ancient Maya site of El Zotz in the Petén Department of Guatemala to help a colleague look for caves. Each evening we’d sit down and take a look at the LIDAR hillside, looking for depressions surrounded by steep areas,” she said.
LIDAR – light detection and ranging – is a remote sensing method that uses light in the form of a pulsed laser to measure variable distances to the Earth. These light pulses, combined with other data recorded by the airborne system, generate precise, three-dimensional information about the shape of the Earth, its surface characteristics, and can provide a high-resolution digital elevation model to study the shape of the landscape at a meter or sub-meter scale.
“Once we found such an area,” she said, “we’d pull up a cross section through the associated point cloud, and check for points that were located below the ground surface.”
The next day, as the day before, as the day before that, and along with her team members, Donn hiked out to the spot to verify.
“We probably identified about 20 potential caves in this manner. We made it out to 10 of these locations over the course of six days, which meant full days of off-trail, largely uphill hiking through densely vegetated and mosquito-infested areas,” she said.
Only three of those 10 of locations were caves, though, and only one of those caves was large and interesting enough to be included in a National Geographic documentary – one in which Donn appeared.
She thought it was particularly unproductive way to pursue the work.
“While hard tropical forest hikes actually are my idea of fun, looking for caves in this manner wasn’t particularly efficient,” she said.
She had an idea.
“It occurred to me that it would be tremendously useful if we could develop a computational method of predicting likely locations of caves,” she said.
At first, she assumed someone had already done the work, for similar efforts to find sinkholes had been done, but she was surprised to find that no one had actually applied machine learning for searches like hers.
So she went to work.
“First, I acquired a training dataset consisting of a shapefile of known caves and 1-meter resolution LIDAR imagery from an area that is geomorphologically similar to my study area in northwestern Belize,” said Donn.
She did this by using Python and ArcGIS to make raster layers to capture morphologic characteristics that were associated with cave entrances, such as the fact that they are frequently located in areas of very steep topography.
“I then extracted the numerical values representing cave morphology, also doing the same thing for a series of randomly generated not-cave points that served the purpose of teaching the computer what the normal cave-free background landscape looks like,” Donn explained.
From there, she created a matrix of all the numerical values associated with the morphologies of all known and unknown cave points.
The result was pretty astonishing.
She “trained” the computer to identify caves with 89-percent accuracy using a random forest classifier.
Oil and Gas Application
Donn, whose doctorate work includes study in geomorphology, caving, soil description and geoarchaeological methods, sees a use for computational modeling in a number of fields, like oil and gas.
“With regard to petroleum geology, these products could potentially be used to determine areas that are more likely to have geomorphology favorable to oil and gas deposits. Machine learning could then be applied to classify the landscape into areas that have similar geomorphology to that of known oil and gas deposits,” she said.
To that end, she recently presented her results – the first time she’s been a presenter – at the recent Geological Society of America conference in Phoenix, Ariz..
“The conference put out a cool press release about my project,” Donn noted.
She is somewhat bemused by all the attention.
“The bit of buzz that this has created is unexpected,” she said, especially because of the excitement and interest of machine learning and artificial intelligence.
Donn, whose master thesis at UT Austin was on the “Human-Environment Impact on Geomorphology and Flood Recurrence of the Belize Rover Valley,” now wants to expand her work to predict caves more accurately.
“I’d like my program to identify only caves versus small cave-like voids. It would also be tremendously helpful if I were able to acquire an additional training dataset from an area of similar geomorphology,” she said.
Donn knows that the more predictive caves she confirms, the larger her training dataset; so fieldwork will play a large component in her work in the years to come. She sees the multilevel nature of it all.
“My project straddles the fields of machine-learning and geomorphology, which is a newly growing area based on the long tradition of using modeling to study geomorphologic trends,” she said. “I think that the combination of the mystery of caves combined with the newness/perceived high-techness of artificial intelligence makes this work something that lots of people think sounds pretty exciting.”