Trees constitute one of the world’s most important carbon sinks, pulling enormous amounts of carbon dioxide from the atmosphere and storing the carbon in their trunks and the surrounding soil. Measuring this carbon storage, however, is extraordinarily difficult: the world is populated by trillions of trees, largely in unpopulated areas, and tree species and density can make large differences in the amount of carbon stored by a given forested area. Moreover, many current efforts ignore trees outside of forested areas (i.e. individual trees and small clusters of trees). Now, NASA has worked with international institutions to develop a supercomputer-powered tool to more effectively map trees and estimate their stored carbon.
“From a carbon cycle perspective, these dry areas are not well mapped, in terms of what density of trees and carbon is there,” said Martin Brandt, an assistant professor of geography at the University of Copenhagen, in an interview with NASA’s Jessica Merzdorf. “It’s a white area on maps. These dry areas are basically masked out. This is because normal satellites just don’t see the trees – they see a forest, but if the tree is isolated, they can’t see it. Now we’re on the way to filling these white spots on the maps. And that’s quite exciting.”
The scientists – led by researchers from NASA’s Goddard Flight Center – began with commercial satellite images that were high-resolution enough to show individual trees in sufficient detail. Focusing on dryland regions like the Sahara Desert, the team began training a deep learning model to recognize trees using a training dataset of nearly 90,000 hand-marked trees across a variety of terrain types. With Brandt alone tasked with identifying all the trees, the training dataset took over a year to develop.
Even then, tuning the model wasn’t an easy task.
“In one kilometer of terrain, say it’s a desert, many times there are no trees, but the program wants to find a tree,” said Brandt. “It will find a stone, and think it’s a tree. Further south, it will find houses that look like trees. It sounds easy, you’d think – there’s a tree, why shouldn’t the model know it’s a tree? But the challenges come with this level of detail. The more detail there is, the more challenges come.”
The model aimed to identify the “crown diameter” of each tree (that is, its diameter when viewed from above) of over 1.8 billion trees across half a million square miles. To run the algorithm at this immense scale, the team turned to supercomputing – namely, the Blue Waters system at the National Center for Supercomputing Applications (NCSA). Blue Waters is a hybrid Cray system with more than 1.5 petabytes of memory that delivers around 13 peak petaflops.
After running the model on those billions of trees, the team correlated how tree crown diameter, coverage and density exhibited variation based on rainfall and land use — finding many more trees than expected in arid and semi-arid regions. The researchers see wide applications for this data in understanding ecosystems and carbon fluxes around the world.
“There are important ecological processes, not only inside, but outside forests too,” said Jesse Meyer, the NASA programmer who led the work on Blue Waters. “For preservation, restoration, climate change, and other purposes, data like these are very important to establish a baseline. In a year or two or ten, the study could be repeated with new data and compared to data from today, to see if efforts to revitalize and reduce deforestation are effective or not. It has quite practical implications.”