AI STRATEGY
CREDIT: NASA
NASA Ames Pleiades
Supercomputer GPU cluster
This wasn’ t about processing pretty pictures from space but about understanding changes to our planet’ s carbon cycle, vegetation patterns and climate systems. The scale of this undertaking was staggering.
NASA trained DeepSat using 330,000 image scenes across the continental US, with average image tiles measuring 6,000 by 7,000 pixels and weighing about 200 megabytes each.
The entire dataset approached 65 terabytes for a single time epoch, with ground resolution down to one metre.
Sangram Ganguly, former Senior Research Scientist at NASA Ames Research Center and now CTO at Rhombus Power, reported the results:“ Our best network dataset produced a classification accuracy of 97.95 % and outperformed three state-of-the-art object recognition algorithms by 11 %,” he said.
DeepSat now enables scientists to quantify carbon sequestration by vegetated landscapes, downscale climate projection variables and assess urban heat island effects. The system also provides critical data for understanding how our planet responds to climate change, information that governments and organisations worldwide rely upon for policy decisions.
The computational core for this work came from Nvidia’ s Tesla GPUs and the NASA Ames Pleiades Supercomputer GPU cluster, equipped with 217,088 CUDA cores. aimagazine. com 133