Case Study: NASA Ames Research Center achieves 97.95% satellite image classification accuracy with NVIDIA GPUs and Deep Learning

An Ai Monitor Of Earth’s Vitals

NASA Ames Research Center worked with NVIDIA to tackle the challenge of monitoring Earth’s changing climate through satellite image classification. The team needed a way to automate analysis of massive, noisy satellite datasets at global scale, which was difficult with conventional tools and required significant compute and memory resources. NVIDIA products used included Tesla GPUs on the NASA Ames Pleiades supercomputer, NVIDIA DIGITS DevBox, cuDNN, and popular deep learning frameworks.

Using NVIDIA GPU-accelerated deep learning, NASA Ames developed DeepSat, a framework for satellite image classification and segmentation trained on millions of parameters and hundreds of thousands of image scenes. NVIDIA helped speed training and large-scale inference, enabling NASA to improve classification and segmentation accuracy; the best network reached 97.95% accuracy and outperformed three state-of-the-art object recognition algorithms by 11%. Training times were also reduced, allowing faster experimentation and innovation.

Nasa Ames Research Center

Sangram Ganguly

Senior Research Scientist