Machine Learning for Data Validation and Natural Event Detection using NASA GIBS Earth Satellite Imagery

NASA's Global Imagery Browse Services (GIBS) provides a full-resolution image archive of over 70 million images and access to services for over 700 NASA Earth science data products covering every part of the world. Most imagery is available within 3-5 hours after satellite overpass and some products span almost 30 years. While the GIBS satellite data remains highly accessible, it remains largely underexploited and analyzed due to its scale. The development of algorithms to ensure the consistency of data is critical for near-real time (NRT) applications. We identify two machine learning tasks to explore: (1) data validation and (2) natural event detection. For data validation we apply machine learning techniques to detect artifacts from satellite drop-out regions, improperly calibrated data, and confounding natural events such as eclipse shadows. For natural event detection we illustrate the promise of deep learning to leverage the scale of GIBS imagery to learn state-of-the-art algorithms for pixel-level wildfire detection that may be used for providing health alerts.