Deep learning used to map storms on the planet Saturn

An article published in the journal “Nature Astronomy” reports the application of a deep learning algorithm to recognize storms on the planet Saturn creating a map of their components and characteristics. Ingo Waldmann of the University College London, England, and Caitlin Griffith of the University of Arizona, USA, developed the algorithm called PlanetNet based on the TensorFlow “engine” to analyze data collected by the Cassini space probe with an efficiency higher than traditional techniques thanks to the recognition of recurrent characteristics across various multiple data sets. PlanetNet could be adapted to analyze observations of other planets.

The Cassini space probe’s mission ended with its destruction in the atmosphere of the planet Saturn on September 15, 2017 but left a wealth of data as a legacy. For this reason, the classical techniques used to analyze them may require very long processing times providing results that are not always accurate. Among Cassini’s instruments there was VIMS (Visible and Infrared Mapping Spectrometer), which in February 2008 collected observations on a number of close storms on Saturn, indicated by the arrow in the bottom image (NASA, JPL, Space Science Institute). It’s a complex situation on a large scale and for this reason it’s suitable to test a deep learning algorithm.

The dataset to be analyzed was fed to PlanetNet, an algorithm based on the TensorFlow “engine” released by Google in November 2015 as a free / open source library and available on GitHub. In a few years TensorFlow has been adapted into very different fields applications and even in astronomy it’s used in various ways, with PlanetNet for the analysis of data collected by the Cassini space probe’s VIMS instrument.

PlanetNet searched in the data for signs of clustering in the cloud structure and gas composition. In the areas of interest, the data were trimmed to eliminate uncertainties at the edges and generated a parallel analysis of spectral and spatial properties. The two data streams were recombined creating a map that quickly and accurately presents the major components of Saturn’s storms with unprecedented accuracy showing the planet’s regions of the atmosphere that got struck and the clouds containing materials swept from the lower layers of atmosphere by vertical winds.

Previous analyzes of the same dataset made using traditional techniques allowed to find traces of ammonia in the atmosphere of Saturn in S-shaped clouds. They’re also present in the map created by PlanetNet and indicated by the arrow in the top image (Courtesy I. P. Waldmann & C. A. Griffith, Nature Astronomy 2019, Nature Astronomy 2019. All rights reserved). The algorithm recognized that type of formation around another smaller storm as well and this suggests that it might be quite common.

To validate PlanetNet’s accuracy on datasets collected by the Cassini space probe, the researchers used other data not included in the training phase. The entire dataset was also rotated and resampled to create synthetic data for further testing. In both cases, PlanetNet achieved over 90% classification accuracy.

The possibility of analyzing large datasets in a shorter time will allow the creation of better storm maps with results that include atmospheric elements that weren’t recognized with classical analysis techniques. PlanetNet can be trained to work with datasets collected during other space missions offering a potentially huge help in the study of other planets.

Leave a Reply

Your email address will not be published. Required fields are marked *