Novelty and Discovery Content Analysis Methods for the Planetary Data System Image Atlas

Monday, 7 December 2020
Paul Horton1, Sanjna Ravichandar2, Jake Lee3, Hannah Rae Kerner4, Anil Natha3, Tariq K Soliman3, Kevin Grimes3, Kiri Wagstaff3, Rishi Verma3 and James McAuley3, (1)Arizona State University, School of Earth and Space Exploration, Tempe, AZ, United States, (2)Massachusetts Institute of Technology, Electrical Engineering & Computer Science, Cambridge, MA, United States, (3)Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA, United States, (4)University of Maryland College Park, Geography, College Park, MD, United States
The Planetary Data System Image Atlas provides access to the public archive of images acquired by NASA interplanetary missions. Like a photographer, mission instruments capture several images of the same subject. These images are by default listed on the Atlas in reverse chronological order yielding an album filled with redundant images. To assist navigation, the Atlas provides tools to filter and classify images by metadata including time of collection, instrument used, and content (e.g., craters, dunes, layered rocks). These tools are valuable for researching specific interests, but less useful when browsing for interesting or unusual examples. To find these examples, users must manually browse through the Atlas which can be a time consuming and impractical process.

To meet these users’ needs, we implemented several image novelty and discovery tools. Novelty detectors analyze large data sets and highlight distinct imagery. These methods detect outliers by calculating how much an image’s content differs from the overall data set. We implemented algorithms ranging from spectrally (composition) sensitive models such as convolutional autoencoders to morphologically (shape) sensitive models such as Reed-Xiaoli detectors. Different models are trained for each instrument which allow users to investigate the most interesting images instead of the most recent.

We also implemented a discovery tool to quickly identify a diverse subset of interesting images. We employed a pre-trained convolutional neural network to extract content-based features from each image, then used the Discovery via Eigenbasis Modeling of Uninteresting Data (DEMUD) algorithm to rank images by unique or interesting content. To detect novel images, DEMUD iteratively builds a model to represent data that have already been displayed. The remaining items are ranked based on reconstruction error, and the item with the largest error is selected for the next iteration. This tool provides users with a highly diverse collection of images that is easy to browse without any mission-specific knowledge.

With the integration of these tools, the capabilities of the Atlas can expand to provide an overview of highly diverse images for each NASA mission, group images by similarity, and sort albums by their most interesting content.