Content-based Classification of Mars Imagery for the PDS Image Atlas

The Planetary Data System (PDS) Imaging Node hosts millions of images acquired from the planet Mars. Missions such as the Mars Science Laboratory (MSL) and Mars Reconnaissance Orbiter (MRO) continually collect new images to enrich our understanding of Mars. These images are served for public access by the PDS Image Atlas (Atlas). With the constantly growing image volume, connecting scientists and other users to images of interest has become a challenge. Images delivered to the PDS Imaging Node contain metadata regarding when and how the images were processed and transferred to the Earth. However, users of the Atlas are often interested in finding images based on content. The content-based information is not included in the metadata and must be extracted through content analysis.

To enable content analysis for efficiently finding images of interest, we employed transfer learning to adapt a previously trained convolutional neural network (CNN) to classify images acquired by the MRO High Resolution Imaging Science Experiment (HiRISE) instrument, the MSL Mast Camera (Mastcam), and the MSL Mars Hand Lens Imager (MAHLI). The initial CNN classifiers were deployed on the Atlas for public use in 2017. Over the past year, we have employed several methods to improve both the accuracy of the classifiers and the reliability of their confidence values. First, we modified our fine-tuning methodology which led to a significant improvement in accuracy for the MSL classifier. Second, we analyzed and categorized each classifier's common errors in the training and validation sets. Discoveries made during error analysis for the initial MSL data set motivated the creation of a newly labeled version 2 MSL data set. Several active learning techniques were evaluated during the creation of the MSL version 2 data set to reduce human labeling effort. Finally, we employed classifier calibration methods to improve the reliability of classifiers' self-reported posterior probabilities. In addition to these improvements, we are also evaluating several methods to enable the classifiers to adapt to changes in image properties over space and time.