Presentation #513.05 in the session Fire and Ice: Europa and Aurorae.
Europa’s young, constantly evolving ice surface and subsurface liquid water ocean has been a source of curiosity among the planetary science and astrobiology community, amplified by the close-up observations of Voyager 1 & 2 and Galileo spacecraft. This project deployed various Machine Learning (ML) models, specifically Deep Learning Convolutional Neural Networks (CNN), to perform object detection of ice blocks on Europa from Galileo images. These models will increase our understanding of the exchange of materials between the surface and subsurface on Europa and could potentially be used for creating a comprehensive surface map helping future space exploration missions select sites for sampling and landing. This understanding and new insight into ice blocks could also give possible insights into the habitability of Europa. According to Leonard et al. 2022, ice blocks are defined on a morphological spectrum of platy (large) to knobby (small). Platy chaos is defined as a unit where the texture and morphology of the pre-existing terrain is clearly visible on the ice blocks. Knobby chaos is defined as a unit where the pre-existing morphologies are not apparent and usually consists of smaller blocks. Both platy and knobby chaos have sharp boundaries with the surrounding ice matrix and are considered ice blocks. After testing several models, the Mask Region-CNN (Mask R-CNN) was able to detect various ice blocks in a bounding box and a pixel-level approach both using a ML technique called instance segmentation. Instance segmentation goal is to detect iterations of objects and demarcate their boundaries. After 20 epochs, the algorithm was able to reach 33% precision with the bounding box approach and 30% with the instance segmentation approach. There are several challenges that were faced to find the best model to classify ice blocks. First, the project is restricted to a relatively “small” dataset compared to usual sizes of ML datasets. This problem was solved by using a technique of ML called Transfer Learning (TL) which uses a model (ResNet50) previously trained on a dataset (ImageNet) of millions of images of common household objects. The use of TL greatly advanced and accelerated the basic training of the model compared to ab initio learning. Other difficulties include low spatial resolution, varying illumination and block morphologies. The Mask R-CNN with the TL backbone provides a dynamic and robust approach that can tackle these difficulties. In the future, we plan to ingest and process upcoming Juno Spacecraft and Europa Clipper images of Europa’s surface to compare to results from Galileo imagery.