Automatic Classification and Segmentation of Patterned Martian Ground Using Deep Learning Techniques
Abstract
Science autonomy onboard spacecraft can optimize image return by prioritizing downlink of meaningful data. Martian polygonally cracked ground is actively studied by planetary geologists and may be indicative of subsurface water. Filtering images containing these polygonal features can be used as a case study for science autonomy and to reduce the overhead associated with parsing through Martian surface images. This thesis demonstrates the use of deep learning techniques in the classification of Martian polygonally patterned ground from HiRISE images. Three tasks are considered, a binary classification to identify images containing polygons, multiclass classification distinguishing different polygon types and semantic segmentation of polygon regions. Due to time and resource constraints, transfer learning is employed on state-of-the-art deep learning networks. Convolutional neural network model architectures are compared for the binary and multiclass scenarios. UNet was used for semantic segmentation. Overall, the models show promising results as a first filter method.