SEGMENTING IMAGES AUTOMATICALLY FOR GRANULOMETRY AND SEDIMENTOLOGY OF MARTIAN SOIL

Terrestrial hydrology, bed surface sedimentology, and sediment core analyses are a few of many terrestrial applications where segmenting photos of unconsolidated sediments has aided their detailed granulometry. Nevertheless, the mechanical sieve and grid-by-number in the field remain the gold standards on Earth, even when analyzing rock piles in the mining industry. Unfortunately, planetary exploration generally lacks physical samples and field access; image segmentation is the only path to sedimentological comparisons with the extensive terrestrial literature. But planetary scientists have been thwarted by the dearth of segmentation algorithms customized for planetary applications, of Mars in particular, and forced to rely on sub-optimal solutions adapted from medical software. We bridge the gap between mature segmentation software used in terrestrial settings and emergent planetary segmentation with an original algorithm optimized to segment whole images from the Mars Exploration Rovers' Microscopic Imagers. Many issues including natural illumination, foreshortening, imbrication, allocation of pore space and unresolvable material, and maximization of grain area bedevil automation. We demonstrate the robustness of our code to such challenges relative to a human operator, summarize granulometric methods that would enable robust comparisons with terrestrial data, and present quantitative limitations to automated segmentation. Consistency and areal distributions are key strengths of the algorithm, potentially enabling robust comparisons across missions and even planets, unlike a human operator. When implemented in Mathematica, the algorithm segments an entire MI image within a minute, far surpassing the extent and speed possible with manual segmentation. Our work enables not only new sedimentological insight from the MER MI data, but also detailed sedimentology with the Mars Science Laboratory’s MAHLI instrument. Furthermore, while the code operates with minimal human guidance, its default parameters can be modified for different geologic settings and imagers on earth and other planets. In addition to strengths in the planetary context, it incrementally improves upon terrestrial segmentation algorithms.