TextureCam: Autonomous Image Analysis for Astrobiology Survey

Introduction: The TextureCam concept is a “smart camera” that aims to improve scientific return by increasing science autonomy and observation capabilities both when the spacecraft is stationary and when it is traversing. This basic onboard science understanding can summarize terrain encountered during travel [1] and direct autonomous instrument deployment to targets of opportunity [2]. Such technology would assist Mars sample caching or return missions where the spacecraft must survey vast areas to identify potential sampling sites and habitat indicators. The instrument will use texture channels to differentiate and map habitat-relevant surfaces. Here we use the term texture not in the geologic sense of formal physical properties, but rather in the computer vision sense to mean statistical patterns of image pixels. These numerical signatures can in turn distinguish geologically relevant elements such as roughness, pavement coatings, regolith characteristics, sedimentary fabrics and differential outcrop weathering. Similar algorithms can perform microstructure analysis and sedimentology using microscopic images. On the scale of meters, surfaces and features can be recognized to summarize rover traverse and draft surficial maps for downlink. Reliable image classification in field conditions is difficult due to factors like variable lighting, surface conditions, and sediment deposition. Future development aims for reliable recognition of basic geological elements. Rad-hardened FPGA processing will instantiate these general-purpose texture analysis algorithms as part of an integrated, flight-relevant prototype. Experimental setup: An initial test demonstrates automatic recognition of stromatoform structures in outcrop. These surfaces have distinctive laminar textures that may indicate previous biogenic activity [3] and would be valuable targets for followup investigation and human review. While it would be unlikely to find such features exposed on the Mars surface, they provide a controlled test to guide initial development and evaluation. We use a machine learning approach, training the system on example pixel labels supplied by the designer. This yields a statistical model of pixel relationships that can classify new surfaces (Figure 1). Stromatolite images were acquired from the Pilbara formation [3]. We consider three outcrops of laminar structures with conical and wavy morphologies. We also include a scene containing much finer-scale localized egg carton laminations. These all show a common lay100 200 300 400 500 600 50