AUTOMATIC DETECTION OF HIDDEN CREVASSES ON ICE SHEETS USING GROUND PENETRATING RADAR AND MACHINE LEARNING

Detection of hidden surface crevasses on glaciers is a vital process involved in over-snow traverses for science and resupply missions in Polar regions. The ability for science teams to safely traverse a glacier or ice cap is becoming significantly more critical with climate change, as increases in temperature lead to both acute and time-sensitive interest in studying the environment, and increased risk to researchers and their associated logistical teams. As temperatures climb and ice caps melt, crevasses begin to penetrate deeper, and grow wider with more erratic geometries. Of particular risk to humans in an already extreme environment are the tenuous snow bridges that completely hide a crevasse from view to a surface observer, even during days with full visibility.

The current protocol for crevasse detection employs a human-operated ground penetrating radar (GPR) on a mid-weight tracked vehicle. Radar surveys are necessary due to crevasses' surface invisibility, and because the application requires non-destructive imaging in order for vehicles and humans to cross over the cracks. To secure safe passage, a GPR scout team must plan an appropriate crevasse-free route by investigating paths across the glacier. This paper presents methods to continue development of a completely autonomous robotic system employing GPR imaging of the glacier sub-surface. Autonomy of the system to classify GPR images is achieved using machine learning algorithms and appropriate un-biased processing, particularly those which are also suitable for real-time image analysis and detection. We tested and evaluated three pre-processing schemes in conjunction with a Support Vector Machine (SVM) trained with Antarctic GPR imagery, collected by our robot and a Pisten Bully tractor in 2009 and 2010 at McMurdo Station. We achieved 86% classification rate for a radial basis kernel SVM trained and evaluated on down-sampled GPR images of crevasses, compared to 92% using a texture feature coding method, and 83% using raw data for contrast. We also discuss independent versus sequential classification of GPR images, and suggest improvements to or combinations of the most successful training models. Our experiments demonstrate the promise and reliability of real-time crevasse detection with robotic GPR imaging surveys.