Crevasse Detection in Ice Sheets Using Ground Penetrating Radar and Machine Learning

Abstract

This paper presents methods to automatically classify ground penetrating radar (GPR) images of crevasses on ice sheets. We use a combination of support vector machines (SVMs) and hidden Markov models (HMMs) with down sampling, a preprocessing step that is unbiased and suitable for real-time analysis and detection. We perform modified cross-validation experiments with 129 examples of Greenland GPR imagery from 2012, collected by a lightweight robot towing a GPR. In order to minimize false positives, an HMM classifier is trained to prescreen the data and mark locations in the GPR files to evaluate with an SVM, and we evaluate the classification results with a similar modified cross-validation technique. The combined HMM-SVM method retains all of the correct classifications by the SVM, and reduces the false positive rate to 0.0007. This method also reduces the computational burden in classifying GPR traces because the SVM is evaluated only on select prescreened traces. Our experiments demonstrate the promise, robustness, and reliability of real-time crevasse detection and classification with robotic GPR surveys.