Multi-Class Crevasse Detection Using Ground Penetrating Radar and Feature-Based Machine Learning

摘要

This paper describes a system for automated processing of glacial Ground Penetrating Radar (GPR) data to locate and label deep and shallow crevasses. In 2014 and 2015 robotic GPR surveys were conducted within the heavily crevassed McMurdo Shear Zone (MSZ), a shear margin between the Ross and McMurdo Ice Shelves. A feature-based machine learning method to process the GPR data uses Histogram of Oriented Gradients (HOG) feature vectors with a Support Vector Machine (SVM) to detect deep and shallow crevasses. A deep HOG model is used for general crevasse classification and achieves a True Positive (TP) rate over 99%, less than 2.6% False Positive (FP) rate, and less than 1% False Negative (FN) rate. A shallow HOG model differentiates between shallow and deep features of interest with a TP rate of 92%, a FP rate of 6%, and a FN rate of 8%.