Multi-Class Crevasse Detection Using Ground Penetrating Radar and Feature-Based Machine Learning
Abstract
The purpose of this thesis is to describe a new automated system for processing of glacial Ground Penetrating Radar data to locate and label glaciologic features of interest. The data generated from expert examination of these labeled glaciologic features of interest support a team from the University of Maine in building a numerical model of the dynamics of the Ross Ice Shelf. In 2014 and 2015, GPS strain field and GPR surveys were conducted within a heavily crevassed shear margin between the Ross Ice-Shelf and McMurdo Ice-Shelf known as the McMurdo Shear Zone. The autonomous ground robot Yeti towed a radar-equipped sled to collect the GPR datasets. Automation of data collection has resulted in a size increase of glacial GPR datasets, the processing of which are time and personnel intensive. The 2014 GPR dataset required over 200 hours to label.This dissertation presents a feature vector based machine learning method to automatically process glacial GPR data using two Histogram of Oriented Gradients (HOG) feature vectors with a Support Vector Machine (SVM) algorithm to detect glaciologic features of interest. The deeper HOG model is used for general feature of interest classification, and a shallow HOG model is used to differentiate between shallow and deep features of interest.The deeper HOG model is used in a SVM to label GPR data. The radar data were filtered with a Triangle FIR filter. The filtered deeper HOG SVM model had a True Positive rate of 92.3\%, 6 False Positive events, and a 7.7\% False Negative rate.The shallow HOG SVM model achieved a True Positive rate over 90.3\%, 80 False Positive events, and a False Negative rate of 9.7\% in differentiating between shallow and deeper glaciologic features of interest.This work produced a system that can reliably detect and differentiate glaciologic features of interest in glacial GPR datasets. The system also delivers a method of review for labeled data images of selected data types, allowing the user to refine the label location relative to data scan number. The system also outputs a two dimensional GPS-referenced map based on the automated label.
- Publication:
-
Ph.D. Thesis
- Pub Date:
- 2019
- Bibcode:
- 2019PhDT........36W
- Keywords:
-
- Geophysics;Artificial intelligence;Applied physics