Ice Crevasse Detection with Ground Penetrating Radar using Faster R-CNN

To further enhance the security capability of polar scientific research, ground penetrating radar (GPR) has been used for ice crevasse detection. In this paper, we proposed a crevasse detection method based on Faster Region Convolutional Neural Network (Faster R-CNN), which can detect crevasses and continuous snow layers automatically in a short time. First, feature maps of GPR images extracted by convolutional layers of Faster R-CNN and region proposals generated by Region Proposal Networks (RPN) are input into regions of interest (ROI) pooling layers. Then the proposal feature maps can be obtained and used to identify crevasses and continuous snow layers. Finally, classification results are obtained according to analyzing confusion matrix. Experimental results show that our method can detect crevasses with high accuracy rate, F1 score and Kappa coefficient while low false positive rate, false alarm rate and missing report rate. The proposed method will be useful for automatic ice crevasse detection in real-time.