Glacial Cirque Identification Based on Convolutional Neural Networks

Abstract

Cirques provide important information about the characteristics of past glaciers and their associated palaeoclimate conditions. However, mapping cirques is challenging and time-consuming due to their fuzzy boundaries. A recent study tested the potential of using a deep learning algorithm, Convolutional Neural Networks (CNN), to detect the boundary boxes of cirques. Based on a similar CNN method, RetinaNet, we use a dataset of > 8,000 cirques and various combinations of digital elevation models and their derivatives to detect cirques. We also incorporate the Convolutional Block Attention Module (CBAM) into RetinaNet for training and prediction. The precision of cirque detection is evaluated for various input data combinations, training sample sizes, and with or without the addition of the CBAM, based on comparison with mapped cirques in two test areas on the Kamchatka Peninsula and the Gangdise Mountains. The results show that the addition of CBAM increases the average precision by 4-5%, and the trained model can detect the cirque boundary boxes with high precision (84.7% and 87.0%), recall (94.7% and 86.6%), and F1 score (0.89 and 0.87), for the two test areas on the Kamchatka Peninsula and in the Gangdise Mountains, respectively, and significantly reduce the number of undetected cirques. The performance becomes relatively stable after the training dataset increases to > 6000 samples. The trained model can effectively detect boundary boxes that contain cirques to help subsequent cirque outline extraction and morphological analysis.

Keywords: Cirques, Object detection, RetinaNet, CBAM