Glacial cirque identification based on Convolutional Neural Networks

Abstract

Cirques provide important information about the palaeoclimate conditions that produced past glaciers. However, mapping cirques is challenging, time-consuming, and subjective due to their fuzzy boundaries. A recent study tested the potential of using a deep learning algorithm, Convolutional Neural Networks (CNN), to predict boundary boxes containing cirques. Based on a similar CNN method, RetinaNet, we use a dataset of >8000 cirques worldwide and various combinations of digital elevation models and their derivatives to detect these features. We also incorporate the Convolutional Block Attention Module (CBAM) into RetinaNet for training and prediction. The precision of cirque detection with or without the addition of the CBAM is evaluated for various input data combinations, and training sample sizes, 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 % (p < 0.01), 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, respectively. The inclusion of CBAM also significantly reduces the number of undetected cirques. The model performance is affected by the quantity and quality of the training samples: the performance generally increases with increasing training samples and a training dataset of 6000 cirques produces the best results. This trained model can effectively detect boundary boxes that contain cirques to help facilitate subsequent cirque outline extraction and morphological analysis.