Accuracy, Efficiency, and Transferability of a Deep Learning Model for Mapping Retrogressive Thaw Slumps across the Canadian Arctic

Deep learning has been used for mapping retrogressive thaw slumps and other periglacial landforms but its application is still limited to local study areas. To understand the accuracy, efficiency, and transferability of a deep learning model (i.e., DeepLabv3+) when applied to large areas or multiple regions, we conducted several experiments using training data from three different regions across the Canadian Arctic. To overcome the main challenge of transferability, we used a generative adversarial network (GAN) called CycleGAN to produce new training data in an attempt to improve transferability. The results show that (1) data augmentation can improve the accuracy of the deep learning model but does not guarantee transferability, (2) it is necessary to choose a good combination of hyper-parameters (e.g., backbones and learning rate) to achieve an optimal trade-off between accuracy and efficiency, and (3) a GAN can significantly improve the transferability if the variation between source and target is dominated by color or general texture. Our results suggest that future mapping of retrogressive thaw slumps should prioritize the collection of training data from regions where a GAN cannot improve the transferability.