Permafrost thaw sensitivity prediction using surficial geology, topography, and remote-sensing imagery: a data-driven neural network approach

• Published in: Canadian Journal of Earth Sciences
• Main Authors: Oldenborger, Greg A., Short, Naomi, LeBlanc, Anne-Marie
• Format: Article in Journal/Newspaper
• Language: English
• Published: Canadian Science Publishing 2022
• Subjects: General Earth and Planetary Sciences; Arctic; Nunavut; Canada; Rankin Inlet; Ice; permafrost
• Online Access: http://dx.doi.org/10.1139/cjes-2021-0117; https://cdnsciencepub.com/doi/full-xml/10.1139/cjes-2021-0117; https://cdnsciencepub.com/doi/pdf/10.1139/cjes-2021-0117

Seasonal or degradational thaw subsidence of permafrost terrain affects the landscape, hydrology, and sustainability of permafrost as an engineering substrate. We perform permafrost thaw sensitivity prediction via supervised classification of a feature set consisting of geological, topographic, and multispectral variables over continuous permafrost near Rankin Inlet, Nunavut, Canada. We build a reference classification of thaw sensitivity using process-based categorization of seasonal subsidence as measured from differential interferometric synthetic aperture radar whereby categories of thaw sensitivity are reflective of ground ice conditions. Classification is performed using a neural network trained on both dispersed and parcel-based reference data. For Low, Medium, High, and Very High thaw sensitivity categories, generalized classification accuracy is 70.8% for 20.6 km 2 of dispersed training data. In all cases, the majority classes of Low and Medium thaw sensitivity are predicted with higher accuracy and more certainty, while the minority classes of High and Very High thaw sensitivity are underpredicted. Minority classes can be combined to improve accuracy at the expense of a reduced level of discrimination. The two-class problem can be classified with an accuracy of 81.8%, thereby effectively distinguishing between stable and unstable ground. The method is applicable to similar Low-Arctic permafrost terrain with geological and topographical controls on thaw sensitivity. However, generalized accuracy is reduced for parcel-based training, indicating that reference samples are not totally representative for inference beyond the parcel, and any deployment of the network to other geographical regions would benefit from full or partial retraining with local data.