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

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