Cryosphere and Polar Science Seminar
Mapping thermokarst landforms from remote sensing imagery with deep learning by Lin Liu, The Chinese University of Hong Kong
The number and extent of thermokarst landforms in permafrost areas have increased in recent decades. However, their distribution and temporal changes, especially the non-lake ones on the Tibetan Plateau, are poorly understood or quantified. We developed a new strategy to utilize deep learning in the processing of remote sensing imagery. Applying this method to CubeSats optical images over central Tibet, we delineated 196 retrogressive thaw slumps, with an F1 score of 0.829. By collecting and incorporating more training data, this method can be potentially extended to a large area and help to determine the vulnerability of permafrost landforms to warming.