CPP-09. Mapping retrogressive thaw slumps in Alaska North Slope from ArcticDEM and Planet CubeSat imagery using machine learning

Abstract
Retrogressive thaw slumps (RTSs) are one of the most dynamic landforms resulting from the thawing of ice-rich permafrost. As reported by some local studies in Tibet and the Arctic, their occurrences and affected areas have increased dramatically in the past few decades. However, in most permafrost areas, their spatial distribution is still unknown due to the challenges of mapping them either from remote sensing data or in the field. Many RTSs, especially at their early developing stages, have small sizes in the area and only show up on very high-resolution (< 5 meters) imagery. Depending on the local environment, the appearance of RTSs may be very similar to the surroundings. To identify the RTSs in large areas and develop the corresponding automated pipeline, we choose Alaska North Slope (245,520 km2) as our study area and apply machine learning algorithms including convolutional neural networks and quick shift image segmentation to a 2-m digital elevation model (i.e. ArcticDEM) and 3-m optical imagery acquired by Planet CubeSats. Our current experiments show promising results in several sub-regions, and we will scale up to the entire Alaska North Slope for compressive algorithm evaluation and mapping exercise. In this presentation, we will present the technical details of our method and preliminary mapping results.