CPP-07. Feature extraction of ArcticDEM and its application on mapping permafrost disturbance
We processed more than 200 TB multi-temporal elevation data (i.e. ArcticDEM) and used them to locate permafrost disturbance in the Arctic. ArcticDEM is a high resolution (2 m) digital surface model derived from optical stereo imagery that covers all land areas north of 60°N. We used both the strip and mosaic version of ArcticDEM for feature extraction including (1) calculating elevation difference by comparing the most recent digital surface model (DSM) and the oldest ones, (2) segmenting subsidence regions from elevation difference using the quick shift algorithm, and (3) extracting headwall lines that represent the sudden slope changes using thresholding and medial axis. We built an automated pipeline which utilized both workstations in our lab and a supercomputer (i.e. RMACC Summit) to download and process ArcticDEM grid by grid. We composed images using a hillshade calculated from the most recent DSM and the headwall lines for locating retrogressive thaw slumps, which is a type of permafrost disturbance resulting from the thawing of ice-rich permafrost. We used an object detection algorithm called YOLOv4 (You only look once) to identify thaw slumps. We combined YOLOv4’s output and polygons of the subsidence regions to remove false positives. In this presentation, we will present the technical details of the automated pipeline for data processing and preliminary mapping results in Alaska.