Automated Mapping of Rock Glaciers Using Image Segmentation with Deep Learning

Rock glaciers are critical landforms in periglacial environments. They play a significant role in regional hydrology and provide valuable insights into climate and geomorphological processes. Mapping rock glacier extent is an important step for quantifying their hydrologic and geomorphic role in the landscape, but this process is labor intensive. To automate the process of mapping rock glaciers in the western U.S. (total area ~ 30000 km2), we present a methodological framework that relies on a combination of Google Earth Engine and TensorFlow cloud computing. Using existing rock glacier inventories, we trained a Compact Residual U-Net Convolutional Neural Network (CNN) that uses 14 input bands, including Sentinel-2 optical imagery, USGS elevation models, Sentinel-1 backscatter SAR imagery, and Landsat 8 thermal imagery. The model was trained across 5 US states (Utah, Colorado, Wyoming, Idaho, Montana) which have different rock types and climates. With 2597 rock glacier outlines from the Portland State University Active Rock Glacier Inventory used for training, the model achieved a moderate Intersection over Union (IoU) of 0.495 when tested on a new dataset. Precision and recall values were 0.735 and 0.602, respectively. The model successfully mapped 206 out of 290 rock glaciers as well as 41 false positives and 84 false negatives in the eastern Uinta Mountains across an area of 3037 km2. The model struggled to map slower-moving rock glaciers, which are more geomorphologically subtle. Our research advances the application of machine learning in rock glacier mapping, offering a high-dimensional method for mapping rock glaciers, which will ultimately enhance our understanding of these important landforms in a changing climate.