Large-Scale Mapping and Graph-Theoretic Characterization of Arctic Tundra Capillary Networks From Submeter Satellite Imagery

Michael Pimenta , Chandi Witharana, Amal Perera, Anna Liljedahl

Abstract— Tundra capillary networks (TCNs) are visible surface-drainage features associated with ice-wedge polygon terrain that can influence lateral surface-water redistribution across Arctic landscapes. However, TCN systems remain poorly characterized at regional scales because their narrow morphology, variable surface expression, and submeter scale have limited the development of scalable mapping and characterization methods. To address this gap, we present the first scalable GeoAI segmentation-to-graph framework for regional mapping and structural characterization of visible TCN expressions directly from very high spatial resolution satellite imagery. Using the first labeled remote-sensing dataset developed specifically for TCN segmentation, we implemented a systematic model-development workflow comparing convolutional neural network and transformer-based architectures across four spatial-context configurations and tuned SegFormer MiT-b3 with 1024 x 1024-pixel inputs as the highest-performing model (F1 = 0.89). Independent assessment at three Alaskan sites produced F1-scores ranging from 0.80 to 0.93, demonstrating transferability across diverse tundra landscapes. The selected model was deployed across a 728,400 km² northern Alaska WorldView-2/3 mosaic composed of 1,821 20 × 20-km sub-grid cells. The workflow mapped approximately 2.7 million km of TCN centerlines and converted the resulting detections into graph representations describing nodes, edges, connected components, and network extent. The resulting products provide the first regional baseline for characterizing the distribution and structural variability of visible TCN expressions across northern Alaska. Our framework demonstrates how submeter satellite imagery, deep learning, and graph theory can be combined to transform large Earth-observation archives into interpretable regional-scale products for Arctic landscape analysis.