Comparing Multiscale Object-Based Image Analysis and Deep Learning for High-Resolution Classification of Ice Wedge Polygon Tundra

Rising Arctic temperatures are making polygonal tundra increasingly vulnerable, primarily due to high ground ice contents. These landscapes form through soil-hydrology interactions, leading to ice wedge formation and degradation. Understanding the future of ice wedge polygon (IWP) landscapes requires detailed land cover classifications, as soil properties vary significantly across IWP sub-features like rims and centers. Existing classifications often distinguish between high-centered (HCP) and low-centered (LCP) polygons but fail to capture finer sub-feature distributions. This study provides high-resolution land cover datasets for two IWP sites on the Canadian Beaufort coast using WorldView-3 imagery. Ptarmigan Bay features well-defined landforms, while Komakuk Beach exhibits greater permafrost degradation. We compare two land cover-mapping approaches: object-based image analysis (OBIA) with segmentation and random forest classification, and a deep learning U-net model. Results show that the OBIA-random forest method performed better, and substantial differences in the landform type distribution between study areas and methods exist. Both methods identify 60% of HCP centers at Ptarmigan Bay and 50% at Komakuk Beach, but mapped IWP sub-feature proportions (HCP troughs, LCP centers, LCP rims) vary across areas and methods, reflecting classification uncertainties. Furthermore, the transferability of models between regions is constrained when there are pronounced differences in degradation of landforms.