Leveraging computer vision on Earth observation data for quantifying Arctic permafrost landscape characteristics

In the rapidly warming Arctic, permafrost ecosystems are undergoing profound transformations, impacting surrounding ecosystems, communities, and infrastructure, while thaw-related geomorphic changes alter hydrological processes within the landscapes. The complex interplay between frozen ground and surface water in permafrost regions significantly influences water movement, storage, and distribution. A particular role is ascribed to the small water bodies, such as trough ponds in landscapes characterized by ice-wedge networks, as their distribution and connectivities play a crucial role not just in local-scale interactions but in biogeochemical cycling with regional to global-scale implications. To comprehensively grasp these complex interactions, monitoring and modeling the hydrology and topography of permafrost regions is of high importance. Given that permafrost spans extensive regions across the Arctic, Earth observation emerges as a valuable tool for comprehensive assessments. Satellite remote sensing data offers broad temporal and spatial coverage, but may lack detailed spatial resolution. Aerial imagery, while providing superior spatial detail, is often limited in spatial and temporal coverage due to its high acquisition cost. The first study in this thesis introduces a new extensive aerial imagery and point cloud dataset of Alaskan and Canadian permafrost landscapes that is used in the subsequent analyses. The following studies detail data science methods that bridge the gap from remotely-sensed local-scale observations to a pan-Arctic extent, facilitating efficient upscaling from small areas of interest to large datasets and broad regions. Central to this research is the introduction of cross-resolution U-Nets, a novel segmentation network that processes lower spatial resolution aerial images, generating segmented maps of surface water bodies with enhanced spatial detail. Applying this methodology to imagery at resolutions comparable to PlanetScope or Sentinel-2 data, enables the mapping of surface water bodies, even the smallest trough ponds, at finer detail across vast regions beyond the coverage of aerial datasets. The focus then narrows to polygonal-patterned ground, where even minor changes in a warming permafrost environment can significantly alter surface hydrological regimes. In particular, I look at the evolution of polygonal networks in recent fire scars, where the abrupt disturbance has triggered rapid permafrost thaw and thus the evolution of a highly dynamic network of polygon troughs. Here, computer vision methods utilizing high-resolution aerial data are introduced to model these patterns and their characteristics as graphs, providing a quantitative assessment of ice-wedge polygon patterns, their evolution over time, and the implication to the changing landscape’s surface hydrology. Through a space-for-time-substitution remote sensing analysis with high-resolution elevation models on recent and historical fire scars in Western Alaska, the subsequent study reveals that despite severe wildfires in the study areas, the degradation of ice-wedge polygons and the associated hydrological networks can stabilize after a few years post-disturbance. For each study, I show how automation and efficient upscaling of these approaches is achieved through computational optimization, such as algorithm parallelism, or the use of scientific workflow managers, facilitating analysis of the growing data archives anticipated in the upcoming years. In conclusion, this thesis lays a crucial data science-driven foundation for quantitatively evaluating surface hydrology in permafrost landscapes from high-resolution remote sensing imagery, by bridging detailed local-scale observations to broad pan-Arctic spatial scales.