Seeing Beyond Visible Ice: AI-Powered Detection and Segmentation of Rock Glaciers
Research Poster Engineering 2025 Graduate ExhibitionPresentation by Nicholas Kraabel
Exhibition Number 134
Abstract
Rock glaciers are vital cryospheric features that serve as physical manifestations of permafrost, influencing localized geohazards like landslides while offering insights into global climatic processes. Despite their importance, traditional satellite reflectance methods fail to detect these landforms, necessitating the reliance on time-consuming manual interpretation for inventory creation, a subjective process that limits their integration into climate models compared to readily-observable exposed ice features. This project introduces an AI-driven solution to automate rock glacier detection and analysis, combining high-resolution satellite imagery, digital elevation models, and glacier masks from Alaska Chugach Mountain range with a novel Segment Anything with In-Context Spatial Prompt Engineering model. The framework specifically targets geomorphological characteristics, such as surface texture, sediment patterns, and kinematic signatures, to distinguish functional permafrost-bearing rock glaciers from inherited relict forms. Although work on data collection and model training is still ongoing, our eventual goal is for the resulting inventories to bridge a persistent gap in geomorphological mapping. This will offer a scalable tool to delineate permafrost boundaries, validate climate models, and quantify water resources in Arctic catchments. Applied initially to southern Alaska's mountain ranges, the methodology has potential for global adaptation. We have particular interest in current un-mapped regions where rock glaciers may store water volumes comparable to traditional glaciers. We hope that this approach helps transform rock glaciers from niche indicators into central components of cryospheric monitoring.
Importance
Current climate models overlook rock glaciers despite their critical role in mountain permafrost systems and as indicators of climate change. These landforms remain unmapped in many regions because traditional satellite methods relying on reflectance can't detect them, forcing scientists to rely on time-consuming manual interpretation. Our project aims to develop an AI system that automatically identifies rock glaciers by analyzing distinctive surface patterns and movement signatures in satellite imagery and elevation data from Alaska's Chugach Mountains. The goal is for resulting inventories to help scientists define permafrost boundaries, validate climate models, and quantify water resources in Arctic regions. Our methodology is designed to easily expand to other regions.