EO-based rock glacier mapping and characterisation

Rock glaciers are tongue-shaped complex landforms that indicate present or past permafrost conditions and are commonly found in many high-latitude and/or high-elevation environments. They consist of poorly sorted angular debris and ice-rich sediments formed by gravity-driven creep. Information about the location, extent and characteristics of rock glaciers is important for several reasons, for example, for estimating their hydrological significance as water resource (e.g., for alpine huts) and the geohazard potential because of rock glacier destabilisation due to climate change. While other cryosphere components, such as snow and glaciers, are spectrally distinct from the surrounding terrain, rock glaciers are spectrally inseparable and, as such, difficult to automatically detect and delineate from Earth observation (EO) data. Thus, rock glaciers are usually mapped through laborious subjective manual interpretation of EO data. This leads to inhomogeneous, incomplete, and inconsistent mapping with large differences among different regions. Therefore, there is a need for objective, automated, and efficient mapping of rock glaciers using globally applicable satellite datasets, such as Sentinel-1 and Sentinel-2.

Modern machine learning methods, such as deep learning (DL), offer new possibilities for automating mapping tasks. DL works by recognising recurring patterns and textures through an artificial neural network. However, very limited work has been done on rock glacier mapping, and there is a lack of consensus regarding the best parameters for this purpose. Moreover, features that share similar surface textures with rock glaciers, such as landslides and avalanche or fluvial deposits, can be misclassified by DL. Therefore, there is a need to conduct a thorough investigation of the DL model architectures and input data types that produce the best results for mapping rock glaciers.

The overall goal of ROGER is to reliably map and characterise rock glaciers using optical and synthetic aperture radar (SAR) EO data. We will assess the performance, robustness, and reliability of DL models for automated EO-based rock glacier mapping in study areas in Austria and Svalbard, Norway, and quantify the accuracy of the results in comparison with reference data. Moreover, we will derive velocity rates of the identified rock glaciers using differential synthetic aperture radar interferometry (DInSAR) and classify them according to their activity status. ROGER represents an important contribution to the field of cryospheric research by evaluating methods for the automated identification and characterisation of rock glaciers and expand our knowledge of the potential of DL to efficiently map complex natural phenomena using EO data. The project findings will contribute to increasing the trustworthiness of DL methods, which is of high importance for many applications and especially when communicating and explaining results to stakeholders and decision makers.