Rock Glacier Detection Using Satellite Embeddings and Machine Learning on Google Earth Engine

Rock glaciers serve as vital indicators and reservoirs within mountain permafrost systems, but their mapping is challenging due to complex terrain and limited labelled data. In this study, we present a novel automated detection method that combines Google DeepMind’s pre-trained 64-dimensional satellite embeddings with machine learning classifiers within Google Earth Engine. Using expert-labelled and partially ground-truthed polygons, we evaluated classifiers including Random Forest, k-Nearest Neighbours, Gradient Tree Boosting, and Support Vector Machine over an area of 3900 Km2 in the Southwestern Maritime Alps. The Support Vector Machine achieved the highest accuracy at 85.7% and demonstrated balanced performance with an F1-score of 0.80. The resulting rock glacier classified raster pixels covers 32 km², with a post-processing inventory of 621 rock glaciers, consistent with regional knowledge and field validation. The embedding-based Support Vector Machine minimises the need for extensive feature engineering and large training datasets. It offers a data-efficient, reproducible framework for regional rock-glacier screening, with broader transferability to be validated through specific cross-region tests. This open-source method supplies an operational tool to enhance permafrost monitoring and climate change impact evaluations.