Satellite-Based Anomaly Detection using GeoAI Embeddings: A Scalable Workflow (Marcin Kluczek)

The Open-Earth Monitor Global Workshop 2026 "Next-Generation EO & GeoAI Built on Openness and Distributed Data" 7-9 October 2026, Barcelona, Spain https://earthmonitor.org/global-workshop-2026/ This study explores the application of high-dimensional embeddings derived from Sentinel-2 imagery for automated anomaly detection in environmental monitoring. By utilizing the SSL4EO self-supervised learning framework, we transform raw satellite data into compact, informative representations that capture essential spatial and temporal features. The entire workflow is integrated within the Copernicus Data Space Ecosystem (CDSE), ensuring efficient access to large-scale, analysis-ready archives and enabling rapid processing of planetary-scale datasets. The core of our approach lies in leveraging SSL4EO to bypass the need for massive labeled datasets, creating a standardized intermediate format that bridges raw imagery with advanced AI tasks. These embeddings serve as the foundation for an anomaly detection pipeline designed to pinpoint deviations from expected seasonal or spatial trends, such as flooding, wildfires, or shifts in vegetation health. To ensure interoperability and ease of use in geospatial analytics, results are stored in the GeoParquet format, which supports both reproducibility and high-performance data handling. To confirm the framework's robustness, we conducted extensive validation across diverse geographical regions and seasonal cycles, including challenging winter conditions with snow cover and low solar illumination. The pipeline demonstrated high resilience, producing consistent embeddings even in the presence of partial cloud cover. Furthermore, we evaluated the system’s portability across heterogeneous computing environments. Testing on the CREODIAS cloud platform (using both CPU and GPU nodes) alongside high-performance computing (HPC) infrastructures like EOHPC and SpaceHPC proved that the solution scales effectively and maintains functional integrity across different hardware architectures. The results indicate that combining self-supervised embeddings with anomaly detection creates a powerful tool for environmental intelligence. The proposed framework is suitable for a wide range of operational applications, from tracking urban growth to monitoring climate-induced environmental changes. By providing a portable and scalable bridge between raw Copernicus data and actionable insights, this study highlights the transformative potential of GeoAI in supporting global and regional decision-making processes.