Operational GeoAI for Scalable Satellite Analytics with Foundation Models on EuroHPC/MeluXina Supercomputer

About the Workshop

Recent progress in geospatial foundation models has accelerated Earth Observation (EO) research, yet many GeoAI methods are still evaluated at limited scale, obscuring issues of generalization, reproducibility, and operational readiness. This workshop provides a research-oriented but operationally realistic introduction to scalable GeoAI workflows, combining open EO foundation models with supercomputing resources such as EuroHPC/MeluXina.

Focusing on large-scale land cover segmentation from Sentinel-2 imagery, the workshop adopts a hybrid format. Participants first run lightweight inference of EO foundation models on their own laptops using small, pre-packaged data samples to build model intuition. The workshop then transitions to an instructor-led live demonstration on MeluXina, showing how the same workflows work at scale.

Two complementary open models are used: Prithvi-EO-2.0, representing temporal geospatial representation learning for downstream EO tasks, and TerraMind, illustrating multimodal and generative GeoAI capabilities. By explicitly linking local experimentation with HPC-scale execution, the workshop bridges GeoAI research and space-industry practice, while positioning supercomputing as an essential research instrument.

Note: this workshop will be provided under the EuroHPC-funded EPICURE project.

Learning Objectives

By the end of the workshop, participants will be able to:

• Recognize performance, reproducibility, and efficiency challenges in operational GeoAI
• Understand the role of geospatial foundation models in modern EO research
• Run lightweight EO foundation model inference locally to gain practical intuition
• Design an end-to-end GeoAI pipeline for land cover segmentation
• Understand how GeoAI workflows scale from laptops to HPC systems