Embed2Scale’s EGU 2025 Session Merges into a Broader Session on Foundation Models for Earth and Climate

Hours

The Embed2Scale-organized original session at EGU General Assembly 2025 has been merged with another session to form a broader discussion on Foundation Models in Earth Observation and Climate Science. The newly combined session, titled “Advancing Foundation Models for Earth and Climate: Benchmarking, Best Practices and AI-Enabled Scientific Understanding” will take place on Monday 28 April 2025, at 08:30–10:15 (CEST) in Room -2.92.

The session will cover key aspects revolving around GeoFoundation Models (GeoFMs), including:

  • Benchmarks & Evaluation – Defining standardized metrics and benchmarks to ensure fair and reliable assessment of GeoFM performance.
  • Pre-Training Strategies & Best Practices – Exploring efficient data sampling, proxy tasks, and scalable training approaches for pre-training and application of GeoFMs.
  • Sensor Independence – Leveraging GeoFMs to integrate data from multiple sensors for a holistic understanding of Earth’s dynamics.
  • Multi-Modal & Temporal Analysis – Advancing multi-modal data integration and spatio-temporal change detection.
  • Scientific Insights – Investigating the role of GeoFMs in uncovering geo-physical principles and causal relationships.
  • Community Involvement & Impact – Building an open-science community for GeoFMs while considering their societal, environmental, and economic implications.

The Embed2Scale consortium remains actively involved in shaping the session’s discussions, contributing insights into AI-driven embedding techniques and compression strategies for geospatial data.

The session has also been awarded the EGU’s Equality, Diversity, and Inclusion (EDI) logo, recognizing its commitment to fostering an inclusive and diverse scientific community. This distinction highlights the session’s dedication to ensuring equitable participation and representation across different backgrounds, career stages, and perspectives within the Earth and climate science domains.

ESSI1.9 | Session description

Geospatial Foundation Models (GeoFMs) have shown great promise in a wide range of applications for Earth Observation (EO) and Earth System Modelling (ESM), as well as for Weather and Climate. With the increasing number of models being published, model inter-comparison is key to identify the best GeoFM for deployment. This session aims to highlight efforts on model development, benchmarking, fine-tuning, and their best practices for utilizing GeoFMs in real-world applications. We invite submissions focused on creating GeoFMs to leverage multi-modal, multi-temporal, and multi-resolution datasets towards sensor-independence. Diverse FMs for EO, ESM, Weather, and Climate can revolutionize data analysis by handling text, imagery, and time-series, enabling insights into natural hazards and climate resilience. Our session will cover advances in data curation, model architecture, scaling, benchmarking, pretraining, fine-tuning, and MLOps for GeoFMs, including use cases and deployment strategies.

The topics of our session revolving around GeoFMs are:

  1. Benchmarks & Evaluation: Establish standardized fair evaluation metrics and benchmarks to assess the performance and capabilities of GeoFMs in multi-modal data analysis, ensuring reliability and efficiency.
  2. Pre-Training Strategies & Best Practices: Discuss efficient data sampling strategies, proxy tasks, and scalable model training for efficient pre-training of GeoFMs. Guidelines for using existing pre-trained GeoFMs for a diverse set of applications, with focus on how to decide which models are best for certain use cases.
  3. Sensor Independence: GeoFMs can process data from various sensors, enabling comprehensive analysis of the Earth’s dynamics holistically.
  4. Multi-Modal/Temporal: GeoFMs offer novel approaches to multi-modal data analysis and spatio-temporal change detection.
  5. Scientific Insights: Highlighting the scientific insights enabled through the creation of GeoFMs, particularly in relation to geo-physical principles and causal relations.
  6. Community Involvement & Impact: How to build an open-science community around GeoFMs that is easily accessible to all while keeping an eye on potential societal, environmental, and economic impacts when deploying GeoFMs.

We aim to foster discussions on current applications, challenges, and opportunities of GeoFMs seeking contributions from AI and domain researchers, climate modelers, industry experts, and stakeholders in AI, HPC, and Big Data.

For more details, please visit the official EGU session page.