How Embeddings Are Changing Earth Observation Workflows

In late May, members of the Embed2Scale consortium from IBM Research and UZH attended the ESA–NASA Workshop on Foundation Models for Earth Observation. While foundation models were naturally central to the agenda, one theme surfaced repeatedly across talks, discussions, and hallway conversations: geospatial embeddings have become a major topic in the EO community. What stood out most was the strong interest from practitioners. Across research institutions, government agencies, startups, and industry, many people are exploring how AI can support their geospatial workflows. Embedding-based approaches are particularly appealing because they offer both ease of use and computational efficiency. The Embed2Scale tutorial on embedding-based EO workflows confirmed this interest, sparking extensive discussions around practical adoption, tooling, and real-world use cases. 

Despite this momentum, a gap remains between research advances and operational use. The rapidly expanding landscape of foundation models, embedding products, and benchmark results can make it challenging for practitioners to determine which approaches best suit their data, how different products compare, and what adaptation is required for a specific application. 

Bridging Research and Practice

As a result, educational resources, transparent evaluation, and practical tooling are becoming just as important as developing new models. A recurring question throughout the workshop was: where do embeddings already provide clear value today? While interest in embeddings is high, many practitioners are still looking for concrete examples that demonstrate their benefits in real workflows. Sharing successful applications, as well as lessons learned and remaining limitations, will be critical for broader adoption. The Embed2Scale use case Forest Biomass Disturbance Monitoring, presented by UZH, sparked exactly these kinds of conversations. 

The workshop also highlighted an important trade-off between accessibility and flexibility. Precomputed embedding products can significantly lower the barrier to entry, yet many use cases require specific spatial resolutions, temporal granularity, modalities, or update cycles. Supporting both easy access and customisation remains a challenge for the community. Open-source tools such as TerraTorch aim to help bridge this gap by making it easier to generate and work with embeddings tailored to specific data and workflows. 

Interoperability emerged as a closely related concern. As new embedding products continue to appear, maintaining compatibility, discoverability, and usability over time becomes increasingly important. While individual models and products will inevitably be replaced by newer approaches, shared standards, common tooling, and community efforts can help ensure that embeddings remain accessible and useful beyond the lifetime of any single release. 

In conclusion, Romeo Kienzler and Isabelle Wittmann left Alabama encouraged by the momentum around EO embeddings and confident that the next phase of progress will depend not only on better models, but also on better tools, benchmarks, standards, educational resources, and community collaboration.