NeuCo‑Bench: A Step Toward Standardised Benchmarking for Earth Observation Embeddings

At the CVPR EarthVision Workshop 2026, and soon at IGARSS 2026, the Embed2Scale consortium presented NeuCo‑Bench, a model‑agnostic framework designed to address the often fragmented benchmarking landscape of Geospatial Foundation Models and other embedding approaches in Earth Observation (EO). While the EO community increasingly adopts embedding‑based methods for their efficiency and versatility, the lack of standardised evaluation has made it difficult to compare results fairly and reproducibly. NeuCo‑Bench tackles this challenge by providing a benchmarking framework that anyone can use to perform reproducible, standardised benchmarking, regardless of the embedding method or model.

The framework treats encoders as black boxes, focusing solely on how well embeddings perform on downstream tasks. This ensures that evaluations are transparent, fair, and independent of the specific model or training process. By using linear probing and a scoring system that balances accuracy and stability, NeuCo‑Bench offers a robust foundation for comparing embeddings in a consistent way.

Beyond standardisation, NeuCo‑Bench is designed to evolve with and by the community. Its open‑source nature invites researchers and practitioners to contribute new tasks, datasets, and evaluation methods, ensuring the framework adapts to emerging needs and use cases. This collaborative approach not only fosters innovation but also helps maintain the relevance and usability of the benchmark as the EO field advances. By providing a shared, extensible tool, NeuCo‑Bench supports the community in building a sustainable ecosystem for embedding evaluation, one that grows alongside the rapidly developing landscape of EO applications.

As embedding‑based workflows become more central to EO, frameworks like NeuCo‑Bench play a crucial role in ensuring that advancements are built on a solid, comparable foundation. This paves the way for more reliable research, better‑informed adoption of new methods, and ultimately, greater impact in real‑world applications.