Embed2Scale research paper presented at EGU General Assembly 2024

Hours

The full paper, detailing the NEC methodology and its implications for future EO systems, can be accessed here:

We are thrilled to announce an initial milestone in the Embed2Scale project with the presentation of a research paper at the prestigious EGU General Assembly 2024. Authored by Carlos Gomes and Thomas Brunschwiler from IBM Research – Europe, the paper titled “Neural Embedding Compression for Efficient Multi-Task Earth Observation Modelling” introduces an innovative approach to handling large-scale Earth Observation (EO) data.

The research addresses critical challenges in EO data management by leveraging Neural Embedding Compression (NEC). This technique significantly reduces the data volume and computational overhead associated with traditional methods, enabling more efficient storage, transmission, and processing of multi-task EO models. The paper highlights the deployment of foundation model techniques, achieving compression ratios of up to 1000x while maintaining high utility for various downstream tasks.

Presented in a session focused on scalable machine learning deployments on High-Performance Computing (HPC) infrastructure, the research was well-received.
This work not only supports Embed2Scale’s vision of enhancing geospatial data accessibility and efficiency but also sets a new benchmark in the integration of AI in Earth observation.

Join us in celebrating this remarkable achievement and stay tuned for more updates as we continue to push the boundaries of geospatial data technology.