Earth Observation & Weather Data Federation with AI Embeddings

Informazioni relative al progetto

Embed2Scale

ID dell’accordo di sovvenzione: 101131841

DOI

10.3030/101131841

Data della firma CE 11 Dicembre 2023

Data di avvio 1 Gennaio 2024

Data di completamento 31 Dicembre 2026

Finanziato da

• Digital, Industry and Space

Costo totale

€ 2 618 047,50

Contributo UE

€ 2 618 047,50

Coordinato da

MARTEL INNOVATE BV
Netherlands

Periodic Reporting for period 1 - Embed2Scale (Earth Observation & Weather Data Federation with AI Embeddings)

Periodo di rendicontazione: 2024-01-01 al 2025-06-30

Sintesi del contesto e degli obiettivi generali del progetto

Embed2Scale (E2S) successfully delved into the world of neural compression (NC) for Earth observation (EO) and weather data (WD). The petabyte-scale of spatio-temporal data generated by EO&WD from multiple sensors and hydrodynamic simulation requires a tool to comprehend the information contents impossible to digest by human inspection. With the aid of self-supervised artificial neural networks (Foundation Models) that generate embeddings, we opened the door to efficiently access geo-information.

Lavoro eseguito dall’inizio del progetto fino alla fine del periodo coperto dalla relazione e principali risultati finora ottenuti

* WP 1:
- Draft extension of openEO API to support neural compression endpoints.
- Analysing possible impact of transforming PAZ SAR data into ARD format.

* WP 2:
- Published an IEEE GRSM review on neural compression for Earth observation, catalyzing new use‐case development.
- Launched NeuCo-Bench at CVPR 2025 as an open-source benchmarking framework, engaging 50+ international teams.
- Open-sourced up to 7000× lossy compression methods on multi-sensor imagery.

* WP 3:
- Designed an end-to-end workflow and federated deployment pipeline for model training, compression, and inference.

* WP 4:
- Co-developed four prototypes in biodiversity, cloud characterization, precision agriculture, and maritime awareness.
- Released SSL4EO-S12-downstream datasets on HuggingFace
- Provided some PAZ SAR and AIS sample data to train vessel detection.

Progressi oltre lo stato dell’arte e potenziale impatto previsto (incluso l’impatto socioeconomico e le implicazioni sociali più ampie del progetto fino ad ora)

- First dedicated comprehensive review linking lossy neural compression to advance real-world Earth-observation tasks.
- NeuCo-Bench: the inaugural, task-centric EO compression benchmark standardizing metrics and datasets including setup of the community "Earth2Vec" with 32 members (and counting) to define embedding interoperability and best practices.
- SSL4EO-S12-downstream: the first public suite coupling compression evaluation with 12 diverse EO tasks on which the "Earth2Vec" community will build on.
- Pioneered an openEO API extension for on-demand neural compression, enabling seamless integration in EO workflows across data centers.