Embed2Scale

Embed2Scale Introduces TerraCodec: Neural Compression for Optical Earth Observation Data

marioame — Mon, 16 Mar 2026 09:37:31 +0000

Embed2Scale Introduces TerraCodec: Neural Compression for Optical Earth Observation Data

The Embed2Scale project has released TerraCodec, an open-source family of neural compression models designed for optical Earth observation data.

Earth observation satellites continuously produce massive streams of multispectral and temporal imagery, creating growing challenges for storage, transmission, and large-scale analysis. TerraCodec addresses these challenges using neural compression models that encode multispectral imagery into compact latent representations and generate efficient bitstreams through learned entropy coding.

While most neural compression research has focused on RGB imagery, TerraCodec is designed specifically for multispectral satellite data. The model family includes efficient image compression baselines and temporal models that further improve compression by exploiting redundancy across satellite time series. All models are pretrained and evaluated at scale on the public SSL4EO-S12 v1.1 dataset and released open-source.

What TerraCodec enables

TerraCodec introduces neural compression models tailored for Earth observation workflows. The models allow users to:

Compress multispectral Sentinel-2 imagery using learned neural codecs
Exploit temporal redundancy in seasonal satellite-image sequences
Use flexible-rate compression with a single model checkpoint
Integrate compression directly into geospatial ML pipelines

TerraCodec achieves 3–10× higher compression than classical codecs such as JPEG2000 or WebP while maintaining comparable reconstruction quality. Beyond compression, the temporal models also enable applications such as cloud inpainting in satellite-image sequences, and we study how compression affects downstream Earth observation tasks.

Model family

The release includes models for different compression scenarios:

TEC-FP – compact factorised-prior image codec for efficient image compression
TEC-ELIC – improved entropy model with better rate–distortion performance
TEC-TT – temporal transformer codec for multispectral time-series data
FlexTEC – flexible-rate temporal model supporting multiple compression levels with one checkpoint

Open-source release

TerraCodec is released as an open-source library with pretrained checkpoints and example notebooks. Models can be used as standalone via PyPI or integrated into the TerraTorch model registry.

GitHub repository: https://github.com/IBM/TerraCodec
Pretrained models: https://huggingface.co/embed2scale
Paper: https://arxiv.org/abs/2510.12670

Towards scalable EO data pipelines

As Earth observation datasets continue to grow, neural compression methods like TerraCodec can help enable more efficient storage, transfer, and sharing of satellite data. By releasing models and code openly, the Embed2Scale project aims to accelerate research on compact representations for Earth observation.

The post Embed2Scale Introduces TerraCodec: Neural Compression for Optical Earth Observation Data appeared first on Embed2Scale.

Embed2Scale Consortium gathers in Ljubljana for General Assembly

marioame — Wed, 11 Mar 2026 12:41:48 +0000

Embed2Scale Consortium gathers in Ljubljana for General Assembly

The Embed2Scale consortium met in Ljubljana on 3 and 4 March for a two‑day general assembly hosted by project partner Sinergise, bringing partners together as the project moves through its third and final year.

Across the meeting, partners reviewed recent progress and aligned on the next steps for the months ahead. Advisory Board members joined both in person and online, contributing valuable perspectives on the positioning of Embed2Scale and the potential future use of its products.

A central focus of the general assembly was the project’s use cases. Dedicated breakout sessions explored the Maritime Awareness, Forest disturbance monitoring, and Crop Stress and Yield Early Detection use cases, each supported by input from the relevant technical work packages. These discussions helped consolidate the work required to bring all use cases to completion and prepare for upcoming demonstrations.

The consortium also worked on finalising dissemination, communication, and exploitation planning for the remainder of the year, ensuring a coordinated approach as the project enters its final phase.

The consortium will reconvene for a final in‑person meeting in the fall.

The post Embed2Scale Consortium gathers in Ljubljana for General Assembly appeared first on Embed2Scale.

Embed2Scale Introduces New TerraTorch Extension for EO Embedding Workflows

marioame — Mon, 02 Feb 2026 08:27:21 +0000

Embed2Scale Introduces New TerraTorch Extension for EO Embedding Workflows

The Embed2Scale project has launched a TerraTorch extension designed to enable Earth Observation (EO) embedding workflows. The extension adds low‑code and no‑code tools to generate, analyse, and reuse EO embeddings across a wide range of downstream tasks.

What the extension enables

In typical TerraTorch workflows, models are trained end‑to‑end on raw imagery. The new embedding workflows separate feature extraction from downstream training, making experimentation more flexible. They allow users to:

Use any supported dataset and backbone for embedding generation
Choose extraction layers and optional aggregation (Mean, Max, CLS)
Use embeddings in all existing TerraTorch downstream workflows
Test different decoder architectures
Compare different EO embeddings on the same task

Explore the demo notebooks

The extension is available through the open‑source TerraTorch library, with example notebooks demonstrating end‑to‑end workflows now published on GitHub: https://github.com/terrastackai/terratorch/tree/main/examples/embeddings

The notebooks provide practical examples, from generating embeddings to evaluating them on downstream tasks, and show how to adapt the workflow to new datasets.

The post Embed2Scale Introduces New TerraTorch Extension for EO Embedding Workflows appeared first on Embed2Scale.

Embed2Scale Expert to Present at AI for Good GeoAI Workshop

marioame — Mon, 12 Jan 2026 14:07:58 +0000

Embed2Scale Expert to Present at AI for Good GeoAI Workshop

On 25 February 2026, Isabelle Wittman (IBM) will speak at the AI for Good GeoAI Workshop in a session dedicated to embedding‑based workflows for Earth Observation (EO) applications.

She will introduce TerraTorch, an open‑source library designed to streamline the extraction and visualisation of EO embeddings, and NeuCo‑Bench, a benchmarking framework developed within Embed2Scale that enables rapid evaluation of embeddings across a wide range of EO tasks. Together, these tools demonstrate how embedding‑based approaches can significantly accelerate model training and inference, supporting more scalable and efficient EO workflows.

This workshop is the first in a series of online events running throughout 2026 that will highlight key outcomes of the Embed2Scale project, from NeuCo‑Bench to practical use cases across the EO domain.

The post Embed2Scale Expert to Present at AI for Good GeoAI Workshop appeared first on Embed2Scale.

Embed2Scale to Present at EUSPA AI Week on Advancing AI in Space Downstream Applications

marioame — Tue, 06 Jan 2026 09:59:33 +0000

Embed2Scale to Present at EUSPA AI Week on Advancing AI in Space Downstream Applications

The Embed2Scale project will take part in the “Advancing AI in Space Downstream Applications” session, held on 19 January 2026 as part of EUSPA AI Week. This online event brings together Horizon Europe projects to showcase how artificial intelligence is enhancing the space downstream services sector, with a particular focus on solutions for Copernicus and Earth observation(EO) applications.

During the session, experts will present innovative approaches to processing and scaling EO data, followed by an open discussion on future opportunities for AI‑driven applications in the space domain.

The session will take place from 10:15 to 11:30 CET, featuring presentations from three AI‑driven projects — Embed2Scale, DaFab and ThinkingEarth — followed by an open Q&A.

As part of the programme, Embed2Scale will be represented by Thomas Brunschwiler (IBM), who will deliver a 15‑minute presentation on the project’s latest advancements in Earth Observation and weather data federation using AI‑based embeddings.

The post Embed2Scale to Present at EUSPA AI Week on Advancing AI in Space Downstream Applications appeared first on Embed2Scale.

Embed2Scale partners at EurIPS 2025 REO Workshop

marioame — Tue, 30 Dec 2025 10:34:50 +0000

Embed2Scale partners at EurIPS 2025 REO Workshop

On December 7, 2025, Erik Scheurer and Gabriele Cavallaro from the Jülich Supercomputing Center attended the inaugural Representation Learning for Earth Observation (REO) workshop at EurIPS 2025 in Copenhagen, Denmark.

They presented the poster “LEPA: Learning Geometric Equivariance in Earth Observation with a Predictive Architecture” during the poster session slots between talks. Their work explores the geometric interpolation of embeddings and how to improve equivariance through a dataset pretraining strategy. The main goal behind this research is to address a practical challenge in deploying an API for downloading embeddings: how to align user and embedding data without repeated encoder passes for generating new embeddings.

Their contribution aligned with the workshop’s broader goal: advancing representation learning methods tailored to Earth observation, bridging machine learning innovation with environmental and climate applications.

Their attendance was an opportunity to showcase the latest Embed2Scale research to a highly engaged scientific audience, reinforcing the project’s role in shaping scalable and interpretable AI for Earth monitoring.

Read their contribution Have a look at their poster

The post Embed2Scale partners at EurIPS 2025 REO Workshop appeared first on Embed2Scale.

Embed2Scale at Big Data from Space 2025

marioame — Tue, 02 Dec 2025 16:00:41 +0000

Embed2Scale at Big Data from Space 2025

Early October, as part of the Big Data from Space 2025 conference, in Riga, Latvia, Embed2Scale partner IBM Research contributed a keynote by Thomas Brunschwiler titled “Tokens and Embeddings: The New Lingua Franca for Exploring Planet Earth.”

The presentation showcased the Embed2Scale consortium’s work on Earth Observation (EO) embedding workflows and benchmarking, alongside IBM’s advancements in multi‑modal EO foundation models. Brunschwiler emphasised how embeddings can streamline data workflows — compressing information, enabling federation, and accelerating exploration and model adaptation.

This work supports the Embed2Scale vision of transforming the scalability of EO data. By showcasing how embeddings unlock precision and efficiency across environmental applications, IBM Research reinforced the project’s mission to drive AI‑based knowledge discovery and impact across Europe.

The post Embed2Scale at Big Data from Space 2025 appeared first on Embed2Scale.

Embed2Scale Showcases Maritime Awareness Research at SPIE Security + Defence

marioame — Wed, 12 Nov 2025 14:04:27 +0000

Embed2Scale Showcases Maritime Awareness Research at SPIE Security + Defence

On September 17, 2025, Miguel A. Belenguer-Plomer, representing project partner SatCen, presented new research on maritime awareness at SPIE Security + Defence 2025.

His presentation, “Insights into deep learning-based vessel detection and characterization using SAR and AIS data”, introduced an approach that combines Sentinel‑1 radar imagery with AIS vessel records to detect and classify ships using a YOLO deep learning model. This work is part of Embed2Scale’s Maritime Awareness use case, which aims to strengthen Europe’s ability to monitor maritime activity.

The combined dataset included:

23,121 vessels
2,008 Sentinel-1 images (taken between 2020 and 2024 from 12 major US ports)
22,770 patches (500×500 px)

Key observations from the research:

YOLOv8, Sentinel‑1 and AIS together enable vessel detection and classification.
The medium YOLO model offers the best balance between accuracy and efficiency.
The model shows strong performance with cargo and service vessels.
Fishing, leisure, and military vessels were harder to detect due to data imbalance and resolution limits.
Data augmentation, class-weighted loss functions, and longer training would improve the monitoring process.

The post Embed2Scale Showcases Maritime Awareness Research at SPIE Security + Defence appeared first on Embed2Scale.

Introducing NeuCo-Bench: Benchmarking Compact Earth Observation Embeddings with Purpose

marioame — Wed, 25 Jun 2025 07:28:05 +0000

Introducing NeuCo-Bench: Benchmarking Compact Earth Observation Embeddings with Purpose

We are pleased to announce the release of NeuCo-Bench, a lightweight, task-driven benchmarking framework for evaluating compact Earth Observation (EO) embeddings. Originally developed to support submissions for the EARTHVISION 2025 Challenge at CVPR, NeuCo-Bench is now available as a standalone toolkit for local evaluation and research use.

What is NeuCo-Bench?
NeuCo-Bench is a model-agnostic evaluation framework designed to test how well compressed, fixed-size EO embeddings preserve task-relevant semantics. In contrast to traditional pixel-level reconstruction metrics, NeuCo-Bench focuses on downstream utility—measuring what embeddings can do in real-world EO tasks.

Key Features

Compact and Efficient: Evaluates fixed-size embeddings (e.g. 1024-dimensional vectors) suitable for resource-constrained pipelines.
Task-Driven Evaluation: Applies linear probes to practical EO tasks including land-cover classification, cloud detection, and biomass estimation.
Model-Agnostic Comparison: Supports embeddings from foundation models, compression techniques, or handcrafted baselines.
Robust Metrics: Includes signal-to-noise ratios and dynamic rank aggregation for comprehensive benchmarking.

Why NeuCo-Bench?
As EO applications move toward real-time and edge processing, embedding efficiency matters. NeuCo-Bench helps researchers and practitioners systematically evaluate semantic retention under compression and standardize comparison across approaches.

Useful link
Explore the repo and get started (with example scripts): github.com/embed2scale/NeuCo-Bench

We welcome your feedback and contributions to accelerate EO embedding research together!

CVPR Challenge: https://www.grss-ieee.org/events/earthvision-2025/?tab=challenge
NeuCo-Bench Repo: https://github.com/embed2scale/NeuCo-Bench

The post Introducing NeuCo-Bench: Benchmarking Compact Earth Observation Embeddings with Purpose appeared first on Embed2Scale.

Embed2Scale to Present AI-Driven Earth Observation Research at ESA’s Living Planet Symposium 2025

marioame — Mon, 23 Jun 2025 07:32:48 +0000

Embed2Scale to Present AI-Driven Earth Observation Research at ESA’s Living Planet Symposium 2025

We’re proud to announce that the Embed2Scale project will be strongly represented at the upcoming ESA Living Planet Symposium 2025 (LPS25), the international event gathering the Earth Observation community to advance the transition from satellite-based observation to climate action and sustainable innovation.

Dr. Conrad Albrecht from the German Aerospace Center (DLR) will present a series of contributions developed within the project’s research activities, focused on advancing AI-based data processing techniques for satellite imagery.

June 25

Spaceborne SAR Compression with AI for Data-Efficient Vessel Detection

Session: D.02.05 – EO-based solutions to address civil security-related scenarios

June 26

Spectral Super-Resolution for Greenhouse Gas Detection
Self-Supervised Learning for Multispectral and Hyperspectral Remote Sensing (Poster)
Towards Efficient Neural Compression for Earth Observation Data (Poster)

Session: D.02.06 – Foundation Models for Earth Observation: Current solutions with less labelled data to improve environment monitoring and future perspectives to revolutionize geospatial data discovery and utilization

June 27

Self-Supervised Learning for Earth Observation (SSL4EO) Put to the Test – a perspective through the lens of two Horizon Europe projects

Session: D.05.04 – Digital Copernicus – Shaping Europe’s Digital Future in the Age of AI, 08:30–10:00 am, Hall M1/M2
Theme: Copernicus Operations & Long-term
Chair: Anna Burzykowska (European Space Agency)
This high-level invited session explores how AI can revolutionise data processing and access in the Copernicus ecosystem, dramatically reducing barriers to EO data usage. Dr. Albrecht’s talk will provide insight into how Embed2Scale and sister Horizon Europe projects are pioneering this transformation using foundation models and self-supervised learning.

All these presentations underscore Embed2Scale’s commitment to scalable, AI-powered techniques that enhance data accessibility, compression, and interpretation across key environmental domains, from maritime surveillance to greenhouse gas monitoring.

We look forward to engaging with researchers, policymakers, and innovators at LPS25 to share insights and foster collaborations that push the boundaries of Earth Observation technology.

More about the event: LPS25 Official Website

The post Embed2Scale to Present AI-Driven Earth Observation Research at ESA’s Living Planet Symposium 2025 appeared first on Embed2Scale.