As an Amazon Associate we earn from qualifying purchases.
Key Takeaways
- TerraMind converts Sentinel observations into reusable numerical representations.
- Embeddings could reduce the cost of building specialized Earth observation products.
- Production value depends on validation, efficient computing and accountable model use.
A Copernicus Example Makes Embeddings More Accessible
On September 25, 2026, the Copernicus Data Space published a TerraMind openEO example. The notebook shows how to combine Sentinel-1 radar and Sentinel-2 optical observations, prepare them in the cloud and run the TerraMind 1.0 model.
The workflow loads the satellite collections, masks unsuitable Sentinel-2 pixels, builds a temporal composite and aligns the sensors in a multimodal data cube. TerraMind then processes neighborhoods of 224 by 224 pixels and produces a 768-dimensional embedding for each patch.
An embedding is a numerical representation that preserves patterns useful to machine-learning systems. Areas with related physical or visual characteristics may occupy nearby positions in the embedding space. A developer can use those representations for classification, similarity search or change detection without engineering every feature from raw imagery.
The notebook uses K-means clustering to test whether the features capture meaningful land-cover structure. It returns the embeddings as a reusable NetCDF product that can feed later processing.
Copernicus describes the example as a proof of concept rather than an optimized production implementation. That distinction matters. The notebook proves that the components can work together, but commercial delivery requires faster inference, quality controls and predictable computing costs.
Embeddings Move Value Away From Individual Images
Traditional Earth observation products often begin with corrected images or physical variables. Embeddings add an intermediate layer between source observations and customer applications. They compress complex sensor patterns into vectors that software can compare and reuse.
A land-cover company could train a small classifier over the embedding rather than build a large model from raw radar and optical inputs. An insurer could search for places that resemble known flood damage. A forestry service could compare representations across dates to identify unexpected change.
The Earth observation foundation model market is forming around this promise of reusable learned features. Training one large model requires substantial data and computing. Many organizations may then adapt its representations with smaller labeled datasets.
The commercial analogy is a standardized feature layer. Satellite operators supply observations, model developers produce embeddings, and application companies attach customer-specific meaning. Each layer can use a different revenue model.
This structure could change competition in the Earth observation data marketplace. Firms may compete on representation quality, geographic coverage, refresh frequency, licensing and integration rather than on imagery alone.
TerraMind Brings Several Observation Types Together
IBM and ESA developed TerraMind as a multimodal foundation model for Earth observation. The wider model family was trained to connect information across several data types rather than treat each sensor as an isolated source.
The Copernicus example uses Sentinel-1 ground-range-detected radar and Sentinel-2 Level-2A optical data. Radar supplies information through clouds and darkness, and optical imagery provides spectral detail related to vegetation, water and built surfaces. Combining them can reduce dependence on clear-sky optical scenes.
Multimodal processing creates engineering demands. Sensors have different resolutions, noise properties and acquisition times. A workflow must align pixels, dates and physical meaning before passing them to a model.
The openEO process graph helps separate data preparation from model execution. Once defined, the workflow can be reused for another area or time period. That supports repeatability and reduces manual scene management.
The satellite data analytics sector already combines observations with models and customer records. Embeddings may make the satellite component easier to reuse, but they do not supply local labels, business rules or ground truth.
A Commercial Data Layer Needs Stable Economics
Generating hundreds of embedding bands across large regions can consume substantial computing and storage. The 768-dimensional TerraMind output is smaller than some raw inputs but larger than a simple land-cover map. Cost depends on area, revisit frequency and the number of model runs.
Commercial providers will need to decide whether to generate embeddings on demand or maintain a precomputed archive. On-demand processing reduces storage but may increase latency. Precomputation supports fast delivery but can create large inventories that need version management.
Pricing could follow area processed, computing consumed, subscription access or application programming interface calls. A provider might also bundle embeddings with trained classifiers and validation services.
The downstream Earth observation market rewards outputs that fit customer workflows. Most users will not purchase a 768-band cube for its own sake. They will pay for a verified answer derived from it.
Open models and public Sentinel data reduce entry costs, yet deployment expertise remains scarce. Companies that can manage cloud processing, model versions and customer-specific testing may occupy a valuable position between public infrastructure and end users.
Verification Will Separate Products From Demonstrations
An embedding may group land-cover patterns effectively in one region and fail in another. Seasonal differences, sensor artifacts and geographic bias can change results. Production systems need tests that reflect the intended market and operating area.
Ground truth remains necessary. A crop classifier needs reliable field labels. Flood mapping requires reference observations that distinguish temporary inundation from permanent water. Infrastructure monitoring needs thresholds tied to inspected conditions.
Model versioning adds another issue. Updating TerraMind or changing the preprocessing workflow can alter the embedding space. A downstream classifier trained on an earlier version may no longer behave the same way.
Providers must also explain what an embedding can and cannot show. A numerical similarity does not establish causation or legal proof. High-consequence applications may require a human review and supporting observations.
The commercialization of space data depends on trust as much as technical performance. Documentation, reproducibility and uncertainty estimates can become commercial differentiators.
Embeddings Could Become Infrastructure Without Becoming Commodities
If several platforms offer reusable Earth observation embeddings, basic access may become inexpensive. That does not mean every representation will be interchangeable. Models differ in training data, sensor coverage, resolution and geographic performance.
A broad public embedding may support exploration and low-cost applications. Specialized providers can add higher-resolution commercial imagery, local training data or sector-specific models. Sovereign customers may require controlled infrastructure and transparent model provenance.
The strongest products will connect embeddings to a repeatable decision. A bank may need verified land-use change around financed assets. A government agency may need an auditable crop-area estimate. A logistics company may need road-disruption alerts.
New Space Economy’s analysis of value-added space services places commercial value in interpretation and delivery. Embeddings can lower the cost of building those services without replacing the need for market knowledge.
Earth observation embeddings are likely to become a common technical layer. Whether they become a large standalone market will depend on licensing, performance and the extent to which customers recognize value in the representation itself.
Summary
The TerraMind openEO example shows how Sentinel radar and optical data can be converted into reusable embeddings through cloud processing. That approach could shorten development for classification, search and change-detection products.
A commercial embedding layer needs efficient computation, stable versions and application-specific validation. The durable revenue opportunity may sit less in selling vectors and more in delivering trustworthy decisions built from them.
Appendix: Useful Books Available on Amazon
- Deep Learning for the Earth Sciences
- Remote Sensing and Image Interpretation
- Fundamentals of Satellite Remote Sensing
- Artificial Intelligence and Machine Learning in Satellite Data Processing and Services
- GIS Fundamentals
Appendix: Top Questions Answered in This Article
What Is an Earth Observation Embedding?
It is a numerical representation generated from satellite data by a machine-learning model. Similar scenes or physical patterns may receive representations that are close in the model’s feature space.
What Is TerraMind?
TerraMind is a multimodal Earth observation foundation model developed through work involving IBM Research and ESA. It can learn shared representations from several observation types.
Which Satellites Does the Copernicus Example Use?
The example uses Sentinel-1 radar and Sentinel-2 optical data. The workflow aligns both collections before model inference.
How Large Is Each TerraMind Embedding?
The published notebook produces a 768-dimensional representation for each 224 by 224 pixel neighborhood. It returns those dimensions as embedding bands.
What Can Embeddings Be Used For?
Potential uses include land-cover classification, crop mapping, similarity search and change detection. Each operational application still requires testing and suitable reference data.
Why Use openEO?
openEO lets users define reusable processing instructions that execute near large Earth observation archives. This reduces manual downloading and local storage.
Is the Published Workflow Ready for Production?
No. Copernicus identifies it as a proof of concept. Production use would require performance optimization, monitoring and stronger quality controls.
Do Embeddings Eliminate the Need for Labels?
No. Many downstream tasks still need labeled examples or ground observations. Embeddings can reduce the amount of feature engineering and training data required.
How Could Companies Charge for Embeddings?
Possible models include subscriptions, area-based processing and application programming interface access. Many providers may bundle embeddings with classification or monitoring services.
What Are the Main Commercial Risks?
Computing cost, model drift, regional bias and weak customer validation are material risks. A technically strong representation may still fail if it does not improve an operational decision.
Appendix: Glossary of Key Terms
Embedding
An embedding is a vector of numbers representing patterns learned from complex data. Machine-learning systems use distances between vectors to compare items or train downstream models.
Foundation Model
A foundation model is trained on large and diverse datasets for reuse across many tasks. Users can adapt its representations rather than train every application from the beginning.
Multimodal
Multimodal processing combines different data types in one model or workflow. In Earth observation, this may include radar, optical imagery and atmospheric measurements.
Data Cube
A data cube organizes geospatial measurements across dimensions such as location, time and spectral band. It supports repeatable processing without treating every scene as a separate file.
Ground Truth
Ground truth is trusted reference information used to train or evaluate a model. It may come from field surveys, inspections, gauges or carefully reviewed imagery.
