Satellite Imagery Licensing for AI Training: Commercial EULAs, Open EO Licenses and Model Ownership

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why satellite imagery licensing is different from other image data

Satellite licenses are written for map products and analytics, not for model weights, so most of them say nothing explicit about training. A commercial EULA typically grants a limited, non-exclusive, non-transferable license to use imagery for internal purposes and then defines what you may do with derived outputs [1]. That structure was designed for orthomosaics, change maps and NDVI layers. Whether a segmentation checkpoint or a geospatial foundation model fits inside "internal use" or "Value-Added Product" is a drafting question you have to settle, not assume.

Three features make the category harder than consumer photos. First, one archive often blends constellations from different owners: a distributor can resell SPOT and Pléiades scenes under a separate EULA from its own imagery [3]. Second, open Earth observation programs use public-sector licenses rather than standard Creative Commons, and open aggregations then layer their own terms on top [4]. Third, provenance gets lost fast: a large audit of public dataset hubs found licenses omitted for over 70% of datasets and miscategorized for over half [8]. For resolution and capture trade-offs before you reach licensing, see aerial vs satellite vs drone imagery, and for the wider category, the image data hub.

Open EO sources: what Sentinel, Landsat and WorldStrat actually allow

Sentinel and Landsat both allow commercial training; WorldStrat does not. The terms differ in ways that matter for redistribution and model release, and the SeeFar dataset on the AWS open data registry shows all three side by side [4].

Copernicus Sentinel. In the SeeFar foundation-model corpus, Sentinel imagery is released under CC BY-SA 3.0 IGO, which allows commercial use but requires attribution and share-alike on adaptations [4]. As of October 2026, Sentinel data fetched directly from Copernicus services comes with the program's own free, full and open legal notice and its attribution wording ("Contains modified Copernicus Sentinel data [year]") [9], so read the license attached to the copy you actually downloaded and record its version.

Landsat. Landsat imagery is unrestricted, with acknowledgment of the source requested rather than required as a license condition [4]. Collections that mix Landsat with purchased commercial scenes can carry the commercial provider's terms, so check the collection, not just the mission name.

WorldStrat and other NC sets. WorldStrat's high-resolution imagery is distributed under CC BY-NC 4.0 [4]. Using it to train a model you sell, host as an API or use in a paid service is a commercial use, so keep it out of production training corpora. Treat research tiers of commercial providers the same way: one provider's education and research terms define "Noncommercial" as uses from which the licensee derives no profit [2]. For a cross-license view, use the open data license compatibility matrix.

Share-alike and the "is the model a derivative" question

No settled rule says a trained model is or is not a derivative of its training imagery, so the contract has to answer it. Share-alike terms require adaptations to carry the same license, and a University of Amsterdam IViR study examines how far that obligation reaches when works are used to train models [6]. The practical risk for a CC BY-SA 3.0 IGO Sentinel copy is less about weights and more about outputs: tiled training chips, super-resolved scenes or labeled masks you redistribute are plausibly adaptations.

Commercial EULAs solve this through definitions. Look for how "Value-Added Product" or "Derived Product" is defined, whether it requires the imagery to be irreversibly transformed, and whether you own it outright or hold it under the same license [1]. A case study of archive imagery licensed for AI training shows the failure mode: the parties had not defined training scope, whether the model was a derivative, or what happened on consent withdrawal, and resolved it only with a written scope, removal tiers and a chain-of-license warranty [5]. For model-rights drafting in general, see derivative and successor model rights.

Commercial EULA clauses to negotiate before signing

The clauses that decide training rights are scope of use, derivative ownership, transferability and term. Ask for each one in writing, because a license that is silent on training leaves the question open to dispute.

• Permitted use. "Internal use" may or may not cover a model served to customers. Name training, fine-tuning, evaluation and deployment explicitly.
• Value-Added Product definition. Confirm that model weights, embeddings and inference outputs fall inside it and that you own them [1].
• Transferability. Non-transferable licenses [1] can block a sale of the company, an affiliate's use or a cloud partner's hosting. Ask for affiliate and successor rights.
• Seat or entity scope. Decide whether the grant is per model or enterprise-wide; see per-model vs enterprise-wide licenses.
• Termination and removal. Agree what happens to trained weights if the license ends or a scene is withdrawn: retrain, quarantine or keep [5].
• Chain of title. Where a distributor resells another operator's constellation, require a warranty that it may grant training rights [3][5].
• Database rights. For imagery compiled in the EU, a separate sui generis database right under Directive 96/9/EC may protect substantial extractions; ask counsel whether it applies to the specific archive.

Our AI data license terms guide covers the general vocabulary, and the dataset licensing glossary entry defines the core terms.

Per-scene license manifest for mixed EO corpora

Track the license at the scene or tile level, because a single training corpus routinely mixes Sentinel, Landsat, NC research sets and commercial scenes [4]. A manifest field per scene lets you filter a commercial training run, generate attribution notices and answer a diligence request without re-auditing.

Illustrative example: invented to show structure; it does not describe an available dataset.

A companion decision table helps triage sources quickly:

Illustrative example: invented to show structure; it does not describe an available dataset.

Diligence questions to put to any imagery supplier

Ask suppliers to document rights at the same granularity you will train at. The training data due diligence checklist and the open dataset license audit cover the general process; for satellite imagery add these:

1. Which constellations and which distributors are in the delivery, and under which EULA version each [3]?
2. Does the license name model training, and who owns weights and outputs?
3. Can the license transfer to affiliates, acquirers and hosting partners?
4. Are any scenes from research, NC or commercially purchased augmentation programs [2][4]?
5. What attribution text must ship with redistributed chips or masks [4]?
6. What happens to models if a scene is withdrawn or the term ends [5]?

Some commentators urge operators to release some derived products openly to meet analytics demand [7], so check for a newer, more permissive tier before negotiating a custom grant. If you also need licensed location records alongside imagery, see geospatial and location data for AI training, or describe the operational data you want to SourceX buyers.

License imagery-linked operational data for your models

SourceX sources operational datasets from US companies on request, such as documents, engineering records and new recordings of hands-on work, and manages the licensing process; categories are not inventory and a request does not guarantee a match. Every dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery, with each release approved by the supplying company. Describe the data you need at SourceX for buyers.