Where AI is actually being used in Earth observation

From spotting a fishing boat that is not broadcasting its position to helping emergency managers decide where to send people after a flood, artificial intelligence is becoming part of the Earth observation production line. The important question is not whether a service uses AI. It is whether it produces a result that someone can understand, check and act on.

In Earth observation (EO), Artificial intelligence (AI) covers a practical collection of tools including software that finds a feature in an image, estimates the likelihood of a change, combines satellite data with other evidence or helps a person ask a question of a very large archive.

That is important because EO now produces more data than most organisations can inspect manually but the commercial opportunity is not simply to make more maps. It is to turn observations into a timely, explainable answer to a real operational question such as which field needs attention, which ship warrants investigation, where is a flood likely to affect policyholders or which emission source should be checked first?

AI is already used in the basic EO workflow

The most mature use is machine learning. This is software trained on examples rather than written as a long list of rules, give the machine enough labelled images of crops, vessels, burn scars or flooded land and it learns patterns that can be applied to new imagery.

The result is usually a classification or an estimate but certainly not a final fact. A model might label a pixel as maize, identify a shape as a vessel or estimate the depth of flooding at a building. Good services attach confidence information, retain the underlying imagery and define when a human should review the result.

The next step is the foundation model. Rather than training a separate model from scratch for every job, a foundation model is trained on very large and varied datasets first. It can then be adapted using much less local data for a specific task. The European Space Agency (ESA)’s Tessera model, for example, was trained using Copernicus Sentinel-1 radar and Sentinel-2 optical observations. It produces compact numerical representations known as embeddings, which capture useful patterns in a scene and can support work such as crop monitoring, burn-scar mapping and forest analysis.

The newest category is agentic AI software that can plan and carry out several steps towards a task, such as finding data, preparing an analysis and reporting a result. This is still early-stage territory in EO but already it can make complex data platforms easier to use whilst still needing an analyst to check the question, sources, assumptions and output.

At sea: detecting what vessels do not report

Automatic Identification System (AIS) signals are a central source of maritime information. AIS is a safety and tracking system through which vessels broadcast details such as position, course and identity but smaller fishing vessels often do not use it and larger vessels’ signals can be absent, unreliable or deliberately manipulated.

Global Fishing Watch illustrates how AI fills part of that gap. Its small-vessel work uses high-resolution imagery and machine learning to search for boats that may occupy only a few pixels. The system first produces a deliberately broad set of possible detections and further models use image patterns and context such as distance from shore and weather to separate likely vessels from glare, waves or wakes, then help distinguish fishing from other activity.

This is a useful reminder that AI rarely replaces other data and the strongest result comes from comparing the satellite view with AIS, vessel registries, fishing logbooks and human review. Global Fishing Watch’s own recent explanation says analysts review potentially suspicious behaviour flagged by experimental AI agents, acknowledging that biases in the inputs can pass through to the result.

For EO suppliers, the lesson is clear that there is value in solving an evidence-fusion problem. A buyer may care far more about a defensible explanation of an unusual vessel movement than a standalone image or detection score.

Methane: finding the sources worth fixing

AI also helps turn specialist measurements into a practical maintenance queue. Carbon Mapper’s Tanager programme uses imaging spectroscopy, which measures reflected light across many narrow wavelength bands and, because different gases absorb light in distinctive ways, it provides a spectral fingerprint. Processing identifies and estimates methane and carbon-dioxide plumes before linking them to a likely facility or piece of equipment where possible.

Tanager-1 has been used to publish observations of emissions from oil and gas, waste and coal sources. Carbon Mapper reported methane observations in Russia, Argentina and China during 2026, alongside data made available to users who can investigate or act on the findings.

The AI component is only part of the service. The broader workflow includes sensor calibration, atmospheric processing, emission quantification, attribution, notification and follow-up on the ground. Carbon Mapper is explicit that satellite observation should be one part of a portfolio of measurement methods. That is the right commercial framing as EO can direct scarce inspection and repair resources, but it should not be sold as a complete substitute for verification.

Floods and wildfires: faster situational awareness

Natural-hazard services show the value of combining EO with models and local information. ICEYE uses synthetic aperture radar (SAR) satellites to send microwave signals towards Earth and measure their return. Because radar does not depend on daylight and can see through cloud, it is valuable during storms and smoke when optical imagery is limited.

ICEYE says its flood products combine SAR observations, ground observations, third-party datasets and machine learning to estimate flood extent, depth and likely impacts. In October, California’s Office of Emergency Services began integrating ICEYE flood and wildfire products into its operations, after using its data during the 2025 Los Angeles fires through a Federal Emergency Management Agency collaboration.

That is a real operational use case but it comes with an important caution that a rapid impact layer is a prioritisation tool. It may tell an emergency manager or insurer where to look first and how to allocate teams but it is not the same as a ground inspection, a claims decision or a complete account of damage. Providers should be precise about the resolution, update frequency and circumstances in which their predictions are reliable.

ESA-backed wildfire projects point to the next frontier. FIREWISE, for instance, aims to run EO foundation models onboard satellites so that alerts can be produced with less dependence on downlink and ground processing. It is a research and development effort, not a deployed public safety service, but it highlights a real market pressure which is that in fast-moving hazards, speed can matter as much as spatial detail.

Agriculture: applying inputs more precisely

Agriculture has used EO for years but AI is making it easier and faster to turn a time series of images into field-level advice. ESA’s WorldCereal system provides crop and irrigation mapping based on Copernicus Sentinel-1 and Sentinel-2 data and Landsat observations. Its models can be reused or adapted for a user’s area and crop types although the project warns users to test models on small sites before scaling them up.

There are more targeted commercial examples. ESA Space Solutions describes Messium’s approach, which trained machine-learning models against 20,000 geographically referenced wheat samples combined with hyperspectral satellite observations. It then combines the resulting crop-health estimates with weather, soil and farm-management data to support nitrogen decisions.

For farmers, the useful question is not whether the algorithm is clever but whether the recommendation is sufficiently timely, local and trustworthy to change an input decision. The model must cope with cloud, variable growing conditions, unusual crop calendars and the fact that similar-looking fields can need different treatment which means that ground samples remain essential, both to train the model and to check it.

Making EO questions easier to ask

Large language models can help non-specialists navigate EO information and processing systems, something that is behind ESA’s Earth Virtual Expert, which uses a retrieval-augmented generation approach. In plain terms, it searches selected source material at the time of a query and uses that material to construct an answer, rather than relying only on information absorbed during training. ESA has also added a separate AI review step intended to identify unsupported content.

The Copernicus Data Space Ecosystem’s openEO Agent has a more operational role and it can turn a plain-language request into a proposed cloud-processing workflow, show the process chain and submit a job after approval. The service is explicitly in an initial testing phase and currently supports the Python client, with its own guidance warning users to check suggested analytical approaches.

These tools could widen access to EO especially for teams that understand a business or environmental problem but do not yet have deep data-engineering skills. However, maybe they should be treated like an eager junior analyst who is helpful in finding a route through a large body of information but not an authority whose work can be accepted without review.

What a credible AI-in-EO service looks like

For buyers and suppliers, five questions cut through most of the hype:

1. What decision changes? Name the user, the action and the time window.
2. What is the evidence chain? Explain how satellite, ground and third-party data are combined.
3. How was the model tested? Performance should be reported across locations, seasons, sensors and relevant edge cases.
4. How is uncertainty shown? A confident-looking map is not enough. Users need confidence scores, thresholds and clear escalation rules.
5. Who is accountable? Define who checks exceptions, corrects errors and communicates the limits of the service.

Right now, the most promising AI applications in EO make the link between observation and action shorter whilst leaving enough evidence for users to understand and challenge the answer.