A utility receives a satellite alert: possible change near a transmission tower.
The alert arrives sooner than the imagery it was derived from. That sounds useful until the operations team asks what changed, how reliable the detection is, and whether the tower is actually affected.
If the answer is simply “AI detected an anomaly,” the team still has a decision to make without enough context.
As more processing moves onto satellites, this problem will become more important. Onboard AI can help identify relevant observations before data reaches the ground. But the value of a faster alert depends on what accompanies it and what remains available for verification.
The question is not whether satellites should stop sending imagery. It is how they should package observations so that organizations can act quickly and retain the ability to check what the system concluded.
Computing closer to the observation
Two recent developments illustrate different parts of this shift.
On 30 October 2026, Google reported that its Project Suncatcher prototype had generated and returned an AI response using an onboard Gemma model and a data-center-class TPU. The initial response was a poem, followed by other prompts, including questions about payload telemetry. Google says further thermal and radiation testing is planned. blog.google
That is a computing experiment. It is not a demonstration that the prototype detected a real-world event or delivered decision-ready Earth observation evidence.
Earth observation satellites are also exploring onboard AI for more directly related tasks. ESA’s Φsat-2 mission has tested applications including cloud detection and image selection. Its aim is to examine how processing in orbit can improve the usefulness of information sent to Earth. Introducing Φsat-2
Together, these developments raise an architectural question:
When a satellite can interpret an observation onboard, what should reach the person responsible for the next decision?
An alert is a starting point
Imagine a satellite monitoring a transmission corridor after heavy rainfall. An onboard model identifies an area of possible ground change near a tower.
This is an illustrative scenario. Assume the model has been designed and validated for that terrain, sensor and type of change.
A minimal message might say:
Possible landslide near Tower T-17.
That could justify a closer look. But it does not tell the operator whether the tower foundation is affected, whether the change is new, or whether the image was clear enough to support the detection.
A more useful evidence event could contain:
- The candidate location and estimated positional uncertainty.
- The observation time, sensor and image footprint.
- The type of change the model was designed to identify.
- The model and processing version used.
- Relevant image-quality information, such as cloud or visibility limitations.
- A confidence score, if available, together with its meaning and validation limits.
- A reference to the supporting imagery, its availability and retention period.
- The comparison date or baseline, if the detection depends on one.
The event should also state the claim carefully:
Possible ground change detected near Tower T-17. Impact on the tower has not been verified.
That wording matters. A model may detect a change in the landscape without establishing its cause or consequence.
The result is still an observation to investigate. It has not yet become a confirmed assessment of asset condition.
What gets sent and what remains available?
Sending every image at full resolution may be unnecessary for some routine monitoring tasks. Sending only compact alerts can make later review difficult, especially when a model misses something or an operator challenges a result.
A practical approach is to select the transmission package according to the decision:
1. Send an alert first when time matters.
A compact message can give the operations team an early indication of a potential issue. It should include the observation time, location, detection type and major limitations.
2. Send supporting evidence when review is needed.
A relevant image crop, comparison image or other supporting data can help an analyst assess the detection and connect it to nearby assets.
3. Retain or downlink broader data where feasible.
Full scenes or representative samples may be needed for independent review, investigations, model evaluation or future questions that were not anticipated when the image was captured.
The right balance depends on bandwidth, storage, access rights, retention policy, the cost of a missed event and the needs of the customer. In some missions, retrieving an image later may be constrained or impossible. That limitation should be visible to the receiving team.
Efficiency is valuable, but selective transmission also creates a selection risk: events the model fails to identify may never appear in the alert stream. Teams therefore need a way to review samples beyond the alerts and measure missed detections not just the accuracy of the alerts they receive.
The asset context belongs on the ground
A satellite may observe a change close to a transmission tower. The utility must determine whether that change matters to the asset and the people responsible for it.
That assessment may require information the satellite does not have:
- The tower’s verified location and foundation details.
- Recent maintenance or inspection records.
- Terrain and drainage information.
- The age and reliability of the asset record.
- Access constraints and current weather.
- The consequences of a failure at that location.
This is where geospatial platforms and digital twins can contribute. They can connect the candidate observation to asset records and other evidence, while preserving the difference between what was observed, what the model inferred, and what has been verified.
For Tower T-17, the next step might be a desk review. It might be a request for another observation. If the potential consequence is high, an authorized operator might commission a UAV or field inspection.
The satellite’s role is to contribute information to that process. It does not need authority to decide that the tower is unsafe or to dispatch a crew.
Confidence must change the workflow
A confidence score is only useful if the team understands what it describes.
For example, a model score of 0.9 may reflect the model’s confidence in a classification. It does not automatically mean there is a 90% probability that a tower is at risk. Its reliability also depends on how the model was evaluated and whether current conditions resemble those evaluation conditions.
A decision workflow should therefore consider more than a single score:
- Observation quality: Was the relevant area visible, and was the data suitable for this task?
- Detection relevance: Is the observed change connected to the asset or area of interest?
- Decision consequence: What is the cost of missing the event or responding unnecessarily?
- Evidence sufficiency: Can the intended decision be made now, or is another observation needed?
- Authority: Who can approve an inspection, operational restriction or intervention?
This is the distinction between a faster detection and a more reliable operational decision.
Measure the decision, not only the downlink
For a satellite operator, reduced data volume may be an important performance measure. For a utility, the more meaningful question is whether the system helps teams identify and resolve material risks sooner.
A pilot could therefore track:
- Time from observation to alert and from alert to review.
- The proportion of alerts that lead to a verified finding.
- Missed events identified through independent checks or sampled imagery.
- The number and cost of avoidable inspections, where a credible comparison is available.
- Time from a verified finding to an authorized action.
- Whether supporting evidence remained accessible when needed.
These measures must be interpreted together. A system that generates fewer alerts may be efficient or it may be missing important changes.
The next design question
Moving AI into orbit changes where the first interpretation happens. It also means that decisions about filtering, confidence and retention move closer to the point of observation.
That makes the transmission policy part of the evidence system.
For some decisions, a timely alert may be enough to trigger a review. For others, an analyst will need imagery, context and independent verification before anyone acts. The strongest architecture will make that distinction explicit and allow the receiving team to understand what was sent, what was retained and what remains unknown.
The goal is not simply to receive less data from space.
It is to give the right people enough evidence to make the next decision and a clear route to obtain more when they cannot.
Start a conversation


