In our most recent webinar at the IoA, we gave satellite first AI an honourable mention as a potentially exciting technical breakthrough in the coming years.
We often talk about edge AI as something that happens on a device on the ground, close to the point of measurement. In 2026, there is a second version of that idea that deserves more attention. The edge can also be in orbit.
For disaster response, environmental monitoring, and maritime safety, the limiting factor is not always the quality of the model. It is the time between an event happening and a decision being supported with reliable evidence. Satellite imagery and sensor data are powerful, but they have historically followed a slow pattern: collect in orbit, transmit to the ground, process in a central environment, then distribute insights. On board processing changes that sequence. The satellite does not just observe, it begins to interpret.
There is a practical reason this is resurfacing now. The volume of Earth Observation data is rising quickly. ABI Research is cited as estimating that there are over 900 active Earth Observation (EO) satellites, and some industry estimates suggest each EO satellite can generate around 100 terabytes of data per day, which would become extremely difficult to move and process centrally at scale. When data is abundant but downlink capacity and response time are constrained, filtering and prioritisation in orbit becomes less like an optimisation and more like a necessity.
What changes when we process data before downlink?
On-board AI is not primarily about making satellites ‘smarter’ in an abstract sense. It is about changing what comes down to Earth.
Instead of downlinking everything and sorting later, the satellite can discard low value captures, compress effectively, detect specific patterns, and transmit alerts that are immediately useful. ESA’s Φsat programme is a clear example of this direction. Its published impact assessment describes cloud detection that processes images in orbit so that only clear, usable images are sent to Earth, and it also describes on board compression and reconstruction to reduce file sizes before downlink.
The same document is explicit about why this matters beyond efficiency. It frames on board AI applications as supporting disaster response, environmental protection, and security, with examples including street map generation for emergency navigation, maritime anomaly detection for oil spills and harmful algal blooms, and wildfire detection intended to supply real time information to response teams.
This is the shift in plain terms. The satellite becomes part of the decision pipeline, not just the data pipeline.
Why this matters for disaster response and environmental monitoring
Disasters are defined by time pressure and incomplete information. Emergency teams need situational awareness that is current enough to act on, and consistent enough to trust.
A useful way to see the gap is to look at established systems that work well but still have unavoidable delays. Even if you treat any single study cautiously, the basic reality is well known: the observation, downlink, processing, validation, and distribution chain introduces delay.
On board processing can shorten that chain in two ways:
First, it can reduce the amount of data that must be transmitted by sending only what is likely to be actionable. Cloud filtering is a simple example, because cloud cover is a common reason EO imagery is not immediately usable. If the satellite can identify this in orbit, the ground system spends less time processing ‘noise’.
Second, it can detect and flag-relevant events so that high priority data is sent first, and sometimes so that the satellite adjusts what it observes next. NASA JPL’s Autonomous Sciencecraft work on EO 1 provides a useful historical precedent here, describing on board analysis that can detect features such as clouds, flooding, ice formation, or volcanic activity, and then trigger additional data collection, while also deciding when to send data back based on what was observed. That is an early version of an idea that is now becoming more feasible for more missions: interpret, prioritise, then transmit.
In disaster response, this is not a minor improvement. It means analysts and responders can spend more time validating meaningful signals, rather than searching for them in a backlog.
Unfortunately, there are constraints
It is important not to oversell what orbit based AI can do, because the constraints are real and shape what works.
Compute and power are limited. Radiation and reliability requirements affect hardware choices. Models need careful validation because false positives can waste resources and false negatives can create harm. Updating models is also non-trivial, because you are operating in an environment where any change has operational risk. These are reasons why many on board deployments focus on bounded tasks, such as filtering, segmentation, anomaly detection, classification of known phenomena, and generation of compressed data.
This is also why the value case is strongest where there is a clear operational payoff. Wildfire growth, flood extent, oil spill detection, rapid map generation after an earthquake, and vessel awareness are all domains where a timely approximate answer can be more useful than a perfect answer delivered too late. ESA explicitly positions Φsat 2 applications in that space, describing crisis management benefits and real time monitoring use cases.
In conclusion
Satellite processed AI is best understood as an extension of edge computing into an environment where time, bandwidth, and attention are constrained. The growth in Earth observation activity makes it harder to rely on ‘downlink everything then decide’, and the use cases in disaster response and environmental monitoring make it easier to justify ‘decide enough in orbit to act sooner on Earth’.
As AI becomes embedded into more critical decision pipelines, the location of inference matters.
