What a satellite chooses to send home is becoming an AI decision

ESA's Φsat-2 demonstrator tests image selection and analysis in orbit, where bandwidth and reversibility matter.

An Earth-observing satellite can collect more imagery than it can immediately send to the ground. Onboard AI changes the order of work: some selection and analysis happen before downlink. That can make the communications link more useful, but it also raises a difficult scientific question. What evidence is lost when the system decides an image is not worth sending?

ESA reported on 15 July 2025 that Φsat-2 had completed commissioning and begun delivering science data. Launched in August 2024, the miniature satellite tests onboard applications including cloud-aware image selection, compression and environmental analysis. ESA describes it explicitly as a demonstrator and reports data access through a registration-based platform. That is evidence of a mission and research capability, not proof of accepted emergency-response performance. [Science-phase report](https://www.esa.int/Applications/Observing_the_Earth/FutureEO/Phsat-2/Phsat-2_begins_science_phase_for_AI_Earth_images), [launch report](https://www.esa.int/Applications/Observing_the_Earth/FutureEO/Phsat-2/New_satellite_demonstrates_the_power_of_AI_for_Earth_observation).

Cloud filtering illustrates the tradeoff. A mostly obscured image may be less useful for one task, yet contain a clear region important to another. A binary decision to discard the whole scene can remove that information. Evaluation should therefore examine the consequence of filtering for the intended users, rather than only the accuracy of a cloud classifier.

An illustrative onboard pipeline might first assess image quality, then calculate task-specific detections and finally prioritize transmission. Keep those stages identifiable. An analyst should know whether a missing image was not acquired, filtered, delayed or lost during transmission. Combining all four into a blank map makes it impossible to interpret coverage honestly.

Retain enough information to audit selection. A mission may not be able to downlink every raw image immediately, but it can design representative retention and ground-validation procedures. Compare selected and rejected samples under different geography, lighting and cloud conditions. Rare scenes deserve attention because they may be exactly the events that motivate the mission.

Compression presents a related issue. Smaller files can make transmission easier, but reconstructed images must be assessed for the downstream scientific task. A picture that looks acceptable to a human may have altered the small spectral or spatial features a detector relies on. Evaluate the analysis before and after compression, with clear limits on what the reconstruction preserves.

Onboard deployment is also a resource-management problem. Compute competes with power, memory, thermal limits and other spacecraft activities. A model that works in a ground workstation environment may exceed the satellite's available budget. Test realistic input sizes, execution time and degraded conditions. Keep a defined response for failed inference so a software problem does not disrupt unrelated mission functions.

The Φ-sat programme's broader role is testing AI technologies in flight. That flight context matters because radiation, communications and mission scheduling create conditions that a desktop benchmark cannot reproduce. A research model should not inherit a claim of flight readiness simply because a related algorithm has flown. [Programme context](https://philab.esa.int/flagship-programmes/phi-sats-programme/).

Developers exploring the data can begin with an offline task and clearly separated ground labels. State acquisition dates, processing levels and spatial coverage. Do not interpret a demonstration image as proof that the model reliably identifies every instance of a phenomenon. An alert intended for a public service needs an agreed threshold, verification route and a delivery time that matches the service's decision.

Costs include spacecraft integration and operations, not merely the price of model inference. Ground storage, validation, updates and analyst review also remain necessary. ESA's free registered research access does not establish a commercial service price or a guaranteed operational support level. Those require specific mission and supplier terms.

Audit the discarded-data policy itself. For example, retain a defined sample of scenes rejected as cloudy and check whether useful coastal regions were removed with them. Report that loss separately from compression quality. A filtering algorithm can be accurate on its own label and still make the mission less useful for a particular scientific question.

Onboard AI is compelling because it can make scarce communications capacity more useful. Its strongest implementation preserves visibility into what was selected, what was withheld and how uncertainty was handled. The satellite's intelligence should improve the evidence reaching the ground while leaving scientists able to understand the decisions that shaped it.