The number of satellites observing the Earth is growing exponentially, collecting increasing amounts of data critical for scientific purposes and risk assessment. Provided with dedicated processors, these satellites can also be turned from simple remote sensors to intelligent actors. Indeed, running ML models directly on board of spacecrafts presents a real opportunity to accelerate knowledge extraction and ultimately decision making. Low latency predictions would benefit a variety of human activities, above all disaster response, by detecting floods and wildfires as an example. In this challenge, we tackle the fundamental task of detecting whether a location has significantly changed over time so that its data is prioritized for downlink and verification. Our analysis not only assesses the quality of the knowledge extracted by our model but also its suitability for deployment on board, due to specific computational constraints.