FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations
FedOrbit is an adaptive personalized federated learning framework designed for non-IID Low Earth Orbit satellite constellations that combines continuous inter-satellite training, class-aware hierarchical aggregation, and adaptive feature decomposition to significantly outperform existing baselines in accuracy and consistency across diverse remote-sensing benchmarks.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a classroom where the teacher can't walk around to check everyone's homework because they are stuck in a tiny office with a broken door. Instead, the teacher asks students to solve problems on their own and only send back their final answers. This is the basic idea behind Federated Learning, a clever way for computers to learn together without ever sharing their private data. Usually, we imagine these computers as phones or laptops sitting on a desk, all connected to the internet at the same time. But what if the "students" were actually satellites zooming around the Earth at 17,500 miles per hour?
This is the wild new frontier of Low Earth Orbit (LEO) satellite constellations. These are swarms of satellites working together to take pictures of our planet, from tracking wildfires to mapping cities. The problem is that these satellites are constantly moving. Sometimes they are right over a ground station (the "teacher's office") and can send data; other times, they are on the other side of the world, invisible and cut off. Furthermore, because they fly in specific paths, one satellite might only see deserts while another only sees oceans. This means their "homework" (the data they collect) is completely different from each other, a situation scientists call non-IID (not independent and identically distributed). If you try to teach a single, one-size-fits-all model to all these satellites, it fails miserably because the "global" answer doesn't make sense for any single orbit. The question becomes: How do you train a smart AI team when the team members are constantly flying in and out of range, and they all see different parts of the world?
Enter FedOrbit, a new method proposed by researchers Satwat Bashir, Tasos Dagiuklas, and Muddesar Iqbal. Think of FedOrbit as a super-organized study group for these flying satellites. Instead of waiting for the teacher to call on them, the satellites have their own rules for how to learn, even when they can't talk to the ground.
The researchers discovered that the old way of doing things had two big flaws. If the satellites saw totally different things (like one seeing only deserts and another only forests), a single global model would collapse into confusion. But if they tried to make every satellite totally unique, they would miss out on useful shared knowledge. FedOrbit solves this by acting like a flexible coach. First, it keeps every satellite learning, even when they are invisible to the ground, by letting them pass notes to each other through inter-satellite links (satellite-to-satellite radio). Second, it uses a "class-aware" system: if a satellite is the only one that sees a specific type of cloud, its opinion on that cloud gets extra weight, while its opinion on things it doesn't see gets ignored.
To handle the fact that some satellites are gone for long periods, FedOrbit uses a "return-rate dampening" trick. Imagine a student who has been absent for a month; when they return, they don't just jump into the advanced class immediately. They take a few steps back to catch up so they don't disrupt the flow. Similarly, when a satellite reappears after a long absence, FedOrbit slows down its learning speed slightly so it doesn't overwrite the group's progress with outdated information. Finally, the system uses a smart "mixing" strategy. If the satellites are all seeing very different things, the system keeps their personal knowledge separate. But if they are seeing similar things, it blends their knowledge together automatically.
In their tests, the authors simulated this system using three different remote-sensing datasets (EuroSAT, So2Sat, and RESISC45) and two different ways of splitting the data. The results were impressive. In five out of six test scenarios, FedOrbit achieved the highest accuracy, beating the next best method by as much as 16.1 percentage points in difficult situations. Even in the one case where it didn't win, it was within just 0.9 percentage points of the best result. Perhaps most importantly, FedOrbit made the whole team fairer: the difference in performance between the best-performing satellite and the worst-performing one was the smallest in five out of six cases. This suggests that FedOrbit isn't just making the satellites smarter; it's making sure no satellite gets left behind, even if it spends most of its time in the dark.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.