Earth observation constellations capture 363,563 images per day at maximum rate. But due to downlink constraints, only 11.7% of that data ever reaches ground stations within 24 hours. Microsoft researchers asked: What if we trained models in space instead? ## Enter OrbitalBrain Instead of satellites as passive data collectors, OrbitalBrain turns nanosatellite constellations into distributed training systems. Models train, aggregate, and update directly on orbit — using onboard compute, inter-satellite links, and predictive scheduling. ### Core Philosophy The framework recognizes three key satellite characteristics: - Constellations are typically single-operator, enabling raw data sharing - Orbits, power, and ground visibility are predictable - Inter-satellite links (ISLs) and onboard accelerators are now practical ### How It Works Each satellite performs three actions under a cloud-computed schedule: - Local Compute: Train on stored imagery - Model Aggregation: Exchange parameters over ISLs - Data Transfer: Rebalance data distribution between satellites A cloud controller predicts orbital dynamics, power budgets, and link opportunities to optimize the schedule. ## Why Federated Learning Fails in Space Standard FL approaches (AsyncFL, SyncFL, FedBuff, FedSpace) break down under real satellite constraints: - Intermittent connectivity: Updates become stale before aggregation - Power limits: Computing competes with essential operations - Non-i.i.d. data: Each satellite sees different scenes Result: 10–40% accuracy degradation compared to idealized conditions. ## OrbitalBrain Results Simulated on real constellations (Planet: 207 sats, 12 ground stations; Spire: 117 sats): | Task | Baseline Best | OrbitalBrain | Improvement | |——|—————|————–|————-| | fMoW (Planet) | 47.3% | 52.8% | +5.5% | | fMoW (Spire) | 40.1% | 59.2% | +19.1% | | So2Sat (Planet) | 42.4% | 47.9% | +5.5% | | So2Sat (Spire) | 42.2% | 47.1% | +4.9% | Time-to-accuracy: 1.52×–12.4× faster than ground-based approaches. ## The Bottom Line OrbitalBrain proves that satellite constellations can act as distributed ML systems, not just data sources. This enables: - Real-time models for forest fire detection - Fresh flood monitoring data - Climate analytics without multi-day delays The future of Earth observation isn’t just better sensors — it’s better coordination.