Globe Satellites

The project

Several thousand active satellites circle above our heads, and their orbits are public. Celestrak publishes orbital elements in the TLE format (Two-Line Element), which is enough to compute the position of any object at any instant.

Globe Satellites shows that ballet on a 3D globe, refreshed every sixty seconds.

Features

• Filtering by category — starlink, oneweb, communications, navigation, stations, weather, Earth observation, science, other — clickable in the HUD, with "Show all" / "Hide all" buttons
• Hover tooltip — name, altitude and category for each satellite
• Logarithmic radial scale — visibly separates low, medium and geostationary orbits instead of stacking them on top of each other
• Automatic WebSocket reconnection (exponential backoff), with a watchdog that catches a frozen connection with no clean TCP close

Computing on the server

The structural choice is to compute server-side. SGP4 orbital propagation for several thousand objects, once a minute, is not work for the browser — all the more so since every connected client shares exactly the same result. The server computes once (SatrecArray + NumPy, a single vectorised call rather than a Python loop — roughly 24x faster), broadcasts to everyone over WebSocket in parallel, and the client only renders.

Network resilience

TLEs lose accuracy as the days pass — propagating from week-old elements drifts noticeably — hence an automatic refresh every two hours. Loading follows a three-tier chain: the full Celestrak group first, a fallback to 37 groups downloaded in parallel if that fails or returns a 403, then a backup proxy if both fail — a sign of a network block rather than an isolated incident. It returns to the direct route on its own, with no manual action. Data freshness is exposed through a /health endpoint, which pushes best-effort Vigie monitoring for the refresh cycle.

The hardware constraint

The project is built for a Raspberry Pi 4 on ARM64. That constraint drove several decisions: a multi-architecture Docker image, vectorised NumPy computation rather than a Python loop, a single Three.js InstancedMesh rather than one object per satellite, and broadcasting one shared state rather than computing per connected client.