Satellite planning
From a question to a satellite plan
AID-Mission turns a request like "warn me of landslides here within 48 hours" into a checked satellite and processing plan, and explains every decision in it.
Five checks between a question and a plan
Studies and a language model can advise on the choice of method, but the physics rules always decide.
Understand the area
Open global maps and image models describe the place and split it into terrain zones. Every fact keeps its source; a failed source leaves a fact unknown rather than guessed.
Choose the method: the rules decide
15 analysis methods are judged per zone against physics limits. Studies back each choice with a quote checked word for word, and a language model may only rank what the rules allow.
Find images that will actually exist
Orbits for five constellations, ESA recording plans, cloud forecasts and ground-station contacts show whether and when each image will be ready.
Schedule
An optimiser picks the best set under the deadline, budget and geometry rules, and records why each option was or was not chosen.
Plan with the user
The user sees the plan, zones, evidence and map footprints, then approves it, asks for faster or more certain data, or edits the deadline. Each change is a new version.
The checks also catch requests that do not fit the land. Asked for a landslide warning on flat farmland in Punjab, the planner replied: "Only 0.2 % of the area is moderate or steep slope... Did you mean subsidence?"
A sample plan
Three requests at once: infrastructure deformation in Gävle, groundwater subsidence in Punjab, and flood extent on the Uppsala plain. From 631 possible observations across 56 satellites, 168 were feasible and 7 were chosen.
A simple earliest-first plan met 2 of the 3 requests at a simulated cost of 2 units. AID-Mission met all 3 at a cost of 4, because a radar stack is only useful when each image has a partner from the same geometry.
Does it predict real images?
Yes, for the missions we could check. We predicted passes over past windows and matched them against the public archives, and modelled when data would be ready against real product times.
| Mission | Real images predicted | Modelled time to data | Real median |
|---|---|---|---|
| Sentinel-1 | 127 / 127 | 2 h (assumed) | not published |
| NISAR | 51 / 51 | 44 h | 44.6 h |
| Sentinel-2 | 115 / 115 | 3.4 h | 3.7 h |
| Landsat 8/9 | 41 / 41 | 6 h | 6.2 h |
Sentinel-2 and Landsat also matched on a separate window (108/108 and 41/41). The check runs one way: some predicted passes are never recorded, for example 32 for Sentinel-1 and 21 for NISAR in the 28 days. For Sentinel-1, ESA's recording plans address this.
Would it have warned in time?
We replayed five real slope failures through today's satellites, the planner and an alarm rule. Settings were tuned on four events and tested on the fifth. Hand settings missed every failure that could be warned of.
| Failure | Simulated warning, tuned | Missed runs | Real warning |
|---|---|---|---|
| Brienz/Brinzauls, Switzerland, 2023 | 20 days | 1 of 3 | 34 days |
| Brumadinho, Brazil, 2019 | 19 days | 1 of 3 | none |
| Xinmo, China, 2017 | 10 days | 2 of 3 | none |
| Blatten, Switzerland, 2025 | 1.8 days | 0 of 3 | 9 days |
| Stenungsund, Sweden, 2023 (quick clay) | correctly no warning | – | none |
A demonstration, not a measurement: motion curves are estimates from the literature, with simulated noise. There was one false alarm in 3 quiet controls for Brumadinho and one in 6 runs for Stenungsund. Brienz also relied on simulated commercial images.
The main finding is a gap. Slopes moving roughly 1–40 m per year are too fast for radar interferometry and too slow for Sentinel-1 image matching. Only high-resolution commercial radar, optical matching on clear days, or ground sensors close it.
What did not work as hoped
- Learning from similar sites. In a library of 1,092 located studies, site similarity did not predict a study's method better than simply picking the most used method. Published studies are biased: 94% report success. Research therefore advises and the physics rules stay in charge.
- Few test cases. The physics check agrees on all 8 labelled cases, but that is too few to prove it on new terrain.
- The first landslide-mapping model is a baseline. It reached F1 0.56 on its test set and overfits after about 13 training rounds.
- Narrow rivers. The water model agreed well with a reference map at Gävle and Punjab (IoU 0.65 and 0.77) but missed a narrow river at Uppsala (0.10).
What comes next
Running one real event through the radar pipeline and alarm, using real tasking under a provider contract, and deciding where each processing step should run. Plans will also gain more alternatives and fuller editing.
Built on open data including CelesTrak orbits, ESA Sentinel-1 plans, Copernicus DEM, ESA WorldCover, Global Sentinel-1 coherence (Kellndorfer et al. 2022), ASF, Copernicus Data Space and USGS archives, ERA5 via Open-Meteo, OpenAlex and Crossref, with Orekit, OR-Tools, Prithvi-EO-2.0 and SAM 2.1. Plan map outlines: Natural Earth.