Making real-time model updating transferable across geoscience and remote sensing — through commercial tasking and on-board edge computing
Every simulation drifts. Whatever produces the forecast — a physics-based model, a big-data statistical model, or a modern AI / geospatial foundation model — its state slowly diverges from the real world. Observations are what pull it back. Data assimilation (DA) is the general, model-agnostic machinery for that correction: it fuses a model's prediction (the prior) with noisy observations (the likelihood) to produce a better estimate (the posterior), together with an honest measure of the uncertainty that remains.
My doctoral work applies Bayesian, ensemble-based DA to glacier mass balance in the Arctic. But the mathematics never knew it was looking at a glacier. The same update that corrects an ice model with a satellite albedo retrieval can correct a flood model with a water-extent map, an air-quality model with a plume detection, or an ML surrogate with any stream of measurements. This project explores that transferability — and argues that the real frontier is no longer the assimilation math, but the latency of the observation pipeline that feeds it.
DA is not a one-shot calculation — it is a loop. Its value compounds when the loop closes quickly and, crucially, when the final action can include re-pointing the sensor that feeds the next cycle.
Sensor captures a scene (optical, SAR, thermal, hyperspectral)
Turn raw pixels into the observable the model needs
Fuse model prior with the observation → corrected state + uncertainty
Propagate the updated state forward in time
Trigger a decision — and re-point or re-orbit the sensor for the next pass
In today's workflows the arrows between Observe and Assimilate can take hours to days: the satellite must be in range of a ground station, downlink a large raw image, and wait for ground processing. The two-tier architecture below attacks exactly those arrows.
A common misconception is that DA belongs to physics-based modeling. It does not. DA sits above the model layer: it only requires (1) something that produces a prediction with an uncertainty, and (2) an observation operator that maps model state to what the sensor sees. The forecast engine underneath can be anything.
Energy-balance, hydrological, atmospheric, or ocean models — the classic DA setting.
Fast neural approximations of expensive physics — assimilated exactly like the model they replace.
Large pre-trained Earth-observation models whose predictions still need observations to stay anchored to reality.
Physics core with learned corrections — increasingly the state of the art.
AI is not a competitor to DA here — it enters at three distinct layers, and edge computing makes each of them deployable on-orbit:
The proposal is deliberately staged: an incremental, low-risk tier that is feasible with today's flight hardware, and an ambitious tier that turns a satellite into an autonomous scientific agent.
The satellite does not downlink terabytes of raw imagery. On-board edge compute runs the observation operator in orbit and sends home only the extracted observable — a few numbers or a thin map per scene — ready to drop straight into a ground-side assimilation.
The model and the assimilation live on the spacecraft. Observations are captured, assimilated, and used to update the forecast entirely in orbit; only the analysis or the decision is downlinked. The loop closes without ever touching the ground — enabling autonomous action: change the observing mode, re-point, or maneuver to chase the phenomenon.
Comparing the conventional pipeline against the two tiers makes the trade-offs explicit. Each step up buys latency and autonomy at the cost of on-board complexity.
| Dimension | Conventional ground DA | Tier 1 · Edge pre-processing | Tier 2 · On-board DA |
|---|---|---|---|
| Observation → update latency | Hours to days | Minutes to hours | Seconds to minutes |
| Downlink volume | Full raw imagery | Extracted observable only | Analysis / decision only |
| On-board compute burden | Minimal | Moderate (inference) | High (model + assimilation) |
| Autonomy | None — human/ground in the loop | Low — smarter downlink | High — self-tasking, maneuvering |
| Main cost driver | Tasking + ground processing | Tasking + flight compute | Flight compute, power, validation |
| Maturity | Operational | Demonstrated in orbit | Early / experimental |
| Best fit | Slow, large-scale processes | Bandwidth-limited monitoring | Fast, transient, decision-critical events |
The reasoning below rests on a set of assumptions that already hold, or are credibly on their way, in 2026.
Commercial constellations (Planet, BlackSky and peers) offer sub-daily to intraday tasking, so temporal gaps are now a budget question, not a physical limit.
AI accelerators now fly: ESA's Φ-sat proved on-board neural inference, and Planet's Pelican fleet operates NVIDIA Jetson Orin modules running detection models in orbit today. This assumption is no longer forward-looking.
What DA actually needs — a snow-cover fraction, a flood mask, a plume concentration — is tiny compared with the raw scene, so moving the extraction upstream pays off.
Reduced-order models, ML surrogates, and local sub-domains make it plausible to run a forecast + assimilation step within an on-board power/compute budget.
Ensemble or Bayesian methods provide the error bars that any downstream decision (or retasking trigger) needs to be trustworthy.
For a large class of hazards, the worth of an observation decays fast — a correction that arrives in minutes is worth far more than one that arrives in days.
The same loop, re-pointed at different problems. Each case pairs a concrete observable with the model it would update and the tier that fits best.
On-board snow-covered-area or albedo retrieval updates a mass-balance / melt model — my home domain, and the natural first testbed.
SAR-derived water extent assimilated into a hydrodynamic model; Tier 2 lets the satellite re-point to a breaking flood front in real time.
Thermal hotspot detection updates a fire-spread model; autonomous retasking tracks an active front between passes.
On-board plume detection and quantification updates an emission/transport model — flagging leaks in minutes, not weeks.
Vegetation and moisture indices assimilated into crop / land-surface models for yield and drought forecasting.
Eruption and ash detection updates a dispersion model to re-route aircraft — the exact scenario EO-1's autonomy first proved.
Ocean-colour retrievals assimilated into biogeochemical models to nowcast harmful blooms for fisheries and drinking water.
Aerosol / trace-gas columns update chemical-transport models for rapid pollution nowcasts over cities.
The vision is compelling, but its limits are real and deserve to be named alongside the upside.
High-revisit commercial imaging is a recurring expense; the value of low latency has to justify it, case by case.
On-board budgets are tight and the space environment is hostile — Tier 2 models must be small, robust, and fault-tolerant.
A model that assimilates and acts autonomously in orbit is hard to audit; uncertainty bounds and safe-mode fallbacks are non-negotiable.
Physics, sensors, and calibration drift — the on-board stack must be re-programmable in flight (as Φ-sat-2 showed).
The argument is not about glaciers, and not about any single mission. It is that a method I developed for one narrow scientific problem is, at its core, a general control loop for keeping any model honest against reality — and that two maturing technologies, commercial tasking and on-board edge computing, are removing the latency that has kept that loop slow. The same design drops into agency missions at NASA or ESA and into commercial constellations alike, wherever timeliness and the ability to act on an observation are worth paying for.
Tier 1 is no longer a proposal. A commercial operator is already extracting the observable in orbit and downlinking the product instead of the picture. What has not yet been built is the step that follows: letting the model and the assimilation ride along, so the satellite does not just report what it sees but updates its own forecast and decides what to look at next. The assimilation math is ready. The hardware has arrived. The missing piece is putting them in the same box.
ESA Φ-sat-1 — first on-board AI on an Earth-observation satellite: esa.int
NASA Autonomous Sciencecraft Experiment (EO-1) — on-board detection & autonomous retasking: ntrs.nasa.gov
Planet Pelican-4 — AI object detection run on board with NVIDIA Jetson Orin (March 2026): Via Satellite
Commercial tasking constellations: Planet Labs · BlackSky
Foundational work in my own domain: Bayesian data assimilation on an Arctic glacier (Journal of Glaciology, 2025)