A small Earth-observation satellite in low orbit can talk to the ground only for a few minutes per orbit, at rates around 1 Mbps on a CubeSat. A satellite that images a flood may therefore hold that data for an hour before it can send any of it down. Neighbouring satellites in the same constellation may reach a ground station much sooner, and many constellations now have inter-satellite links (radio or laser connections between spacecraft). This proposal takes an idea from coded distributed computing, the line of work that adds redundant computation so a cluster tolerates slow machines. The capturing satellite compresses an urgent scene into neural features, splits them into $k$ pieces, and encodes these into $n>k$ coded pieces. It hands the coded pieces to $n$ neighbours. The ground recovers the scene once any $k$ of them arrive, from whichever satellites happen to pass over a station first.
The claim is narrow and testable. The coding should pay off only when the timing of future downlinks is genuinely uncertain. Sources of that uncertainty include link failures between satellites, several satellites competing for one ground station, weather fading the ground link, and power-driven scheduling on board. Under those conditions, $n/k$ redundancy should buy a predictable cut in delivery time, at downlink cost near the compressed payload rather than a full copy per satellite. When pass schedules are fully predictable, plain routing to the satellite with the earliest pass should match it. Mapping where that boundary lies is half the contribution.
Why this mapping, and not the obvious one
The research file suggests the obvious intersection: run the XOR-based coded shuffling of Compressed Coded Distributed Computing across inter-satellite links. In that scheme, nodes broadcast XOR combinations that several receivers decode using data they already hold. We think this mostly fails to transfer. Its gain comes from many receivers each holding different side information. The ground segment, however, behaves as one receiver, because ground stations are joined by cheap terrestrial networks. Satellites rarely need all-to-all exchange; they need to get bits to Earth.
The piece of the coded-computing literature that does transfer is the straggler side. A Unified Coding Framework for Distributed Computing with Straggling Servers uses MDS codes (codes where any $k$ of $n$ coded pieces suffice) and stops waiting after the first $q$ servers finish. This yields a continuous tradeoff between latency and communication. It reports that in an 18-server cluster, doubling the allowed latency roughly halves shuffled data. We reinterpret each "server" as a satellite's upcoming ground pass, and "finishing" as that pass succeeding. Writing $t_i$ for the time coded piece $i$ reaches the ground (relay hop, then wait for the host's pass, then transmission), the delivery time is the $k$-th smallest arrival:
$$T_{\text{coded}}(n,k) = t_{(k:n)}, \qquad \text{downlink overhead} = n/k .$$
That paper supplies the method. It also constrains our design: its scheme assumes independent, homogeneous servers, whereas pass times in a constellation are correlated by orbital geometry. We must test whether the order-statistic gain survives that correlation.
The payload comes from the downlink-compression literature. FOOL compresses early, task-agnostic features of a pretrained network on 15W Jetson hardware. It reports a 51% lower bitrate than prior feature-compression methods with under 1% mAP loss on DOTA-2, SpaceNet-3 and xView. We reuse FOOL-style features so that coded pieces are small, and we keep its single-satellite pipeline as a baseline. FOOL motivates the work by stating that it treats each satellite in isolation. COSMIC is an alternative codec: a light on-board encoder paired with ground-side diffusion reconstruction. Its documented hallucination risk below 0.1 bits per pixel is why we prefer feature payloads whose fidelity can be checked against downstream task metrics.
Triage decides which scenes count as urgent. One option is the flood segmenter of Flood Detection On Low Cost Orbital Hardware, which turns 12-bit multispectral frames into 2-bit masks for a 100× volume cut. Another is the uplinked-embedding retrieval of Embedding-Only Uplink for Onboard Retrieval, which reports Recall@5 of 1.00 on hazards with under 700 bytes of telemetry. We treat triage as given and do not innovate on it.
The delivery-time target comes from EO-ALERT (Kerr et al.), which reported civil alerts delivered from orbit in under five minutes. Their workload was lightweight alerts. Ours is full scenes, because masks discard the spectral data scientists later need (Towards global flood mapping onboard low cost satellites).
Research Program
We will build a simulator that combines orbits, contacts, and links with an actual compression pipeline. It will model several constellation shapes, from a small CubeSat string to a multi-plane shell, and ground-station networks of varying density. Neighbour links will have configurable failure rates. Each ground pass will be subject to three uncertainties: contention among satellites, random fading, and power-limited duty cycles.
Urgent scenes are WorldFloods tiles flagged by the onboard segmenter and DOTA-2/xView tiles flagged by a detector. Their FOOL-style features are split, MDS-encoded, and spread over $n$ neighbours chosen by predicted pass time. The ground decodes and runs the downstream task.
We will study three things. The first is the latency-overhead curve as a function of $(n,k)$. The second is how correlation between pass times erodes the order-statistic gain. The third is a simple adaptive rule that picks $(n,k)$ from the contact plan's own confidence.
As an extension, we will test graceful coding, in which any subset of pieces yields a coarser but usable reconstruction. This is inspired by Barycentric Coded Distributed Computing, whose decoder improves with every arriving worker response instead of waiting for a fixed threshold. That paper reports a 38% latency gain over polynomial codes, but only on tabular and synthetic data in single-hop edge clusters. Whether this idea carries over to image features is an open question here, not an assumption.
Baselines
The do-nothing reference is bent-pipe FOOL. The capturing satellite compresses with FOOL and waits for its own pass. This is the strongest published single-satellite method, with 51% lower bitrate than prior feature compression and under 1% mAP drop (FOOL). It shares our codec, so any difference is due to spreading alone.
The second baseline is the 2-bit mask downlink of Flood Detection On Low Cost Orbital Hardware, which reports 67.54% IoU with its small network and a 13-minute transmission reduced to seconds. It is the speed ceiling and the fidelity floor.
The strongest competitor is uncoded contact-plan routing. The whole compressed payload goes over neighbour links to the satellite with the earliest predicted pass, and is re-routed if that pass fails. We found no reported result on Lacuna for inter-satellite routing delivery latency, so we will implement it ourselves with the same simulator, payload, and link budget. It gets equal effort on tuning its re-routing timeout.
We add two ablations. Uncoded splitting divides the payload across $k$ satellites with no redundancy, which isolates parallel use of passes from robustness. Replication sends full copies to two or three satellites, which isolates the value of coding against naive redundancy at matched overhead.
Beating the baselines means the following. At matched total downlink bits, the coded scheme must show a higher fraction of urgent scenes fully delivered within five minutes than contact-plan routing and replication. Task metrics must sit within FOOL's sub-1% mAP loss and near the masks' IoU-equivalent on WorldFloods.
Why not just…?
Why not just route everything to the satellite with the next known pass? Orbits are predictable, so if pass outcomes are too, this baseline should tie or win. It costs no redundancy. This is the main threat. Our experiments sweep link and pass failure rates from zero upward and report the crossover point. If the crossover only appears at unrealistic failure rates, we will report that negative result and drop the coding claim.
Why not just downlink the flood mask? It is fast and already flown, but it permanently discards spectral data (Towards global flood mapping). The authors also note it over-predicts water in cloud shadow. Our scheme can send the mask immediately and the coded features alongside it. We measure whether the features arrive soon enough to matter.
Why not just replicate the payload to two neighbours? At low redundancy, replication is simple and may perform about as well. The order-statistic argument predicts that coding pulls ahead as $n$ grows at fixed overhead. The matched-overhead ablation tests exactly this.
Why not just compress harder on one satellite? Harder compression with diffusion decoding risks invented textures (COSMIC). It also does not change the wait for the capturing satellite's own pass, which dominates delivery time in sparse ground networks.
Proposed Experiments
The decisive check is a two-dimensional map: deadline-hit rate as a function of pass and link failure probability and ground-station density, for coded, routed, replicated, and bent-pipe delivery, all at equal downlink bits. The proposal survives if a realistic band of that map exists where coding wins clearly.
Secondary checks address three questions. The first is whether correlation in pass times between neighbours, measured in the simulator, shrinks the gain the independent-arrival formula predicts. The second is whether decoded features preserve WorldFloods IoU and DOTA-2/xView mAP relative to undecoded FOOL. The third is whether graceful coding delivers usable coarse products before $k$ pieces arrive. Coding and decoding cost will be timed on a Jetson-class device, to confirm the scheme fits the 15W budget that FOOL already meets.
References
- Compressed Coded Distributed Computing
- A Unified Coding Framework for Distributed Computing with Straggling Servers
- FOOL: Addressing the Downlink Bottleneck in Satellite Computing With Neural Feature Compression
- COSMIC: Compress Satellite Image Efficiently via Diffusion Compensation
- Flood Detection On Low Cost Orbital Hardware
- EMBEDDING-ONLY UPLINK FOR ONBOARD RETRIEVAL UNDER SHIFT IN REMOTE SENSING
- Advanced Data Chain Technologies for the Next Generation of Earth Observation Satellites Supporting On-Board Processing for Rapid Civil Alerts
- Towards global flood mapping onboard low cost satellites with machine learning
- Barycentric Coded Distributed Computing With Flexible Recovery Threshold for Collaborative Mobile Edge Computing