Earth Observation Research

Space-Based Data Processing

Edge compute on orbit, and why dimensionality matters.

The Bandwidth Problem

A single Earth observation satellite generates 1–5 TB of raw imagery per day. A constellation of 100 satellites produces 100–500 TB/day. The downlink bandwidth available during a ground station pass is typically 1–2 Gbps for 8–12 minutes, yielding roughly 60–180 GB per pass. The math does not work: satellites generate far more data than they can transmit.

ConstellationSatellitesDaily DataDownlink CapacityGap
Planet (SuperDove)~200~30 TB~5 TB6x
Maxar WorldView4~2 TB~0.8 TB2.5x
Pixxel Fireflies6 (growing to 24)~8 TB~1 TB8x
Future mega-constellations500+~500 TB~20 TB25x

The traditional solution is selective tasking: only image what a customer has requested. But this throws away the global monitoring capability that constellations are designed for. The alternative is processing on-orbit -- running ML inference on the satellite itself and transmitting only the results.

Current State of On-Orbit Processing

Several companies are deploying or testing ML inference hardware on satellites:

CompanyApproachHardwareStatus
OrbitsEdgeHardened edge compute module (SatFrame)HPE Spaceborne Computer-2 heritageISS-tested; commercial deployment
PixxelOn-board ML for hyperspectral classificationCustom FPGA + ARM pipelineFirefly constellation operational
SatellogicOn-board cloud detection, change detectionXilinx FPGA acceleratorsOperational on 30+ satellites
Ubotica (now Dott)CogniSAT AI processor for ESA missionsMyriad X VPU, radiation-hardenedFlown on PhiSat-1 (ESA)
XploreEdge ML for autonomous taskingNVIDIA Jetson (rad-tolerant packaging)Development

The ground segment is also evolving. AWS Ground Station provides ground station-as-a-service, and Azure Orbital offers similar capabilities. These services reduce the latency from downlink to processing but do not solve the bandwidth bottleneck -- they just move the processing from the customer's data center to the cloud provider's.

The Embedding Connection

This is the key insight. If you are doing ML inference on a satellite with limited compute, limited memory, and limited downlink bandwidth, you need compressed embeddings. But our dimensionality research shows that naive PCA compression to 16 dimensions destroys domain-specific features. The question for space-based processing is: what is the minimum dimensionality that preserves the distinctions that matter for your specific mission?

Consider two satellite missions:

A single global PCA basis optimized for "variance explained" would keep the dimensions that separate land from water from cloud (the top principal components) and discard the dimensions that separate wheat stress from corn stress, or tanker from container ship. Both missions would fail at their specific tasks while "performing well" on a generic benchmark.

Task-Specific Compression

The solution is not to find the "right" global dimensionality. It is to train task-specific projections that preserve the features relevant to each mission.

MissionCritical FeaturesMin. Dims (Estimated)PCA Would Destroy
Crop species classificationRed-edge slope, chlorophyll indices, phenology64–128Inter-species spectral differences
Crop stress detectionNDVI anomaly, water stress indices, fluorescence128–256Stress-type discrimination
Mineral mappingSWIR absorption features (narrow, diagnostic)256+Mineral-specific absorption bands
Ship detection (SAR)Backscatter intensity, shape, wake32–64Fine vessel type distinctions
Change detectionTemporal difference features32–64Gradual vs. sudden change
Cloud maskingBrightness, texture, thermal contrast8–16Nothing -- this is the easy task

Cloud masking works at 16 dimensions because it is an inter-domain task (cloud vs. not-cloud) -- exactly the kind of task that PCA compression handles well. Mineral mapping requires 256+ dimensions because it is an intra-domain task (distinguishing kaolinite from montmorillonite from illite within the broader category of "clay minerals"). This is the same pattern documented in our dimensionality research for legal terms.

The Matryoshka Approach for Space

Matryoshka Representation Learning (MRL) trains a model to produce useful embeddings at multiple dimensionalities simultaneously. The first 32 dimensions are optimized to be useful on their own. The first 64 include those 32 plus 32 more. And so on up to the full 768 or 1024 dimensions.

For space-based processing, MRL offers an elegant solution:

  1. Train on ground: train the Matryoshka model with domain-specific data for the satellite's mission (crop monitoring, maritime, etc.)
  2. Deploy on orbit: run inference using the full model, producing full-dimensional embeddings
  3. Truncate for transmission: transmit only the first k dimensions, where k is chosen based on the mission's precision requirement and available bandwidth
  4. Reconstruct on ground: use the truncated embedding for immediate queries; request full embedding for high-precision analysis during the next downlink window
Bandwidth Reduction Transmitting 128-dimensional float16 embeddings instead of 768-dimensional float32 embeddings achieves a 12x bandwidth reduction. Even at 128 dims, task-specific Matryoshka embeddings preserve the intra-domain distinctions that matter for the mission. The satellite transmits embeddings (vectors of numbers) rather than raw imagery (pixels), and the ground system queries those embeddings to decide which raw imagery to request on the next pass.

The Bandwidth Math

What's TransmittedSize per Image TileTiles per PassTotal per Pass
Raw imagery (10 bands, 16-bit)~200 MB~300~60 GB
Compressed imagery (JPEG2000)~40 MB~300~12 GB
Embeddings at 768d (float32)~3 KB~300~900 KB
Embeddings at 128d (float16)~256 B~300~75 KB
Embeddings + metadata + alerts~1 KB~300~300 KB

The difference is five orders of magnitude. Embedding transmission fits in the margins of telemetry bandwidth that every satellite already has. The question is not whether to transmit embeddings -- the question is whether the embeddings preserve enough information to be useful.

Federated Learning Across Constellations

As satellite constellations grow, a new paradigm emerges: federated learning across orbiting nodes. Each satellite trains (or fine-tunes) a local model on its observations. Inter-satellite links (ISLs) allow gradient sharing between constellation members without transmitting raw data to ground.

The challenges are substantial:

Cross-Satellite Vector Search

A tantalizing application: maintain a distributed vector index across the constellation. Each satellite holds embeddings for its recent observations. A query from ground ("find all images of this port from the last 48 hours") is broadcast to the constellation, each satellite searches its local index, and the top results are transmitted as metadata (embedding + coordinates + timestamp). Only the relevant raw imagery for the top matches is downlinked.

This transforms the satellite constellation from a data collection system into a queryable sensor network -- a spatial search engine with global coverage and sub-daily revisit.

The Dimensionality Decision Is a System Design Parameter For ground-based systems, embedding dimensionality is a tuning parameter -- you can always go back and re-embed at higher dimensionality. For space-based systems, it is a design parameter locked in at launch. The model and its output dimensionality are burned into the satellite's software (or at best, updatable via slow uplink). Getting it wrong means years of degraded performance. Our dimensionality research provides the framework for making this decision: characterize the intra-domain discrimination requirements of the mission, measure DCRP at candidate dimensionalities, and choose the minimum that meets the precision threshold.

The Future: Compute in Orbit

Three trends are converging:

  1. Radiation-tolerant AI accelerators: purpose-built chips (not adapted consumer GPUs) designed for inference in the radiation environment of LEO. Ubotica's CogniSAT, Syntiant's NDP, and custom FPGA designs are leading this.
  2. Software-defined payloads: satellites whose processing pipelines can be updated after launch. ESA's PhiSat program and DARPA's Blackjack demonstrate the concept. The model deployed at launch is not the model running three years later.
  3. Optical inter-satellite links: SpaceX Starlink has demonstrated laser ISLs at multi-Gbps. When constellation members can communicate at high bandwidth, distributed processing becomes practical -- the constellation becomes a mesh computer, not a collection of independent sensors.

The constraint that does not change is physics: downlink bandwidth to ground stations remains the bottleneck, and power budgets on satellites are finite. Both constraints favor compact representations -- embeddings rather than raw data, lower dimensionality rather than higher. The research question is how to compress without destroying the distinctions that define the mission.

Related Research

The Dimensionality Illusion
Why PCA compression to 16 dimensions destroys domain-specific features. The foundational argument for why space-based embedding systems cannot use naive compression.
Embedding Research
Earth Observation for Trading
The downstream use case: satellite data driving commodity price signals. The pipeline from orbit to alpha.
Earth Observation
GPS/PNT Authentication
Authenticated positioning and timing. PNT is the clock and coordinate system that makes distributed satellite processing coherent.
Earth Observation