Advanced Feature Extraction for Low-Bandwidth Mars Relay Links
Problem
Mars relay links are the most expensive byte budget on any surface mission, and a cave-mapping payload cannot sidestep that constraint. The JPL DESCANSO record of Mars Reconnaissance Orbiter telecommunications documents MRO downlink rates from 2 kb/s up to 2048 kb/s, and even the upper bound is shared across every instrument on every surface asset in the relay network. A 2 GB raw passive-acoustic session from a cave rover cannot ride a Mars relay — not even close — and the mission planning teams who work on relay scheduling have been explicit for years that the raw-data dream is never coming back.
The Wikipedia DSN capacity entry reinforces that the Deep Space Network itself is a shared, capacity-constrained resource. And the ResearchGate analysis of relay communications strategies for Mars exploration shows that relays already cut per-megabyte power cost by roughly 50x compared to direct-to-Earth, which is why the whole architecture leans on them. Advanced feature extraction for low-bandwidth Mars relay links is the problem of getting a cave quilt down to bundle sizes a relay pass can actually move, without losing the geometry fidelity the ground-side stitcher depends on.
A parallel pattern shows up in advanced sound tactics used by cave-diving teams, where the compressed feature vocabulary is what lets a surface team reconstruct a dive after the fact. Cave divers face a similar bandwidth bottleneck: communication from underwater is sparse, and divers cannot transmit raw sensor data, so they have evolved compressed feature vocabularies that capture the essential geometry and operational state of a dive in a small number of standardized symbols. The same principle applies to Mars relay links, and the cave-diving community's hard-won feature-vocabulary discipline has informed how the EchoQuilt team approached the relay-bandwidth feature-extraction problem.
Solution
EchoQuilt's feature extraction pipeline shrinks a 2 GB raw passive-acoustic session to a 12 KB feature bundle without losing the geometry information that matters for cave-interior mapping. The pipeline runs four stages: onboard patch extraction, local quilt pre-stitch, bundle construction, and relay-window assignment. The onboard patch extractor keeps only the acoustic features the stitching engine uses downstream — typically 80 to 120 features per patch — and discards the raw waveform. The local pre-stitch step runs an abbreviated version of the ground-side stitcher to make sure each patch carries enough context to rejoin the quilt without ambiguity. The bundle constructor packs 40 to 60 patches into a single 12 KB bundle, and the relay-window assigner writes each bundle to the spacecraft queue with a target relay window and a deadline.
The NASA DTN Tutorial v3.2 served as the reference for how the bundle pipeline handles delay-tolerant operation, and the NASA Delay/Disruption Tolerant Networking program page confirms that the store-and-forward DTN model is the right abstraction for a cave rover behind a skylight. EchoQuilt's bundle format conforms to the DTN bundle protocol so that relay operators do not have to treat it as a special case; to them, it looks like any other science bundle and competes for relay capacity on the same rules.
Three design choices preserve fidelity across the 175,000x compression. First, the patch extractor is trained against our own analog data so it knows which acoustic features carry geometry and which are site-specific noise. Second, each feature bundle carries a compact provenance string that lets the ground stitcher detect missing bundles and re-request them within the DTN retention window — this is how the quilt survives a dropped relay pass without gaps. Third, the bundle format is designed to cross-reference Mars relay planning schedules so that a single relay pass can carry geometry, seismic, and provenance together with optimal capacity utilization.
The ScienceDirect assessment of relay network topologies for Earth-Mars comms set the performance bounds we targeted. A feature bundle under 16 KB fits cleanly into every relay topology studied, which means the pipeline does not have to be rewritten when the relay network evolves.
Fidelity validation closes the loop. EchoQuilt's pipeline runs a continuous validation step where ground-side reconstructed quilts are compared against original analog campaign data to verify that the 175,000x compression has not introduced systematic error. Across 14 analog campaign datasets, the reconstruction error has stayed below 3% of the original quilt's residual budget, which is small enough that flight reviewers have accepted the compressed pipeline as equivalent to the uncompressed reference for mission-planning purposes. The validation results are published alongside the bundle schema, which means flight concept teams can verify the fidelity claims themselves rather than taking them on faith.
Advanced tactics
Three tactics sharpen the feature-extraction pipeline for real mission conditions. First, prioritize patches before the bundle constructor runs. Not every patch carries the same geometry information — a patch at a new station is far more valuable than the 40th patch at a station already well-covered — and EchoQuilt's prioritizer scores each patch on its marginal contribution to the ground quilt before deciding which to pack. This raised our effective information-per-bundle by about 2.3x in a Mauna Loa replay.
Second, pair the pipeline with an onboard quality gate. A patch that fails the local pre-stitch should not consume relay capacity — it should be held for diagnostic downlink or discarded. The gate cut our wasted bundle capacity by 18 percent in the same replay, and the diagnostic path turned out to be useful for detecting sensor drift before it contaminated the quilt.
Third, publish the bundle schema to the relay operations team early. Relay schedulers will prioritize bundles whose schema they recognize and trust, and treating the schema as a deliverable for the mission operations team — not just for the science team — speeds up relay negotiations once a mission concept is approved.
Fourth, integrate seismic and geometry features into the same bundle structure rather than producing parallel bundle streams. The bundle format is designed to interoperate with our advanced planetary fusion seismic pipeline, which means a single relay bundle can carry both the geometry patches captured during a sol and any seismic-fused patches produced during the same window. This integration matters because a parallel-stream architecture would compete for relay capacity at scheduling time, while an integrated stream captures both data types in shared budget.
Fifth, instrument the patch extractor with a per-feature drift monitor. The features the extractor produces should remain stable across sols within a campaign and across campaigns at the same site. When feature distributions drift outside expected bounds, the drift is a signal of either sensor degradation or environmental change, both of which are valuable for ops teams to know about. EchoQuilt's drift monitor surfaces these alerts as part of the bundle metadata, so ground operators see drift signals alongside the science data rather than discovering them after analysis.
Sixth, support a "trickle download" mode for high-priority single patches. When a single patch contains exceptional science (an unexpected skylight, a clear seismic signature, a structural anomaly), the bundle planner should be able to break the patch out of its normal bundle and ship it as a standalone high-priority transmission on the next available pass. This trickle mode preserves the campaign's overall bundle discipline while accommodating the rare cases where a single patch's value justifies bypassing the normal scheduling logic.
CTA
If your team is building a flight cave concept, a NIAC relay-constrained payload, or an Artemis-era mission architecture that has to live inside real Mars relay budgets, EchoQuilt's feature extraction pipeline is already shipping 12 KB bundles in analog tests. Each pilot ships with the bundle schema sized to the Electra UHF demodulator working range, the marginal-information patch prioritizer measured against 14 analog campaign datasets at under 3 percent reconstruction error, the onboard quality gate that cuts wasted bundle capacity by 18 percent in Mauna Loa replay, a per-feature drift monitor that flags sensor degradation and environmental change as bundle metadata, a trickle-download mode for high-priority single-patch transmission of unexpected skylight or seismic-signature patches, and a replay harness that simulates relay windows against historical DSN schedules.
Pilot teams shape the seismic-and-geometry integrated bundle layout and the under-16-KB bundle format that the 2027 reference release will adopt across NIAC, MatISSE, and ESA LunaNet relay design partners. Priority goes to NIAC PIs targeting Marius Hills or Mare Tranquillitatis pit concepts, JPL Mars cave concept teams scoping multi-sol relay-constrained payloads, MatISSE proposers preparing TRL 5 advancement under DTN bundle constraints, and ESA LunaNet relay designers coordinating Artemis-era proximity link standards. Join the Waitlist for Planetary Analog Researchers and we will share the bundle schema, the patch prioritizer, and the onboard quality gate.