Distinguishing Wind-Driven From Structural Sound in Open Skylights
Why the Aeolian Wind Floor Almost Kills Open-Skylight Cave Mapping
An acoustic payload sitting on the rim of an Arsia Mons lava tube skylight records a spectrum that mission planners and atmospheric scientists have been studying together for years. AGU/Wiley's analysis of wind and turbulence observations with the SuperCam microphone on Perseverance reports 25 kSPS sampling across 20 Hz to 50 kHz, with turbulence features in the 20 Hz band that dominate the low-frequency end of the spectrum and swing in amplitude over the course of a sol. The Springer description of the SuperCam instrument suite on the Mars 2020 rover gives the microphone's 100 Hz-10 kHz design band, and the PMC report on the Mars soundscape shows that 5 hours of surface recordings are dominated by turbulence events even at the relatively quiet Jezero crater landing site.
The physics is well understood. Aeolian sound arises when air passes obstacles at the right Reynolds number, and a skylight rim is geometrically ideal for generating aeolian tones. HAL's experimental wind characterization with the SuperCam microphone under simulated Martian atmosphere pins down the ground-truth relationship between microphone output and low-pressure wind. The structural signal (cave wall reverberation, thermal ticks, rover vibration carried through regolith) sits beneath this floor at typical Martian wind conditions. A mapping payload that cannot separate them produces quilts that drift with the weather.
The drift is not a small effect. In analog campaigns at Surtshellir under variable Icelandic winter wind, a naive classifier produces patch geometry estimates that swing by tens of meters between low-wind and high-wind recording windows. A reviewer who sees that variance immediately questions the underlying inference, even when the geometric ground truth is stable. EchoQuilt's two-stage separator was specifically designed to suppress this kind of weather-driven drift so that the inferred geometry remains stable across the kinds of atmospheric variability flight teams expect at Marius Hills, Arsia Mons, and Mare Tranquillitatis candidate sites. The same physics applies on the Moon at lower magnitude because solar-wind and micrometeorite interactions still drive transient surface vibrations that look like noise to a naive classifier; moonquake signal fusion provides an alternative excitation source that sidesteps the wind-coherence problem entirely on bodies without an atmosphere.
How EchoQuilt Isolates Structural Signal From the Aeolian Wind Floor
EchoQuilt runs a two-stage separator. The first stage is a wavelet-packet decomposition borrowed from the Methods in Ecology and Evolution wind-robust sound event detection work, which separates broadband turbulence from transient structural events by time-frequency localization. Aeolian sound produces long, low-frequency, spatially diffuse signatures; structural reverberation produces short, broadband, direction-coherent signatures. The wavelet-packet step cleans the first-order wind contamination before any geometry inference runs, which means the downstream classifier never has to relearn the wind contribution from raw recordings.
The second stage is a direction-of-arrival gate that uses the passive microphone array's spatial diversity. Wind-driven signal arrives broadly and incoherently across the array; structural reverberation arrives coherently from the direction of the reflecting surface. EchoQuilt's pipeline tags each acoustic event with a coherence score and suppresses patches that are stitched from low-coherence events. The resulting quilt stays stable across a sol that includes Martian afternoon wind gusts, and the system marks uncertainty honestly when wind conditions exceed the classifier's confidence threshold rather than silently producing low-quality geometry.
The coherence-gate approach also produces a useful diagnostic byproduct. The fraction of events that pass the gate per recording window is itself a measurement of local atmospheric quietness, and over a long campaign this fraction maps to known meteorological patterns. PIs running parallel atmospheric science observations can use the EchoQuilt pass-rate trace as an independent measurement of local boundary-layer dynamics, without dedicated wind sensors. This kind of cross-instrument byproduct is exactly the sort of additional science return that helps a payload defend its slot on a contested Phase B integration timeline.
A companion post on skylight acoustics entry planning establishes the baseline geometry inference that wind separation protects, and the patch confidence model depends directly on the classifier's coherence threshold.
The classifier's threshold also has implications for science observation planning. A team that wants to capture an aeolian event for atmospheric science can flip the gate logic and accept low-coherence diffuse events while suppressing high-coherence structural ones. This dual-mode operation lets a single deployment serve both the cartography campaign and parallel atmospheric science objectives without additional hardware. PIs running mixed science packages on Mars analog campaigns appreciate this flexibility because it spreads the campaign cost across more proposal lines and produces a richer dataset from each EVA window, which often translates into stronger Phase B reviews when funding is contested across multiple SMD lines.
Advanced Tactics for Sustained Skylight Mapping
Three tactics keep the quilt honest during sustained skylight observation. First, log a wind-regime index alongside the quilt so downstream analysts can correlate patch uncertainty with atmospheric state. The PMC Mars soundscape work showed wind regimes shift predictably over a sol, and a quilt that carries its wind index as metadata lets analysts re-weight patches during post-processing. This matters most for the first analog campaigns when the classifier is still being tuned against truth.
Second, schedule targeted recording windows around local minimum wind periods. On Mars, these typically fall in the hours around local sunrise; on the Moon, wind is not the constraint but thermal-contraction bursts favor the local dawn or dusk terminator. Concentrating stitching activity in these windows produces denser, lower-uncertainty quilt patches for the same energy budget. The traverse can be longer without a larger energy cost because passive recording continues in any weather.
Third, maintain a rejected-event archive rather than discarding low-coherence events outright. Events that fail the coherence gate sometimes carry useful information about wind regime or skylight geometry once the classifier is retrained with more data. A reference archive built during early analog campaigns at Corona or Surtshellir reduces the reclassification cost for later flight concepts significantly.
A fourth tactic that benefits long-duration observation campaigns is to run a periodic re-baselining pass on the wind regime model, drawing on the inverse problem bat hibernacula teams solve when looking for airflow sound pockets where stable airflow patterns are the signal rather than the noise. Atmospheric conditions at an analog site shift seasonally, and a classifier tuned in winter at Surtshellir can drift out of calibration by summer. EchoQuilt's pipeline supports a quarterly re-baselining workflow that consumes a few hours of audio at a known geometry and updates the classifier weights without invalidating prior quilt patches. This keeps the long-term quilt geometry consistent across seasonal cycles, which matters for reviewers comparing pre-flight analog quilts against early flight returns from Marius Hills or Arsia Mons.
Without periodic re-baselining, classifier drift can introduce systematic biases that mimic real geometric change, and the resulting confusion takes months of debugging to untangle. EchoQuilt's logging format records the active classifier version with each patch so that any inferred change can be traced back to either a true geometric event or a known classifier update.
CTA
EchoQuilt is looking for flight concept teams operating at open skylight geometries (Marius Hills, Mare Tranquillitatis, Arsia Mons candidates, Hadley Rille analogs, Pavonis pit chains) who need a mapping payload that stays quilt-stable during active wind regimes and across seasonal classifier drift over multi-year analog campaigns. We ship the wavelet-packet separator, the coherence gate, the wind-index metadata logger, the seasonal re-baselining workflow, and an analog tuning kit for Surtshellir-class field sites. Join the Waitlist for Planetary Analog Researchers to align the classifier with your target wind regime and analog atmospheric profile before a campaign so the first week of recordings produces signal-grade data instead of training data. PANGAEA instructors, NIAC PIs, and JPL Mars cave concept leads get priority configuration support and direct access to our analog field engineers.