Species-Specific Acoustic Signatures Inside EchoQuilt Surveys

When Two Myotis Species Share the Same Ceiling

A mixed hibernaculum in southern Ohio holds Indiana bat (Myotis sodalis) clusters on the east wall and a scatter of NLEB on bedding-plane overhangs 8 meters north. A biologist walking the room in March cannot reliably tell them apart from the floor without a ladder, and the winter acoustic files from the nearest SM4BAT are not an automatic fix. A PLOS ONE review of automated echolocation classifiers reported classifier accuracy varies significantly by genus, with Myotis calls the most frequently confused group. NPS coverage of Niobrara bat acoustic monitoring notes that software IDs by call frequency and shape, but expert review remains necessary for species-level calls.

The practical problem inside a hibernaculum is that species-specific signatures — the call shape, the inter-pulse interval, the wing-beat envelope of approach and emergence — have to thread through reflections, reverberations, and nearby bats' masking activity. A single AudioMoth node cannot always place a chirp in the right species box, let alone the right cluster. The whole reason to move past node-level ID is that a hibernaculum is a geometric problem, not a single-file problem.

How EchoQuilt Stitches Species Signatures Into Patches

EchoQuilt treats species ID as a property of the patch rather than of the file. Each patch is a reconstructed ceiling or wall tile stitched from overlapping arrivals at multiple nodes. A chirp that registers at three nodes — with known positions relative to the ceiling landmark set — gets a 3D arrival point, a species-signature classification, and an uncertainty score. The patch inherits the species distribution of the chirps that fell on it.

Sonobat's Eastern US acoustic reference table provides the reference metrics the classifier draws from — characteristic frequency, maximum frequency, slope, and duration for Myotis and Perimyotis species. Open-source PAMGuard workflows described in Methods in Ecology and Evolution provide the deep-learning building blocks that modern pipelines (EchoQuilt included) run call-by-call classification against. Inside a cave, reflection off karst fluting can clip the upper end of a tri-colored bat (Perimyotis subflavus) chirp just enough to look like a low Myotis call, so the patch-level stitching has to weight multi-node consensus rather than trusting any single node's ID.

The quilt metaphor is literal here. Imagine each patch as a fabric square colored by the dominant species of chirps that landed on it; overlapping patches share stitched seams where classifier disagreement gets reconciled across sensors. The seam discipline matters: if node 3 has the clearest line-of-sight to the east wall but node 2 is closer, a naive loudest-wins vote will color the east wall patches with whatever species is near node 2. EchoQuilt's stitching resolves by distance-weighted arrival time rather than loudness, which is the same discipline that made swarming chamber signatures readable when pre-hibernation swarming compresses thousands of calls into a few minutes.

A 2022 paper by Bergmann and colleagues on the soundscape of swarming showed that an LFCC (linear frequency cepstral coefficients) classifier identified swarming Myotis calls with 100 percent accuracy in their test set. That result does not generalize — classifier performance drops when species repertoires overlap and when recordings come from highly reverberant ceilings — but the underlying point holds: species signatures exist, and stitching them across sensors is how the hibernaculum's quilt carries species identity from the entrance to the deepest chamber.

The cross-niche parallel is noise profile signatures used in cave-diving rebreather noise mapping. Different rebreather models produce distinct acoustic fingerprints, and the same distance-weighted stitching logic applies: a clean fingerprint at one hydrophone cannot override an ambiguous one at two others.

The mockup shows a species-signature panel pulled from a winter survey at a mixed hibernaculum. The top half is a spectrogram stack — Indiana bat calls in one band, NLEB in a near-overlapping band, with a confidence ribbon below each. The bottom half is the 3D ceiling patch with the species distribution painted on it. The visual immediately shows that the north wall is NLEB-dominant while the south wall is mixed Indiana and NLEB, which a single-node audio file would not resolve.

Advanced Tactics for Species-Specific Quilts

Three tactics sharpen species-specific mapping inside EchoQuilt. First, use wideband nodes where possible (SM4BAT records to 500 kHz at 384 kHz sample rates) so that high-frequency Perimyotis calls do not get truncated below the discriminating slope. Second, anchor to the ceiling landmark set with tight positional tolerance (under 3 cm between winters) so that patch-level species assignments from year 1 are comparable to year 2 without re-training the classifier. Third, run a matched-pair comparison between your pre-swarm acoustic stream and a summer mist-net capture from the same system — the bioacoustics.info cave passive acoustic guidance describes AudioMoth cave-entrance deployments that pair with capture data for exactly this calibration.

Downstream, the species-assigned patches feed advanced counting workflows. Cluster counts inside a mixed hibernaculum only make population sense when each cluster carries species identity, and the patch-level stitching is how that identity travels from chirp to count.

Fourth, retrain classifiers annually against an expanding regional reference library. The Sonobat Eastern US table covers established species at population-mean parameters, but a specific Ohio karst hibernaculum may host a Myotis lucifugus subpopulation whose calls trend 1-2 kHz lower than the regional mean due to founder-effect or local habitat acoustic adaptation. EchoQuilt's classifier-retraining workflow ingests verified species-tagged calls from the site itself, building a site-specific classifier layer on top of the regional model. Cross-classifier validation against Kaleidoscope Pro and SonoBat outputs catches systematic disagreement that flags miscalibration on either side.

Fifth, embed call-confidence thresholds in downstream analyses. A patch where 90 percent of resident calls classify with confidence above 0.85 is reliably species-tagged; a patch where 60 percent of calls fall in the 0.55-0.7 confidence band is uncertainty-flagged and should not feed a deterministic species count. EchoQuilt's analysis layer enforces this distinction by carrying confidence as a first-class field rather than collapsing it into a binary tag at the patch boundary.

Sixth, validate species assignments against USGS Bird Banding Lab band recovery records for the rare cases where a banded individual is recovered at a site with concurrent EchoQuilt coverage. The band recovery establishes ground truth for the patch the recovered individual occupied; the corresponding patch's call assignments either confirm or conflict with the species ground truth. Even one validation event per winter improves the classifier's calibration meaningfully when accumulated across years.

Seventh, distinguish species-level confusion from species-level co-occurrence. Two clusters of similar Myotis species occupying adjacent patches will produce overlapping call detections at any node positioned between them, which a naive classifier may report as a single mixed-species cluster when it is actually two adjacent single-species clusters. EchoQuilt's spatial deconvolution accounts for this by modeling the spatial decay of call detection probability with distance from the source patch, separating "patch A is single-species" from "patches A and B both project calls onto detector C."

Eighth, contribute species-tagged patch records back to the NABat acoustic data archive so the broader community benefits from each site's classifier-validation work. Patch-level species records with confidence metadata are richer than the file-level records that NABat currently aggregates, and a federated patch archive could substantially raise the analytical floor of the program through cross-site classifier sharing.

Bring Species-Accurate Quilts to Your Mixed-Species Hibernaculum

NABat partners and WNS response teams working mixed Indiana bat and NLEB hibernacula need species-level clarity that a single AudioMoth cannot reliably deliver alone. EchoQuilt's distance-weighted stitching threads species signatures through the whole cave, so each cluster on your winter quilt carries a species tag with an explicit uncertainty score. Sites with mixed Myotis sodalis and Myotis septentrionalis cluster zones, mixed Perimyotis subflavus and Myotis lucifugus chambers, or mixed Corynorhinus townsendii and Myotis californicus colonies all benefit from the patch-level species resolution that single-sensor approaches cannot match. The patch-level species records integrate directly into NABat's data submission pipelines, the USGS BatAMP acoustic data portal, and state DNR Section 7 consultation deliverables. Join the Waitlist for Hibernacula Biologists and tell us which of your sites routinely defeats single-sensor classifiers.

We will scope a multi-node deployment tuned to your species list, with classifier retraining scheduled against your existing capture-and-band field calendars so the calibration data builds organically from your existing operations.