Separating Bat Echolocation From Mapping Signal Sources
When the Bats Shout Into Your Map
A February 2024 mid-torpor lull in an Ulster County hibernaculum produced an unexpected spike in mid-frequency audio across a four-sensor EchoQuilt array. For 45 minutes, the acoustic quilt showed geometry errors bloom across the cluster-proximate voxels. Cause analysis identified a single Myotis lucifugus waking, briefly flying, and generating roughly 220 echolocation calls from 40 to 60 kHz before re-roosting. The calls had cross-talked into the geometry-inversion math because the frequency range overlaps the drip-reflection band the inversion relies on. Without proper separation, one aroused bat degrades a full day of mapping data.
The challenge is real. Filter accuracy for bat echolocation identification varies 41-85% depending on call libraries and acoustic filters, meaning even species ID — a simpler problem than signal separation — is hard. And yet both problems share sensors. The same array that catches ambient drip and airflow for geometry catches every wing-beat and every echolocation pass. The question is not whether to install a bat detector alongside a mapping array but how to let the two datasets inform each other without corrupting either.
The frequency overlap is the source of all the trouble. Drip impacts on flowstone produce broadband transients that contain energy from a few hundred Hz up to 80 kHz, with peaks in the 30-50 kHz range that map directly onto the call bands of Myotis lucifugus (centered around 40-45 kHz), Myotis sodalis (40-45 kHz), and tri-colored bat (40-50 kHz). Airflow rumbles add narrowband energy at lower frequencies that occasionally beats against echolocation harmonics. A naive geometry inversion that treats every transient as a drip and every narrowband event as airflow will pull cluster-proximate calls into the geometry math and produce ghost reflections that warp the cluster boundary. Equally bad, a naive call detector that treats every drip as a candidate echolocation pass will inflate species counts and corrupt the NABat record.
Solving either problem in isolation makes the other worse, and the real solution requires both pipelines to negotiate the contested time-frequency cells in the spectrogram.
Splitting the Quilt Into Two Cooperating Layers
EchoQuilt runs a two-layer signal-separation pipeline that patches echolocation data and mapping data into cooperative rather than competing streams. The bat-call layer extracts calls using a deep-learning classifier trained on NABat reference libraries; the mapping layer inverts ambient reflections into geometry after subtracting the classified calls. The two layers exchange information: confirmed bat positions from call triangulation sharpen the occupancy layer, and stable geometry from the mapping layer reduces multipath ambiguity in call localization.
The classifier core uses deep learning reaching 90-92% accuracy on standard species-ID tasks, augmented with the BatDetect2 detection and classification pipeline for time-frequency localization. On hibernaculum data the species set is narrower than in summer surveys — Myotis lucifugus, Myotis sodalis, tri-colored bat, NLEB, and occasional Townsend's big-eared bat or Eptesicus fuscus dominate — so classifier accuracy runs higher, and the separation math benefits. The quilt metaphor is literal: each second of audio gets stitched into the geometry quilt, the call quilt, or both, with a confidence weight that reflects how clean the separation was.
For operational hibernaculum work, the payoff is a single array producing three co-registered datasets. First, a species-resolved call log that feeds directly into NABat reporting via the USGS NABat Acoustic ML classifier. Second, a call-position dataset that shows where individual bats arouse, fly, and re-roost within the hibernaculum across the winter. Third, a geometry and occupancy quilt that is no longer corrupted by call cross-talk. Compared to established workflows that use SonoBat or Kaleidoscope HMM approaches as separate downstream tools, EchoQuilt unifies the detection, classification, and mapping onto a shared time base.
Integration matters for biologists. The need for geometric inference from ambient sound rather than direct visual survey grows every year as WNS-affected species lose numbers. But geometric inference fails if echolocation calls are naively treated as noise. The EchoQuilt pipeline closes that loop, and it pairs cleanly with species signatures work for refining the classifier against site-specific call libraries.
Advanced Tactics for Clean Separation
Three tactics raise separation quality from adequate to robust. First, pre-train the classifier on the site's own swarm-season audio before hibernation begins. Generic NABat models work, but a site-tuned model catches local call idiosyncrasies and dialectal variation that otherwise show up as mis-classifications. A week of September swarm audio typically yields 8,000 to 40,000 clean call examples across the local species set, which is enough to fine-tune the classifier to local conditions. The fine-tuning also captures the cave's acoustic signature — its specific reverberation profile, its drip cadence, its airflow band — which makes the classifier markedly more robust against site-specific false positives that generic models routinely flag.
Second, run the separator in forward and backward modes. Forward mode subtracts detected calls from the mapping signal; backward mode uses stable mapping geometry to reject false positives in the call detector. The two directions catch different failure modes, and combining them reduces both map corruption and species-ID error. The bidirectional approach also exposes ambiguous time-frequency cells — places where neither pipeline can confidently claim a transient as call or as drip — and EchoQuilt routes those cells to a quarantine layer that human reviewers can inspect during quarterly QA. Quarantine cells are typically less than 0.1% of total audio in a well-tuned site, but they catch the edge cases that automated pipelines miss and feed corrections back into the next iteration of the classifier.
The bidirectional separator works particularly well alongside Anabat detector pairings for teams bringing in established detector hardware, since the detector's high-sensitivity classifications give the separator a strong prior on which transients to treat as calls.
Third, preserve full-bandwidth raw audio for a rolling 30-day window. The separator's confidence depends on model version and training data, and both will improve over the life of a long project. Keeping 30 days of raw audio means biologists can rerun the separation with improved models without losing the mapping data. Budget for this: a four-sensor array at 384 kHz produces roughly 3.3 GB per sensor per day, so 30 days across four sensors is about 400 GB of storage — manageable with a small ruggedized drive at the field site.
A fourth tactic is to publish per-cave separator performance metrics alongside the species counts. NABat aggregations are most useful when downstream consumers know how confident the upstream classifier was at each site. EchoQuilt logs precision, recall, and confidence distribution for the separator at each hibernaculum across each winter, and these metrics travel with the data into NABat submissions. State coordinators can see at a glance which sites have reliable acoustic data and which need additional human review before trend analysis. The transparency is uncomfortable in the first year of a deployment, when novel sites typically score lower, but it improves quickly as the site-tuned classifier matures across multiple winters.
A fifth tactic addresses the multi-array case. Some hibernacula warrant more than one sensor cluster — a maternity-adjacent throat array plus a back-chamber cluster array, for example, or separate arrays for parallel passages. EchoQuilt's separator runs independently per array but cross-references confirmed call positions across arrays, which catches cases where a bat flew between arrays during a single arousal episode. The cross-reference also adds a sanity check on time synchronization across the multi-array deployment: if the same call appears at both arrays with a propagation delay that does not match the physical geometry, one array's clock has drifted and needs correction before the next analysis pass. Cave-diving surveyors face a structurally similar problem with noise source interference from rebreather bubble streams, and the same separation math applies across both niches.
These five tactics collectively turn a noisy, contested signal stream into a clean dual-data product that biologists can defend in front of regulators and peer reviewers.
Get Early Access to EchoQuilt
If you run hibernaculum acoustic surveys and are tired of reconciling SonoBat output, Kaleidoscope output, and separate LiDAR geometry on uncoordinated time bases, EchoQuilt's unified pipeline collapses all three into a single co-registered dataset. The system is especially suited to WNS field teams and NABat participants who already rely on automated species classification and want mapping data from the same array. Each pilot kit ships with a hibernaculum-specific calibration set tuned against your existing Anabat detector libraries and Kaleidoscope reference call sets, an AudioMoth-compatible recording configuration that drops into existing detector arrays, and a passive acoustic logger array sized to your chamber count, swarm-season throat geometry, and back-cluster spatial spread.
The site-tuned classifier ingests a week of September swarm audio to capture local call idiosyncrasies for Myotis lucifugus, Myotis sodalis, NLEB, tri-colored bat, and Eptesicus fuscus before mid-torpor windows open, and the bidirectional separator's quarantine layer routes ambiguous time-frequency cells to your QA queue with full provenance. Pilot biologists shape the cross-array call-position cross-reference rules and the per-cave precision-recall reporting format that the 2027-28 NABat reference release will adopt. Priority slots go to USFWS Section 7 consulting teams, multi-state surveillance partnerships covering Priority 1 hibernacula with decade-long visual count records, and IUCN-aligned WNS researchers running multi-year arousal-cascade datasets that can validate the wing-beat detector against simultaneous infrared camera observations. State NABat coordinators receive an additional precision-recall dashboard sized to a 40-site portfolio so cross-site classifier-confidence can travel into trend analyses rather than disappearing in aggregation.
Join the Waitlist for Hibernacula Biologists to get a pilot sensor kit for the 2026-27 winter and help us tune the separator against your site's local call library.