betaFeedback

WorksComputer Science

ChiroEcho: extending automated bat vocalisation classification beyond the learned taxonomy

Abstract

Bats are key indicators of ecosystem health and are protected throughout Europe, making reliable population monitoring a conservation priority. Their cryptic nocturnal lifestyle makes passive acoustic monitoring essential, yet automated identification remains difficult as echolocation calls vary with behaviour and environment and overlap among species. We present a deep learning framework that jointly predicts species and genus and combines genus predictions with geographic species distributions at inference. When only one species of a predicted genus occurs in a region, the framework can resolve species absent from the learned taxonomy. This reframes geographic information as a means of extending, rather than constraining, a classifier's effective taxonomy. Using recordings spanning 35 European bat species, we evaluate closed-set classification, examine the instability of performance estimates for sparsely represented species, and conduct a controlled held-out proof-of-principle experiment. The rare-species analysis shows how limited evaluation data can obscure species-level performance, while the held-out experiment shows that genus predictions and location can recover labels unavailable to the species head. Geographic resolution extends operational coverage from 35 to 41 of the 48 native European bat species, increasing coverage from 73% to 85%. To our knowledge, this is the broadest operational coverage reported for automated European bat classification. More broadly, the bat framework provides proof of principle for resolving unseen fine-grained classes by combining coarse predictions with transparent external constraints.

Introduction

Bats are ecologically important yet among the least visible inhabitants of European ecosystems. Their nocturnal, cryptic lifestyles [17] often obscure their roles in insect suppression, nutrient transport, and trophic regulation [20, 55, 47].

Many European bat populations have also undergone long-term declines [5], and European bats receive strict legal protection under the EU Habitats Directive [14]. Their sensitivity to environmental change makes population trends widely used indicators of ecosystem health [26, 44, 50], placing reliable long-term monitoring at the center of biodiversity conservation [7].

These traits make acoustic monitoring indispensable for surveying bat populations [43]. Modern passive acoustic monitoring systems use inexpensive autonomous detectors to collect ultrasonic recordings continuously across unprecedented spatial and temporal scales [42]. Converting these recordings into ecological information, however, remains challenging [45, 46, 42]. Whereas automated identification of birds and many other taxa often uses the communication signals, bat classification relies primarily on echolocation calls, whose structure varies not only with species identity but also by flight behaviour, habitat, and foraging context [11]. Consequently, echolocation calls overlap among species and vary within species [16], making reliable identification difficult even for experienced bat workers [46].

Deep learning has substantially improved automated acoustic species recognition [52, 37]. However, much of this progress has relied on relatively small, curated datasets suited to controlled comparisons of architectures and training strategies [4, 18, 1]. This focus has accelerated methodological development but often deprioritized taxonomic breadth, limiting the applicability of classifiers for continental-scale monitoring. For bats, broader coverage requires more than acoustic modeling: experts routinely combine acoustic evidence with contextual ecological knowledge [43] during manual identification. Improving automated bat classification need not rely solely on increasingly sophisticated neural architectures; incorporating ecological information at inference offers a complementary path toward greater practical utility.

To demonstrate this principle, we present a framework for automated classification of European bat vocalisations with species- and genus-level classification heads. At inference, each genus prediction is combined with regional species distributions; when only one species of that genus occurs at the recording location, the prediction is resolved unambiguously to that species (Figure 1). This enables identification of geographically restricted taxa, including species absent from training, thereby extending the framework’s operational taxonomy. To our knowledge, this represents a novel reframing of geographic information in bioacoustic classification: it extends the classifier’s effective taxonomy rather than constraining predictions within a fixed one. More generally, this resolve-when-unambiguous strategy is not tied to acoustic data or geography: any hierarchical classifier could uniquely determine a finer class by combining a reliable coarse prediction with external domain knowledge, such as seasonal occurrence or biotic associations. Our bat framework therefore serves as a proof of principle for this broader inference strategy.

Illustration of inference-stage geographic resolution using echolocation call sequences and approximate Southern European distribution ranges

Figure 1. Illustration of inference-stage geographic resolution using the example of Plecotus kolombatovici. (a, c) Echolocation call sequences of P. austriacus and P. kolombatovici, respectively. (b) Approximate Southern European distribution ranges of P. austriacus (orange) and P. kolombatovici (blue), with overlap shown by orange–blue hatching (ranges based on Dietz et al. [13]). Owing to insufficient acoustic recordings, P. kolombatovici is not included in the learned taxonomy. The species head may therefore assign such a recording to an acoustically similar congener class, illustrated here by P. austriacus. For a recording from Cyprus, a genus-level Plecotus prediction above the confidence threshold activates the geographic lookup (Table 1), which resolves the recording to P. kolombatovici because it is the sole regional representative of the genus.

The main contributions of this work are threefold:

  1. We introduce a joint species- and genus-level classifier that provides explicit genus-level predictions for downstream ecological inference.

  2. We present the broadest operational bat vocalisation classifier for European bats to date, extending coverage from 35 trained species to 41 of the 48 native European bat species1.

  3. We demonstrate a resolve-when-unambiguous strategy in which geographic priors transform genus predictions into species-level identifications for taxa absent from training, providing a proof of principle for extending hierarchical classifiers through external domain knowledge.

Table S1: The 55 bat taxa listed under the EUROBATS Agreement on the Conservation of Populations of European Bats [56], presented using current nomenclature, and their representation in ChirosetEurope (‘CSE‘). ‘No.‘ indicates the number of EUROBATS Range States (n=60n = 60) in which a species is listed as present. Species marked with * have no documented occurrence within the territory covered by the latest EU Action Plan for Bats (2018–2024) [3]. Data current as of July 2026.

Scientific nameCSENo.Distribution within EUROBATS area
Barbastella barbastellus✓45Widespread across Europe and adjacent western Asia
Barbastella caspica *4Caucasus and Caspian region
Cnephaeus anatolicus *5Eastern Mediterranean
Cnephaeus isabellinus6Iberia and North Africa
Cnephaeus nilssonii✓30Northern, central and eastern Europe
Cnephaeus ognevi6Caucasus and Middle East
Cnephaeus serotinus✓53Widespread across Europe, North Africa and western Asia
Hypsugo savii✓40Southern and central Europe, North Africa and western Asia
Miniopterus pallidus *5Caucasus, Turkey and western Asia
Miniopterus schreibersii✓28Southern Europe, the Balkans and the Mediterranean region
Myotis alcathoe✓31Central, southern and eastern Europe
Myotis bechsteinii✓31Western, central and southeastern Europe
Myotis blythii/oxygnathus✓40Southern and central Europe, North Africa and western Asia
Myotis brandtii✓37Northern, central and eastern Europe
Myotis capaccinii✓24Mediterranean region and Middle East
Myotis crypticus8Southwestern Europe
Myotis dasycneme✓24Northern, central and eastern Europe
Myotis daubentonii✓42Widespread across Europe and adjacent western Asia
Myotis davidii✓11Eastern Europe, the Caucasus and western Asia
Myotis emarginatus✓45Southern and central Europe, North Africa and western Asia
Myotis escalerai3France and the Iberian Peninsula
Myotis hoveli4Eastern Mediterranean
Myotis myotis✓35Western, central and southeastern Europe
Myotis mystacinus✓43Widespread across Europe and adjacent western Asia
Myotis nattereri✓33Western, central and northern Europe
Myotis punicus7Western Mediterranean and North Africa
Myotis schaubi *2Armenia and Iran
Myotis tschuliensis *5Caucasus and Turkey
Nyctalus azoreum1Azores (Portugal)
Nyctalus lasiopterus✓27Southern, central and eastern Europe
Nyctalus leisleri✓47Widespread across Europe and adjacent regions
Scientific nameCSE No.Distribution within EUROBATS area
Nyctalus noctula✓ 48Widespread across Europe and western Asia
Otonycteris hemprichii∗11North Africa and Middle East
Pipistrellus hanaki/creticus✓ 1Greece
Pipistrellus kuhlii✓ 44Southern and central Europe, North Africa and western Asia
Pipistrellus maderensis✓ 2Madeira and the Canary Islands
Pipistrellus nathusii✓ 43Widespread across Europe and western Asia
Pipistrellus pipistrellus✓ 53Widespread across the EUROBATS area
Pipistrellus pygmaeus✓ 44Widespread across Europe and western Asia
Plecotus auritus✓ 47Widespread across Europe and adjacent western Asia
Plecotus austriacus✓ 33Western, central and southern Europe
Plecotus gaisleri6Central Mediterranean and North Africa
Plecotus kolombatovici8Adriatic, Balkan and eastern Mediterranean region
Plecotus macrobullaris22Mountain regions of southern Europe and western Asia
Plecotus sardus1Sardinia (Italy)
Plecotus teneriffae1Canary Islands (Spain)
Rhinolophus blasii✓ 20Balkans, eastern Mediterranean and western Asia
Rhinolophus euryale✓ 32Southern Europe, North Africa and western Asia
Rhinolophus ferrumequinum✓ 45Widespread across southern Europe and western Asia
Rhinolophus hipposideros✓ 50Widespread across Europe, North Africa and western Asia
Rhinolophus mehelyi✓ 27Mediterranean region, the Balkans and western Asia
Rousettus aegyptiacus9Eastern Mediterranean and Middle East
Tadarida teniotis✓ 36Mediterranean region and western Asia
Taphozous nudiventris∗10North Africa and Middle East
Vespertilio murinus✓ 40Central, northern and eastern Europe

Table S1. Table S1 continued from previous page

Automated bioacoustic classification

Automated bioacoustic classification has undergone a rapid transition from pipelines based on handcrafted acoustic features to deep learning approaches, most commonly operating on time–frequency representations such as spectrograms [23, 27, 52, 49]. Convolutional and transformer-based architectures now enable large-scale analysis across diverse taxa and recording conditions [53, 27, 24, 6, 38]. Building on this, transfer learning from large audio corpora has become a key strategy for improving performance in low-resource settings [15, 33, 21], allowing models to generalize across domains and taxa with limited labelled data [40, 22, 30, 25]. Despite these advances, most existing systems remain inherently closed-set, restricting predictions to a predefined taxonomy determined during training [52, 28]. This assumption limits their applicability in real-world monitoring, where novel or previously unseen species may occur, motivating recent interest in open-set and open-world bioacoustic classification frameworks that can better handle taxonomic uncertainty and expansion [9, 39, 32, 29].

Automated bat vocalisation classification

Bat-specific studies have similarly shown that spectrogram-based convolutional networks outperform handcrafted acoustic descriptors [59, 8, 48]. Subsequent work has addressed joint detection and classification [36], transformer architectures [4, 18], multi-label prediction [12], and behavioural annotation [58]. Open-source frameworks including NABat ML [31], BatDetect2 [36], BatSpot [51], and BSG-BATS [37] increasingly emphasise reproducible deployment and community-driven development for passive acoustic monitoring. Despite these advances, bat classifiers remain constrained by scarce and heterogeneous annotations. Public repositories are typically long-tailed, combine different recording equipment and protocols, and provide weak recording-level labels rather than strongly annotated, temporally localised calls [8]. Community initiatives have expanded access to recordings and annotation tools [37], while data augmentation and transfer learning can improve generalisation under limited supervision [18, 22]. Nevertheless, representative training data remain the principal bottleneck to broad taxonomic coverage.

Ecological knowledge in automated classification

Ecological knowledge remains underused in automated bat classification. In bioacoustic and fine-grained visual species classification, geographic priors have mainly constrained predictions to locally plausible species within a fixed taxonomy [27, 35]. We instead use geography to extend the effective taxonomy, resolving geographically restricted species absent from training when genus and location uniquely determine species identity. To our knowledge, this reframing has not previously been explored in either domain.

Methodology

Dataset

We train and evaluate our classifier using ChirosetEurope2, a curated collection of European bat vocalisations (echolocation and social calls). The dataset comprises 11,517 recordings of 35 European bat species from 32 countries in the western Palaearctic (Table S2). As expected for an opportunistically aggregated bioacoustic dataset, recordings follow a long-tailed distribution: widespread taxa are substantially better represented than geographically restricted species (Figure S1). We retain this natural imbalance to reflect operational passive acoustic monitoring conditions. ChirosetEurope includes native full-spectrum ultrasonic and historical time-expansion recordings; metadata record each recording’s time-expansion factor (TEF), enabling conversion to a common representation during preprocessing. It provides recording-level (weak) species labels rather than temporally localized call annotations; accordingly, the proposed model is trained with weak supervision. See Supplementary Material for details.

Learned speciesTrainValTestTotalDurationCountries
Barbastella barbastellus170773728400:46:2914
Cnephaeus nilssonii79544417700:34:4412
Cnephaeus serotinus2531237945501:51:0817
Hypsugo savii39910016065901:00:0814
Miniopterus schreibersii1121093025100:40:566
Myotis alcathoe521800:00:455
Myotis bechsteinii6131000:01:054
Myotis blythii331700:00:283
Myotis brandtii381235300:12:115
Myotis capaccinii15212216600:19:412
Myotis dasycneme118633922000:48:0310
Myotis daubentonii254819042501:49:2215
Myotis davidii14111600:00:301
Myotis emarginatus35123800:08:205
Myotis myotis66527300:23:357
Myotis mystacinus67817600:23:308
Myotis nattereri119463319800:44:4812
Nyctalus lasiopterus7121000:03:205
Nyctalus leisleri235524533201:12:2417
Nyctalus noctula51818616787103:21:4019
Pipistrellus hanaki76355917000:13:341
Pipistrellus kuhlii4627726480302:05:0513
Pipistrellus maderensis225174400:16:272
Pipistrellus nathusii42615421679603:58:1116
Pipistrellus pipistrellus1048344210160204:56:2319
Pipistrellus pygmaeus2971149750801:26:2720
Plecotus auritus158406025800:27:5813
Plecotus austriacus8151400:00:284
Rhinolophus blasii1351244730601:01:272
Rhinolophus euryale210622729900:17:304
Rhinolophus ferrumequinum2061093635100:26:3614
Rhinolophus hipposideros283616841200:26:1711
Rhinolophus mehelyi269369940400:20:021
Tadarida teniotis49719028797401:42:048
Vespertilio murinus136446724700:57:117
Geo-resolved speciesApplicable region(s)
Cnephaeus isabellinusGibraltar
Myotis punicusMalta, Pantelleria
Nyctalus azoreumAzores
Plecotus gaisleriMalta, Pantelleria
Plecotus kolombatoviciCyprus
Plecotus teneriffaeCanary Islands

Table S2. Per-species composition and operational taxonomic coverage of ChiroEcho. For the 35 species represented in ChirosetEurope, the table reports recording counts by split, total audio duration, and the number of recording countries. The lower block lists six additional species supported in ChiroEcho through geographic resolution and the regions in which genus and location uniquely determine species identity. Dataset statistics are not applicable to these species because they are outside the learned taxonomy.

Recording abundance and relative geographic coverage across the 35 species in ChirosetEurope

Figure S1. Recording abundance and relative geographic coverage across the 35 species in ChirosetEurope. Species are ordered by decreasing recording count, represented by bar height. Bar colour indicates the proportion of each species’ reported EUROBATS range states represented by at least one recording, with darker shades indicating greater coverage. Percentages are calculated against the full set of reported range states, including neighbouring countries and range states not represented in ChirosetEurope.

Audio preprocessing

All recordings are converted to a common input representation before feature extraction. Each recording is loaded at its native sampling rate, time-expanded tenfold unless already at TEF =10= 10, and resampled to 32 kHz. Recordings shorter than 20 s are tiled to 20 s; longer recordings are divided into non-overlapping 20 s segments. Trailing remainders of at least 2 s are tiled to 20 s; shorter remainders are discarded. To represent realistic field conditions, we construct a noise class from bat-free recordings in the open-source audio dataset ECOSoundSet [19]. Clips are drawn from two complementary subsets: recordings sampled above 100 kHz (250–384 kHz), which preserve the ultrasonic range relevant to bat monitoring, and recordings sampled at 96 kHz, which provide greater acoustic diversity (birdsong, geophony, anthropophony). The number of anthropogenic sounds is capped to prevent class imbalance. Each 4 s source clip undergoes the same time expansion and resampling pipeline, yielding 40 s of audio, from which the first 20 s are retained without padding. Bat and noise clips therefore have the same input format. The resulting 20 s clips, hereafter termed segments, serve as classifier inputs. During evaluation, segments are processed independently; for recordings containing multiple segments, per-class scores are max-pooled to produce recording-level predictions.

Model architecture

ChiroEcho3 uses a shared EfficientNet-B3 acoustic encoder [54] with species and genus classification heads. We evaluate two initializations: ChiroEcho_IN uses standard ImageNet-1k weights [10] provided by timm, whereas ChiroEcho_Perch uses weights from Perch 2.0 [57], an EfficientNet-B3 bioacoustic model trained on over 1.5 million labelled recordings of birds, amphibians, insects, and mammals.

To enable end-to-end fine-tuning in PyTorch, we map the released JAX/Flax backbone weights layer by layer to the equivalent timm implementation. Transferring only the convolutional backbone, the Perch linear and prototype-learning classification heads, covering 14,795 species, are discarded. We validate the conversion on matched inputs by comparing intermediate activations at the stem, every MBConv block, and the final convolutional head, obtaining near-exact numerical agreement (cosine similarity >0.999> 0.999) throughout.

Input representation. All models use the Perch 2.0 log-mel spectrogram frontend [57]: a 20 ms window with 10 ms hop (640/320 segments at 32 kHz), a 1024-point FFT, HTK-scaled mel filterbanks with 128 bins spanning 60 Hz–16 kHz, window-sum-normalized magnitude spectra, and log compression with a 10−510^{-5} floor. Using this frontend for both variants isolates the effect of initialization from differences in input representation. Because ChiroEchoIN_{\mathrm{IN}}’s pretrained stem expects three-channel input, we modify its first convolution to accept single-channel spectrograms rather than replicate the input across three channels.

Genus-level auxiliary head. To address acoustic ambiguity among closely related species, particularly within genera such as Myotis, we add an auxiliary genus-classification alongside the species head. Each head is a single linear layer applied to the shared pooled encoder embedding. Genus labels require no additional annotation and are derived by mapping the 35 training species to 11 genera. The heads are trained jointly using

L=Lspecies+Lgenus(1)\mathcal{L} = \mathcal{L}_{\mathrm{species}} + \mathcal{L}_{\mathrm{genus}} \tag*{(1)}

Both terms are binary cross-entropy losses on multi-label targets, consistent with our mixup strategy, which produces multi-label rather than soft-interpolated targets. This is a standard multi-task setup rather than a hierarchical or conditional classifier: the heads share the encoder but produce predictions independently, without conditioning or gating. The genus objective provides additional supervision and supports geographic resolution (Section 3.6), where genus predictions combined with location metadata can recover certain species absent from or underrepresented in the training data.

Training protocol

Training strategy. Both ChiroEcho variants are trained end-to-end, with no frozen backbone layers and randomly initialized linear heads after global average pooling. We use AdamW [34] with a learning rate and weight decay of 1×10−41 \times10^{-4}, a batch size of 32, and no gradient accumulation. The backbone and both heads are optimized jointly using the multi-label binary cross-entropy objective in (1). Training uses automatic mixed precision and gradient clipping with a maximum norm of 1.0. A reduce-on-plateau scheduler (mode max) monitors validation mean average precision (mAP), reducing the learning rate by a factor of 0.1 after four epochs without improvement, to a minimum of 1×10−61 \times10^{-6}. Because species classification is the primary objective, checkpoint selection and early stopping are based on validation species mAP. Training stops after five consecutive epochs without improvement, yielding final checkpoints at epochs 14 for ChiroEchoIN_{\mathrm{IN}} and 22 for ChiroEchoPerch_{\mathrm{Perch}}.

To mitigate class imbalance, examples are sampled with replacement using capped inverse-frequency weights. For class cc, with ncn_c training examples, the class weight is

wc=min⁡{nc,1000}−0.5(2)w_c = \min\{n_c, 1000\}^{-0.5} \tag*{(2)}

The cap assigns equal weights to classes with at least 1,000 examples, focusing oversampling on underrepresented classes. The resulting per-example weights define the weighted random sampler.

Data augmentation. We apply waveform mixup and SpecAugment during training. Mixup is applied with probability p=0.1p = 0.1 per batch, using a coefficient λ∼Beta⁡(10,10)\lambda\sim\operatorname{Beta}(10, 10) to linearly combine pairs of waveforms. Rather than interpolate their labels using λ\lambda, we retain both source labels as a hard multi-label target, consistent with binary cross-entropy loss. SpecAugment [41] is applied to every log-mel spectrogram using frequency and time masks with maximum widths of 48 bins and 192 frames, respectively; no frequency or time warping is used.

Closed-set species and genus classification

We first evaluate the model under the standard closed-set assumption: every test recording belongs to one of the trained species classes. This provides a baseline measure of species and genus discrimination ability before considering the open-set extension enabled by geographic resolution (Section 3.6), where recordings might belong to species entirely absent from training. Following the max-pooling procedure described above (Section 3.2), We report recording-level F1-score, mean average precision (mAP), and area under the receiver operating characteristic curve (AUROC). mAP is the unweighted mean of per-class AP, whereas F1-score and AUROC are averaged using per-class weighting; F1-score is computed by applying a fixed threshold of 0.5 to every class, without threshold tuning, ensuring comparability across settings. Species and genus predictions are evaluated independently using their respective label spaces of 35 species and 11 genera.

Geographic resolution

Prior work has primarily used geographic information to constrain predictions within a fixed taxonomy by favouring locally plausible species. We instead use geography to extend the classifier’s operational taxonomy beyond the species represented during training. This is possible in regions where only one species of a given genus occurs: once a recording is assigned to that genus, its location uniquely determines species identity (Table 1). For example, a Myotis prediction in Malta resolves to M. punicus (Figure 2), while a Plecotus prediction in Malta resolves to Pl. gaisleri. We implement this inference as a lightweight, rule-based post-processing step combining the genus head (Section 3.3) with recording location metadata. Algorithm 1 is applied independently to each predicted genus, allowing predictions for multiple genera within a recording. For a genus prediction g^\hat{g} with confidence c≥τc \ge\tau, the rule queries GeoResolveSpecies(g^,ℓ\hat{g}, \ell), where ℓ\ell is the recording location. The lookup returns a species ss only when that species is the sole representative of genus g^\hat{g} in the region; otherwise, it returns ∅\varnothing. When ss is returned, it inherits the genus confidence cc and replaces the species-head predictions s^g\hat{s}_g for that genus. Predictions for other genera remain unchanged. If c<τc < \tau or the lookup returns ∅\varnothing, the corresponding species-head predictions are retained (Figure 2).

Figure 2. Schematic overview of the ChiroEcho architecture and geographic-resolution step, illustrated for Myotis punicus in Malta. The depicted species-head output, M. blythii, is illustrative rather than observed and represents a plausible confusion with a closely related species whose echolocation characteristics overlap.

GenusRegionResolved speciesIn-region dataTaxonomy extension
CnephaeusGibraltarC. isabellinus✓
MyotisMaltaM. punicus✓
MyotisPantelleriaM. punicus✓
NyctalusAzoresN. azoreum✓
PlecotusMaltaPl. gaisleri✓
PlecotusPantelleriaPl. gaisleri✓
PlecotusCyprusPl. kolombatovici✓
PlecotusCanary Is.Pl. teneriffae✓
PipistrellusPantelleriaP. kuhlii
PipistrellusAzoresP. maderensis
PipistrellusMadeiraP. maderensis✓

Table 1. Region–genus lookup table used to resolve species ss from predicted genus g^\hat{g} and location ℓ\ell (Algorithm 1). A genus prediction is resolved only when a single species of that genus occurs in the listed region. “In-region data” indicates that labeled recordings of the resolved species were available from that region. “Taxonomy extension” indicates that the resolved species was absent from the training taxonomy.

Algorithm 1 Geographic resolution rule

SmodelS_{\mathrm{model}}: species represented during training; SgeoS_{\mathrm{geo}}: species that may be returned by the geographic lookup

Require: genus prediction g^\hat{g} with confidence cc, user-defined threshold τ\tau, recording location ℓ\ell, scored species-head predictions s^g\hat{s}_g for genus g^\hat{g}, lookup function GeoResolveSpecies(g^,ℓ\hat{g}, \ell) (one-to-one mapping)

  1. s←∅s \leftarrow\varnothing

  2. if c≥τc \ge\tau then

  3. s←GeoResolveSpecies⁡(g^,ℓ)s \leftarrow\operatorname{GeoResolveSpecies}(\hat{g}, \ell) ▹s∈Sgeo∪{∅}\triangleright s \in S_{\mathrm{geo}} \cup\{\varnothing\}

  4. end if

  5. if s≠∅s \ne\varnothing then

  6. return {(s,c)}\{(s,c)\} ▹\triangleright geographically resolved single species inherits genus confidence and overrides s^g\hat{s}_g

  7. end if

  8. if s^g≠∅\hat{s}_g \ne\varnothing then

  9. return s^g\hat{s}_g ▹\triangleright fallback to species-head predictions s^g⊆Smodel\hat{s}_g \subseteq S_{\mathrm{model}}

  10. end if

  11. return ∅\varnothing ▹\triangleright no valid genus or species prediction for genus g^\hat{g}

This mechanism can identify species absent from training, provided that their genus is resolvable and they are its sole regional representative. In the current implementation, the lookup contains 11 region–genus mappings across seven regions and resolves eight species, six of which are absent from training (Table 1). It therefore extends the classifier’s operational taxonomy from 35 to 41 European bat species. The rule requires no additional training and remains transparent and readily extensible as species-distribution knowledge evolves. Its applicability is limited to region–genus pairs that uniquely determine species identity and depends on accurate location and distribution data.

To evaluate whether the rule can recover species outside the learned taxonomy, we trained a separate model with Pipistrellus maderensis and P. kuhlii excluded from both the training data and the species-head output classes. The held-out recordings from these two species were retained solely for evaluation, making direct prediction by the species head impossible. For P. maderensis, we applied the rule using observed recording locations in Madeira. We selected P. kuhlii because, aside from P. maderensis, it was the only species in the geographic lookup table with sufficient recordings for a quantitative held-out evaluation (n=803n=803). Because none of these recordings originated from Pantelleria, where P. kuhlii is the sole representative of Pipistrellus, we assigned Pantelleria as their location metadata. This provides a controlled test of the inference mechanism. Together, the two conditions test whether genus predictions and geographic information can extend operational species coverage beyond the learned taxonomy. Results are reported in Section 4.2.

Results

We first evaluate classification described in Section 3.5, where all test recordings belong to classes represented during training, and examine performance on sparsely represented species. We then evaluate the geographic resolution strategy using the held-out proof-of-principle experiment described in Section 3.6.

Closed-set species and genus classification

Table 2 reports recording-level species and genus performance for the two ChiroEcho variants. ChiroEchoPerch_{\mathrm{Perch}} achieved slightly higher F1-scores at both taxonomic levels, whereas ChiroEchoIN_{\mathrm{IN}} achieved higher mAP and AUROC, obtaining the best result for four of the six metrics. Genus-level F1-score and mAP exceeded species-level performance for both variants; genus-level AUROC was higher for ChiroEchoIN_{\mathrm{IN}} and equal to species-level AUROC for ChiroEchoPerch_{\mathrm{Perch}}. Overall, the two initialization strategies performed comparably.

ModelWeights initializationSpeciesGenus
F1-scoremAPAUROCF1-scoremAPAUROC
ChiroEchoIN_{\mathrm{IN}}ImageNet-1k0.8360.6940.9900.8800.9230.992
ChiroEchoPerch_{\mathrm{Perch}}Perch 2.00.8450.6720.9880.8860.9210.988

Table 2. Recording-level closed-set species and genus classification performance of ChiroEcho, trained on 35 European bat species and one noise class with ImageNet-1k or Perch 2.0 initialization. Segment-level predictions are aggregated to recording-level predictions by max pooling before evaluation. Best performance per metric is shown in bold.

Aggregate metrics can obscure uncertainty for sparsely represented species. Table 3 therefore reports AP for the nine species represented by fewer than 10 test-split segments using ChiroEchoIN_{\mathrm{IN}}. Test-split average precision (AP) ranged from 0.001 for M. blythii to 1.0000 for M. davidii, although the latter was based on a single test segment. To assess sensitivity to the small sample sizes, we also calculated AP on the combined validation and test splits, increasing the number of segments per species from 1–9 to 2–50. AP changed substantially for several species, including M. emarginatus (0.004 to 0.335), M. capaccinii (0.501 to 0.817), and Plecotus austriacus (0.825 to 0.374), underscoring the uncertainty associated with these sparse evaluations.

Species# SegmentsAverage precision
TestVal+TestTestVal+Test
Myotis davidii121.0001.000
Myotis alcathoe1110.0190.013
Myotis blythii2150.0010.001
Myotis mystacinus2420.0190.074
Myotis emarginatus470.0040.335
Myotis bechsteinii590.1380.052
Myotis capaccinii6260.5010.817
Myotis brandtii9500.3030.493
Plecotus austriacus9110.8250.374

Table 3. Table 3: Classification results for species represented by fewer than 10 test-split segments using ChiroEchoIN_{\mathrm{IN}}. Average precision (AP) is reported for both the test split and the combined validation+test split.

Because genus labels can also be inferred from species predictions, we compared the auxiliary genus head with the genus implied by the top-1 species prediction. The two agreed for 97.74% of test recordings and disagreed for 2.26%. Top-1 species accuracy was 91.53% when the predictions agreed but only 32.81% when they disagreed (Fisher’s exact test, odds ratio = 22.1, p=1.2×10−29p = 1.2 \times10^{-29}). Disagreement between the heads therefore provides a strong indicator of species-level prediction error.

Geographic resolution

We next evaluate the central contribution of this work: extending species recognition beyond the training taxonomy through inference-stage geographic resolution. Table 4 summarizes the held-out endemic species experiment described in Section 3.6: all Pipistrellus kuhlii and P. maderensis recordings and their species-head output classes were removed from the training taxonomy, leaving 33 trained species plus the noise class. All available segments from the held-out species were used for evaluation.

Species# SegmentsSpecies head (baseline)Geo-resolved species accuracy
Top confusion%
P. kuhlii803P. nathusii37.4%0.695
P. maderensis44P. pipistrellus93.2%0.954

Table 4. Table 4: Geographic resolution of two Pipistrellus species held out entirely from the training taxonomy and species-head label space. All available segments were used for evaluation. Because neither species has an output class, direct identification by the species head is impossible; the “Top confusion” columns instead report the most frequent segment-level top-1 prediction among the trained classes and its share of segments. Geo-resolved species accuracy is reported at the recording level after max pooling and application of the geographic resolution rule (Algorithm 1).

Because neither species had an output class, the species head necessarily predicted incorrect alternatives from the training taxonomy, most often acoustically similar Pipistrellus species. Its most frequent segment-level predictions were P. nathusii for P. kuhlii (300/803 segments; 37.4%) and P. pipistrellus for P. maderensis (41/44; 93.2%). In contrast, combining the genus head with geographic resolution recovered the correct held-out species with recording-level accuracies of 69.5% and 95.4%, respectively. Because the lookup is deterministic, successful resolution depends on the genus head identifying Pipistrellus above the confidence threshold. These results provide proof of principle that genus-level predictions combined with geographic information can recover species absent from the learned taxonomy.

Discussion

Overall, our results show that explicit geographic information can complement acoustic classifiers to extend the species coverage beyond the training taxonomy while maintaining strong closed-set performance. In doing so, the framework also broadens the operational taxonomic scope of automated European bat classification. We discuss the implications of these findings below.

We first consider the influence of initialization on the underlying acoustic classifier. ChiroEchoIN, initialized with ImageNet-1k weights, achieved the highest mAP and AUROC at both taxonomic levels, whereas ChiroEchoPerch achieved slightly higher F1-scores. F1-score depends on a single operating threshold and is sensitive to calibration and class distribution, whereas mAP and AUROC assess discrimination across thresholds. We therefore emphasize mAP and AUROC when comparing the initialization strategies to provide a more robust assessment of the model’s discriminative ability. Based on these metrics, ChiroEchoIN provides the strongest overall acoustic classifier and serves as the default ChiroEcho model in the remaining experiments.

The principal contribution of this work lies in the inference-stage geographic resolution mechanism. The held-out proof-of-principle experiment (Table 4) evaluated two species absent from a separately trained 33-species taxonomy, making direct species-head prediction impossible. Rather than failing arbitrarily, the species head consistently confused held-out recordings with acoustically similar congeners, suggesting that the learned embedding captures meaningful taxonomic structure even beyond the trained output classes. Geographic resolution recovered the correct species with recording-level accuracies of 69.5% and 95.4%, respectively, without requiring any species-specific training data. This mechanism is inherently limited to taxa for which geographic occurrence uniquely determines species identity (Table 1) and therefore cannot be applied to many species with limited acoustic training data. For applicable species, however, it demonstrates how ecological knowledge can complement learned acoustic representations. Because the rule operates only at inference, updated species distributions, taxonomic revisions, or newly recognized endemic populations can be incorporated without retraining the acoustic model.

Outside geographically resolvable cases, performance remains constrained by the availability of acoustic training and evaluation data. Table 3 illustrates the instability of per-species evaluation under severe class imbalance. With as few as one or two test segments, average precision is dominated by individual predictions rather than reflecting general model behavior: a single correct or incorrect classification can shift the reported score from 0 to 1. The perfect AP obtained for M. davidii (1 test segment) should therefore not be interpreted as evidence of reliable performance on this species, but rather as an artifact of the evaluation size. Combining the validation and test splits provides a broader estimate, but several species remain represented by only 2–11 segments, and their AP values vary substantially between splits. We view this as an inherent constraint of working with long-tailed, data-scarce taxa in bioacoustic monitoring, rather than a shortcoming specific to our model, and note that the low training segment counts for these species (10–40 segments) likely contribute directly to their inconsistent performance. This presents a practical challenge for conservation, as many species for which reliable monitoring is most important are also among those with the scarcest acoustic training data, creating a self-reinforcing cycle in which limited data availability constrains classifier performance precisely where automated monitoring could provide the greatest benefit. Targeted collection of recordings from underrepresented species should therefore remain a priority.

A second limitation is domain shift. Both training and evaluation use curated repository recordings, whereas operational passive acoustic monitoring encounters more complex soundscapes with overlapping vocalisations, environmental noise, and recording artefacts. Evaluating the proposed framework on long-duration, uncurated recordings is therefore an important next step toward deployment.

Beyond its role in geographic resolution, the auxiliary genus head provides informative genus-level predictions when species-level discrimination is uncertain. Genus-level F1-score and mAP exceeded species-level values for both initialization strategies; genus-level AUROC was higher for ChiroEchoIN_{\mathrm{IN}} and equal for ChiroEchoPerch_{\mathrm{Perch}}. This pattern is expected as distinguishing among genera is generally easier than distinguishing among closely related species with overlapping vocal characteristics. The genus head therefore supports graceful degradation: when a species prediction is unreliable, the system can report a genus and confidence score, narrowing the set of plausible species for expert review. This is particularly relevant for rare species such as those in Table 3, whose species-level training data may remain scarce.

The genus implied by the top-1 species prediction agreed with the dedicated genus head for 97.74% of test recordings, indicating that an explicit head is not strictly necessary to obtain genus-level predictions. We nevertheless retain it as an architectural choice, for two reasons. First, disagreement between the heads provides a useful uncertainty signal unavailable to a purely species-derived genus score: species accuracy fell from 91.53% on agreeing recordings to 32.81% on disagreeing recordings. Second, the dedicated head provides a directly thresholdable genus output without requiring an additional aggregation rule over multi-label species outputs, at negligible additional computational cost.

Conclusion

We presented a deep learning framework for automated classification of European bat vocalisations that combines species- and genus-level prediction with ecological domain knowledge. Trained on 35 species, the proposed framework extends operational coverage to 41 of the 48 native European bat species by combining learned genus predictions with an inference-stage geographic rule. This increases operational taxonomic coverage from 73% to 85% without additional species-specific training examples and, to our knowledge, represents the broadest operational coverage reported for automated European bat classification.

More broadly, this study illustrates how transparent domain knowledge can extend a learned classifier without increasing architectural complexity. Although we demonstrate the resolve-when-unambiguous principle for genus-to-species acoustic classification in European bats, it could extend the effective taxonomy of other hierarchical classifiers whenever a reliable coarse prediction and an external ecological constraint uniquely determine a fine-grained class. Such constraints might include regional checklists, host associations, or seasonal occurrence, with potential applications to other acoustically monitored taxa, camera-trap imagery, and eDNA surveys. We therefore view our bat pipeline as one instance of a broader strategy for taxonomic resolution through auxiliary ecological information.

Future work will focus on two complementary directions. First, the proposed framework should be evaluated on uncurated passive acoustic monitoring soundscapes to assess its robustness to the domain shift between curated reference recordings and real-world monitoring data. Strongly annotated soundscape datasets will be essential both for this evaluation and for further improving the underlying acoustic classifier as they become available. Second, we will extend the ecological inference framework by incorporating additional forms of ecological domain knowledge. An important next step will be to provide geographically plausible candidate species that are absent from the learned taxonomy, allowing users to distinguish between species that can be identified directly by the classifier and those that remain ecologically plausible but are not yet represented in the training data. This would improve the transparency and practical utility of automated biodiversity monitoring systems while accommodating future taxonomic revisions, refined species distributions, and newly available training data. Rather than treating ecological knowledge as external to machine learning, we envision biodiversity monitoring systems in which learned acoustic representations and explicit ecological reasoning operate as complementary components of a unified inference pipeline.

CRediT authorship contribution statement

BG: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – original draft, Writing – review & editing. WE: Software. MH: Data curation. JSC: Conceptualization, Writing – review & editing, Visualization. VJK: Methodology, Supervision, Project administration, Funding acquisition. DS: Methodology, Supervision, Project administration, Funding acquisition. ALB: Conceptualization, Methodology, Validation, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization, Supervision.

Acknowledgements

This work was supported by funding received via the Horizon Europe projects TETTRIs (grant no. 101081903, supporting work by BG and VJK) and MAMBO (grant no. 101060639, supporting work by BG, MH, VJK, DS, and ALB) as well as the UK Research and Innovation (UKRI) Horizon Europe Guarantee (Grant EP/Y033299/1), as part of the Marie Skłodowska-Curie Actions Doctoral Networks Programme, BioacAI (grant no. 101116715, supporting work by JSC). The views and opinions expressed are, however, those of the authors only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Supplementary Material

ChirosetEurope dataset

Taxonomic coverage. To contextualize the taxonomic coverage of ChirosetEurope, we compared its contents with the 55 bat species currently listed under the EUROBATS Agreement [56]. The Agreement, however, has a broader geographic scope than the latest EU Action Plan for Bats (2018–2024)[3], encompassing non-EU European states as well as parts of North Africa and the Middle East. Seven of the 55 species have no documented occurrence within the area covered by the Action Plan, resulting in a reference set of 48 species. This updated total exceeds the 45 species originally recognized by the Action Plan because it incorporates subsequent taxonomic revisions adopted by EUROBATS [3, 56]. ChirosetEurope currently represents 35 of these 48 species (Table S1).

Composition. Building on the reference framework described above, Table S2 reports the per-species composition of ChirosetEurope and its operational taxonomic coverage. Figure S1 shows recording abundance and relative geographic coverage, highlighting the long-tailed distribution arising from differences in species availability and recording effort. These summaries describe bat recordings only and exclude the separately constructed noise class.

To minimise data leakage, recordings were partitioned using session-level metadata grouping keys. Sessions comprised recordings from the same biological night and approximate location, defined by coordinates rounded to three decimal places (approximately 100 m); recordings made before 12:00 were assigned to the preceding night. Where coordinates were unavailable, recordings were grouped by biological night, recordist, and country. All recordings within a group were assigned to the same split.

ChirosetEurope contains native full-spectrum and historical time-expanded recordings. The time-expansion factor retained in the metadata allows both formats to be transformed to a common input representation. Heterodyne and frequency-division recordings were excluded because these methods irreversibly alter the original ultrasonic signal.

References

  1. [1]Alipek, S., Maelzer, M., Paumen, Y., Schauer-Weisshahn, H., Moll, J.: An Efficient Neural Network Design Incorporating Autoencoders for the Classification of Bat Echolocation Sounds. Animals 13(16), 2560 (Jan 2023). https://doi.org/10.3390/ani13162560
  2. [3]Barova, S., Streit, A.: Action plan for the conservation of all bat species in the european union 2018–2024. Brussels: European Commission, UNEP and Eurobats (2018), https://environment.ec.europa.eu/topics/nature-and-biodiversity/habitats-directive_en
  3. [4]Bellakfir, H., Vogelbacher, M., Gottwald, J., Mühling, M., Korfhage, N., Lampe, P., Frieß, N., Nauss, T., Freisleben, B.: Bat Echolocation Call Detection and Species Recognition by Transformers with Self-attention. In: Bennour, A., Ensari, T., Kessentini, Y., Eom, S. (eds.) Intelligent Systems and Pattern Recognition. pp. 189–203. Springer International Publishing, Cham (2022)DOI
  4. [5]Browning, E., Barlow, K.E., Burns, F., Hawkins, C., Boughey, K.: Drivers of European bat population change: A review reveals evidence gaps. Mammal Review 51(3), 353–368 (2021). https://doi.org/10.1111/mam.12239
  5. [6]Cañas, J.S., Kahl, S., Denton, T., Toro-Gómez, M.P., Rodriguez-Buritica, S., Benavides-Lopez, J.L., Ulloa, J.S., Caycedo-Rosales, P., Klinck, H., Goëau, H., et al.: Overview of birdclef+ 2025: Multi-taxonomic sound identification in the middle magdalena, colombia. In: Conference and Labs of the Evaluation Forum (CLEF 2025). pp. 2909–2919. CEUR-WS (2025)
  6. [7]Cardinale, B.J., Duffy, J.E., Gonzalez, A., Hooper, D.U., Perrings, C., Venail, P., Narwani, A., Mace, G.M., Tilman, D., Wardle, D.A., Kinzig, A.P., Daily, G.C., Loreau, M., Grace, J.B., Larigauderie, A., Srivastava, D.S., Naeem, S.: Biodiversity loss and its impact on humanity. Nature 486(7401), 59–67 (Jun 2012). https://doi.org/10.1038/nature11148
  7. [8]Chen, X., Zhao, J., Chen, Y.h., Zhou, W., Hughes, A.C.: Automatic standardized processing and identification of tropical bat calls using deep learning approaches. Biological Conservation 241, 108269 (Jan 2020). https://doi.org/10.1016/j.biocon.2019.108269
  8. [9]Cramer, A.L., Lostaunen, V., Farnsworth, A., Salamon, J., Bello, J.P.: Chirping up the right tree: Incorporating biological taxonomies into deep bioacoustic classifiers. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 901–905. IEEE (2020)
  9. [10]Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Li, Fei-Fei: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248–255. IEEE, Miami, FL (Jun 2009). https://doi.org/10.1109/CVPR.2009.5206848
  10. [11]Denzinger, A., Schnitzler, H.U.: Bat guilds, a concept to classify the highly diverse foraging and echolocation behaviors of microchiropteran bats. Frontiers in Physiology 4, 15–15 (2013). https://doi.org/10.3389/fphys.2013.00164
  11. [12]Dierckx, L., Beauvois, M., Nijssen, S.: Detection and multi-label classification of bats. In: International Symposium on Intelligent Data Analysis. pp. 53–65. Springer (2022)DOI
  12. [13]Dietz, C., Kiefer, A.: Naturführer Fledermäuse Europas: Alle Arten erkennen und sicher bestimmen. Franckh-Kosmos Verlags-GmbH & Co. KG, Stuttgart (2020)
  13. [14]Directive, H., et al.: Council directive 92/43/eec of 21 may 1992 on the conservation of natural habitats and of wild fauna and flora. Official Journal of the European Union 206, 50 (1992)
  14. [15]Dufourcq, E., Batist, C., Foquet, R., Durbach, I.: Passive acoustic monitoring of animal populations with transfer learning. Ecological Informatics 70, 101688 (2022)DOI
  15. [16]Fenton, M.B., Grinnell, A.D., Popper, A.N., Fay, R.R.: Bat Bioacoustics. Springer (2016)
  16. [17]Fenton, MB., Simmons, NB.: Bats: A World of Science and Mystery. University of Chicago Press (2020)
  17. [18]Fundel, F., Braun, D.A., Gottwald, S.: Automatic bat call classification using transformer networks. Ecological Informatics 78, 102288 (Dec 2023). https://doi.org/10.1016/j.ecoinf.2023.102288
  18. [19]Funosas, D., Massol, E., Bas, Y., Schmidt, S., Bennett, D., Arend, D., Gebhard, A., Barbaro, L., König, S., Font, R.C.: A finely annotated dataset for the automated acoustic identification of European Orthoptera and Cicadidae. Scientific Data (2026)DOI
  19. [20]Ghanem, S.J., Voigt, C.C.: Increasing Awareness of Ecosystem Services Provided by Bats. In: Advances in the Study of Behavior, vol. 44, pp. 279–302. Elsevier (2012). https://doi.org/10.1016/B978-0-12-394288-3.00007-1
  20. [21]Ghani, B., Baier, A.L., Kalkman, V.J., Stowell, D.: Twelve quick tips for applying deep learning to animal sounds. PLOS Computational Biology 22(8), e1014604 (2026). https://doi.org/10.1371/journal.pcbi.1014604
  21. [22]Ghani, B., Kalkman, V.J., Planqué, B., Vellinga, W.P., Gill, L., Stowell, D.: Impact of transfer learning methods and dataset characteristics on generalization in birdsong classification. Scientific Reports 15(1), 16273 (May 2025). https://doi.org/10.1038/s41598-025-00996-2
  22. [23]Goëau, H., Glotin, H., Vellinga, W.P., Planqué, R., Joly, A.: Lifelefler bird identification task 2016: The arrival of deep learning. In: CLEF: conference and labs of the evaluation forum. pp. 440–449 (2016)
  23. [24]Hagiwara, M.: Aves: Animal vocalization encoder based on self-supervision. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2023)arxiv.org/abs/2210.14493
  24. [25]Hummel, H.I., Bhulai, S., van der Mei, R., Ghani, B.: Linear probing enables ship-radiated noise recognition with pretrained audio embeddings. Ecological Informatics p. 103709 (2026)DOI
  25. [26]Jones, G., Jacobs, D., Kunz, T., Willig, M., Racey, P.: Carpe noctem: The importance of bats as bioindicators. Endangered Species Research 8, 93–115 (Jul 2009). https://doi.org/10.3354/esr00182
  26. [27]Kahl, S., Wood, C.M., Eibl, M., Klinck, H.: BirdNET: A deep learning solution for avian diversity monitoring. Ecological Informatics 61, 101236 (Mar 2021). https://doi.org/10.1016/j.ecoinf.2021.101236
  27. [28]Kath, H., Gouvéa, T.S., Sonntag, D.: Novel class identification: A human-in-the-loop approach for efficient species recognition in bioacoustic data. Methods in Ecology and Evolution (2026)
  28. [29]Kather, V.S., Ghani, B., Stowell, D.: Clustering and novel class recognition: evaluating bioacoustic deep learning feature extractors. arXiv preprint arXiv:2504.06710 (2025)arxiv.org/abs/2504.06710
  29. [30]Kather, V.S., Haupert, S., Ghani, B., Stowell, D.: bacpipe: a python package to make bioacoustic deep learning models accessible. arXiv preprint arXiv:2604.11560 (2026)arxiv.org/abs/2604.11560
  30. [31]Khalighifar, A., Gotthold, B.S., Adams, E., Barnett, J., Beard, L.O., Britzke, E.R., Burger, P.A., Chase, K., Cordes, Z., Cryan, P.M., Ferrall, E., Fill, C.T., Gibson, S.E., Haulton, G.S., Irvine, K.M., Katz, L.S., Kendall, W.L., Long, C.A., Mac Aodha, O., McBurney, T., McCarthy, S., McKown, M.W., O’Keefe, J., Patterson, L.D., Pitcher, K.A., Rustand, M., Segers, J.L., Seppanen, K., Siemers, J.L., Stratton, C., Straw, B.R., Weller, T.J., Reichert, B.E.: NABat ML: Utilizing deep learning to enable crowdsourced development of automated, scalable solutions for documenting North American bat populations. Journal of Applied Ecology 59(11), 2849–2862 (Nov 2022). https://doi.org/10.1111/1365-2664.14280
  31. [32]Ko, K., Lee, B., Kim, D., Hong, J., Ko, H.: Open set bioacoustic signal classification based on class anchor clustering with closed set unknown bioacoustic signals. In: 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). pp. 1–4. IEEE (2023)DOI
  32. [33]Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28, 2880–2894 (2020)arxiv.org/abs/1912.10211
  33. [34]Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization (2017). https://doi.org/10.48550/ARXIV.1711.05101
  34. [35]Mac Aodha, O., Cole, E., Perona, P.: Presence-only geographical priors for fine-grained image classification. In: Proceedings of the IEEE/Cvf International Conference on Computer Vision. pp. 9596–9606 (2019)arxiv.org/abs/1906.05272
  35. [36]Mac Aodha, O., Martínez Balvanera, S., Damstra, E., Cooke, M., Eichinski, P., Browning, E., Barataud, M., Boughey, K., Coles, R., Giacomini, G., Swiney G., M.C.M., Obrist, M.K., Parsons, S., Sattler, T., Jones, K.E.: Towards a General Approach for Bat Echolocation Detection and Classification (Dec 2022). https: //doi.org/10.1101/2022.12.14.520490DOI
  36. [37]Meramo, K., Somervuo, P., Rannisto, M., Lauha, P., López-Baucells, A., Gili, F., Roslin, T., Lilley, T.M., Ovaskainen, O.: BSG-BATS: An open-source data annotation portal and classifier for European bat vocalizations. Methods in Ecology and Evolution 17(3), 718–726 (Mar 2026). https://doi.org/10.1111/2041-210X. 70220
  37. [38]Miron, M., Robinson, D., Alizadeh, M., Gilsenan-McMahon, E., Narula, G., Chemla, E., Cusimano, M., Effenberger, F., Hagiwara, M., Hoffman, B., et al.: Avex: What matters for animal vocalization encoding. In: International Conference on Learning Representations. vol. 2026, pp. 113288–113313 (2026)arxiv.org/abs/2508.11845
  38. [39]Morgan, M.M., Braasch, J.: Open set classification strategies for long-term environmental field recordings for bird species recognition. The Journal of the Acoustical Society of America 151(6), 4028–4038 (2022)
  39. [40]Nolasco, I., Singh, S., Morfi, V., Lostanlen, V., Strandburg-Peshkin, A., Vidana-Vila, E., Gill, L., Pamula, H., Whitehead, H., Kiskin, I., et al.: Learning to detect an animal sound from five examples. Ecological informatics 77, 102258 (2023)DOI
  40. [41]Park, D.S., Chan, W., Zhang, Y., Chiu, C.C., Zoph, B., Cubuk, E.D., Le, Q.V.: SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition. In: Interspeech 2019. pp. 2613–2617. ISCA (Sep 2019). https: //doi.org/10.21437/Interspeech.2019-2680DOI
  41. [42]Roemer, C., Haquhart, A., López-Baucells, A., Besnard, A.: Current frontiers in the passive acoustic monitoring of bats. Methods in Ecology and Evolution 16(11), 2534–2544 (Nov 2025). https://doi.org/10.1111/2041-210X.70157
  42. [43]Runkel, V., Gerding, G., Markmann, U.: The Handbook of Acoustic Bat Detection. Pelagic Publishing Ltd (Aug 2021)
  43. [44]Russo, D., Salinas-Ramos, V.B., Cistrone, L., Smeraldo, S., Bosso, L., Ancillotto, L.: Do We Need to Use Bats as Bioindicators? Biology 10(8), 693 (Jul 2021). https://doi.org/10.3390/biology10080693
  44. [45]Russo, D., Voigt, C.C.: The use of automated identification of bat echolocation calls in acoustic monitoring: A cautionary note for a sound analysis. Ecological Indicators 66, 598–602 (2016)
  45. [46]Rydell, J., Nyman, S., Eklöf, J., Jones, G., Russo, D.: Testing the performances of automated identification of bat echolocation calls: A request for prudence. Ecological Indicators 78, 416–420 (Jul 2017). https://doi.org/10.1016/j.ecolind.2017.03. 023
  46. [47]Sánchez-Soto, M.F., Gaona, O., Mercado-Juárez, R., Yanez-Montalvo, A., De León-Lorenzana, A., Borja-Martínez, G., Zaldívar, D., Rodríguez-González, S., Hernández-Villegas, E.N., Moreira-Soto, A., Drexler, J.F., Suzán, G., Vázquez-Domínguez, E., Falcón, L.I.: Ecological-based insights into bat populations in the Yucatán Peninsula under a One Health approach: Coexistence or biophobia. Frontiers in Conservation Science 5, 1488378 (Jan 2025). https://doi.org/10.3389/fcosc.2024.1488378
  47. [48]Schwab, E., Pogrebnoj, S., Freund, M., Flossmann, F., Vogl, S., Frommolt, K.H.: Automated bat call classification using deep convolutional neural networks. Bioacoustics 32(1), 1–16 (2023)
  48. [49]Schwinger, R., Zadeh, P.V., Rauch, L., Kurz, M., Hauschild, T., Lapp, S., Tomforde, S.: Foundation models for bioacoustics—a comparative review. Ecological Informatics p. 103765 (2026)arxiv.org/abs/2508.01277
  49. [50]Siya, A., Matovu, B., Nalukenge, L., Mutebi, M., Nalikka, B., Castle, K., Dewey, T., Williams, K.M., Wickenkamp, N.R., Harris, E.K., Rwego, I.B., Sande, E., Masembe, C., Kading, R.C., Kityo, R.M.: Bat Cave Vulnerability to Anthropogenic Factors: Status and Priorities for Conservation Within the Mount Elgon Region, Uganda. Life 15(12), 1940 (Dec 2025). https://doi.org/10.3390/life15121940
  50. [51]Smeele, S.Q., Hauer, C., Bergler, C., Dechmann, D.K.N., Dietzer, M.T., Elmeros, M., Fjederholt, E.T., Fogato, A., Kohles, J.E., Nöth, E., Brinkløv, S.M.M.: BatSpot: A retrainable neural network for automatic detection and classification of bat echolocation and detection of buzzes and social calls. Biological Conservation 321, 111975 (Sep 2026). https://doi.org/10.1016/j.biocon.2026.111975
  51. [52]Stowell, D.: Computational bioacoustics with deep learning: A review and roadmap. PeerJ 10, e13152 (Mar 2022). https://doi.org/10.7717/peerj.13152
  52. [53]Stowell, D., Wood, M.D., Pamula, H., Stylianou, Y., Glotin, H.: Automatic acoustic detection of birds through deep learning: the first bird audio detection challenge. Methods in Ecology and Evolution 10(3), 368–380 (2019)
  53. [54]Tan, M., Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks. In: Chaudhuri, K., Salakhutdinov, R. (eds) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 6105–6114. PMLR (Jun 2019)arxiv.org/abs/1905.11946
  54. [55]Tuneu-Cuadros, C., Puig-Montserrat, X., Flaquer, C., Mata, V.A., Rebelo, H., Cabeza, M., López-Baucells, A.: Bats and rice: Quantifying the role of insectivorous bats as agricultural pest suppressors in rice fields. Ecosystem Services 66, 101603 (Apr 2024). https://doi.org/10.1016/j.ecoser.2024.101603
  55. [56]UNEP/EUROBATS: Resolution 9.2: Amendment of the annex to the agreement (2022), https://www.eurobats.org/sites/default/files/2025-09/Resolution%209.2%20Amendment%20of%20the%20Annex_0.pdf, ninth Session of the Meeting of the Parties, Brijuni, Croatia, 10–13 October 2022
  56. [57]van Merriënboer, B., Dumoulin, V., Hamer, J., Harrell, L., Burns, A., Denton, T.: Perch 2.0: The Bittern Lesson for Bioacoustics (2025). https://doi.org/10.48550/ARXIV.2508.04665
  57. [58]Vogelbacher, M., Bellafkir, H., Gottwald, J., Schneider, D., Muhling, M., Freisleben, B.: Deep Learning for Recognizing Bat Species and Bat Behavior in Audio Recordings. In: 2023 10th IEEE Swiss Conference on Data Science (SDS). pp. 50–57 (2023). https://doi.org/10.1109/SDS57534.2023.00014
  58. [59]Zualkernan, I., Judas, J., Mahbub, T., Bhagwagar, A., Chand, P.: A Tiny CNN Architecture for Identifying Bat Species from Echolocation Calls. In: 2020 IEEE / ITU International Conference on Artificial Intelligence for Good (AI4G). pp. 81–86 (Sep 2020). https://doi.org/10.1109/AI4G50087.2020.9311084

Paper details

vlnc:W_tn9r37t3
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents