{"paper_id":"487c2cea-6052-4df0-a316-d15616a3c4ab","body_text":"BatSpot: a retrainable neural network for automatic \ndetection and classification of bat echolocation and \ndetection of buzzes and social calls \n \n \nAuthors: Simeon Q. Smeele, Christopher Hauer, Christian Bergler, Dina K. N. Dechmann, \nMelina T. Dietzer, Morten Elmeros, Esben T. Fjederholt, Andrea Fogato, Jenna E. Kohles, \nElmar Nöth, Signe M. M. Brinkløv \n \nAffiliations: \n SQS, ME, ETF, SMMB: Department of Ecoscience, Aarhus University, 8000 C, \nAarhus, Denmark \n CH: Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, \nErlangen, Germany, Faculty of Electrical Engineering, Media and Computer Science, \nOstbayerische Technische Hochschule Amberg-Weiden, Germany \n CB: Faculty of Electrical Engineering, Media and Computer Science, Ostbayerische \nTechnische Hochschule Amberg-Weiden, Germany \nDKND: Max Planck Institute of Animal Behavior, 78315, Radolfzell, Germany; \nDepartment of Biology, University of Konstanz, 78464, Konstanz, Germany; Center for \nAnimal Behavior, Smithsonian Tropical Research Institute, 0843-03092, Panama City, \nPanama \nMTD: Department of Wildlife Ecology and Management, Albert-Ludwigs-Universität \nFreiburg, 79104, Freiburg, Germany \nAF: Department of Migration, Max Planck Institute of Animal Behavior, 78315, \nRadolfzell, Germany; Department of Biology, University of Konstanz, 78464, Konstanz, \nGermany; International Max Planck Research School for Quantitative Behaviour Ecology \nand Evolution, 78464, Konstanz, Germany \nJEK: Center for Animal Behavior, Smithsonian Tropical Research Institute, \n0843-03092, Panama City, Panama; Department for the Ecology of Animal Societies, Max \nPlanck Institute of Animal Behavior, 78467, Konstanz, Germany; Department of Biology, \nUniversity of Konstanz, 78464, Konstanz, Germany \n EN Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, \nErlangen, Germany \n \nORCID: SQS, 0000-0003-1001-66159; CH, 0009-0004-6137-1101; DKND, \n0000-0003-0043-8267; MTD, 0000-0002-1793-5713; AF, 0009-0009-7258-9989; JEK, \n0000-0002-9087-4849; EN, 0000-0002-3396-555X, SMMB, 0000-0002-7289-3536 \n \nSubject areas: conservation \n \nData and code availability \nAll training examples, model files and scripts to produce the results are maintained on \nGitHub: \nhttps://github.com/simeonqs/BatSpot_a_retrainable_neural_network_for_automatic_detectio\nn_and_classification_of_bat_echolocation and a permanent version is shared on Zenodo: \nlink for review: \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nhttps://zenodo.org/records/18607461?preview=1&token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6Ij\nhhM2EwYjNmLTQ0YWItNGU1My1iYTE4LTdmMzZjODZkNGM3NSIsImRhdGEiOnt9LCJyY\nW5kb20iOiI0NzFlYWVjZTVlZDRiOGRhM2ZlMDliNmQ3Mjc5MTY2YiJ9.a8v5Y3t4pgn-f8ViyL\nUZZNEELiIZ3lc4WpI_MG2blO8w0oqqSwSh3a3gPiSWD1_tAK_BIevNcckGuzpLHguAdA. \nThe source code for BatSpot, the GUI and the README describing all steps are maintained \non GitHub: https://github.com/Hauechri/BatSpot. \n \nKeywords: passive acoustic monitoring, deep neural network, bats, conservation, feeding \nbuzz, automated classification \n \nAuthor for correspondence: Simeon Q. Smeele, simeonqs@hotmail.com \n__________________________________________________ \n \nAbstract \n \n1. Bats are a diverse taxonomic group that display a wide range of interesting \nbehaviours. Many bats are keystone species for their ecosystem, are IUCN \nRed-listed as vulnerable to critically endangered, and subject to human-wildlife \nconflicts arising from anthropogenic expansion. Yet bats remain understudied both \nwith respect to behaviour, population ecology and conservation status. One of the \nmajor challenges when studying bats is obtaining data. Their nocturnal lifestyle and \nuse of ultrasonic echolocation makes them difficult to track and record using \ntraditional methods. Recent advances in passive acoustic monitoring have allowed \nresearchers to record large amounts of data, but the detection and classification of \nvocalisations remain a challenge. Most available tools are either for profit or are \nlimited to a narrow geographic range, and mostly focus on echolocation search \nphase calls.  \n2. Here we present BatSpot, a convolutional neural network trained to detect search \nphase calls, buzzes and social calls. It also offers the option to classify the search \nphase calls to species(-complex) level. We provide a GUI that allows researchers to \nretrain or transfer-train the models for their specific needs and validate the \nperformance.  \n3. We test the performance of all models and show that they perform better than both \ncommercial and open-source solutions (search phase file level F1: 0.97 vs 0.96, buzz \ndetector F1: 0.95 vs 0.11). We furthermore show that retraining the search phase call \ndetector for a new country with examples from just 59 recordings massively improves \nthe performance (F1: 0.48 to 0.79).  \n4. BatSpot will enable bat researchers globally to automate detection and classification \nwith minimal effort and includes novel options for social call and buzz detection, \ntypically not featured in other automated tools for bat monitoring.  \n \nIntroduction \n \nBats are the second most diverse group of mammals (Burgin et al., 2018) with 1,500 species \ndescribed (Simmons & Cirranello, 2020), and display a wide variety of complex vocal \nbehaviours. They actively sense their surroundings using echolocation during both individual \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nand social foraging (Jones & Holderied, 2007; Petrites et al., 2009; Egert-Berg & Hurme et \nal., 2018; Krivoruchkoetal et al., 2024), and display complex vocalizations related to social \nstructure and mating (Knörnschild et al., 2017; Knörnschild et al., 2019; Smotherman et al., \n2016). Bats are keystone species that provide important ecological services such as pest \ncontrol, soil fertilisation, seed dispersal and flower pollination (Ghanem & Voigt, 2012; \nAncillotto et al., 2017;Ramírez‐ fráncel et al., 2021). Loss of bat ecosystem services can \nhave direct and profound effects on human health (Frank, 2024). Globally, numerous bat \nspecies are currently threatened by extinction (IUCN, 2024) and all European bat species \nare listed on Annex IV of the European Union’s Habitats Directive (92/43/EEC), requiring \nmember states to monitor them and ensure favourable population status. There is increasing \nevidence that migrating bat species face increased mortality from expanding wind farms \n(Arnett et al., 2015; Voigt et al., 2022), as countries around the world have to balance the \nconservation of flying species against the need to shift energy production to renewable \nsources. Other threats that bat species face include climate change (Sherwin et al., 2012), \ndisease (Hoyt et al., 2021) and declining insect populations (Forister et al., 2019). \nAddressing these challenges requires the development of robust methods for monitoring bat \nbehaviour, activity, and populations.  \n \nRecording behaviour of bats which fly in the dark is challenging and monitoring population \nsizes becomes especially difficult for bats that roost and breed in cavities and potentially \nmigrate long distances (Russo et al., 2018; Martínez-Fonseca et al., 2024; Rebolo et al., \n2024). The most cost-efficient way to record bat activity on a large scale is the use of \npassive acoustic monitoring (PAM), which has become a well established method (Biffi et al., \n2024, Roemer et al., 2025, Toshkova et al., 2025). Most bat species produce ultrasonic \necholocation calls and social vocalisations that may span the audible and ultrasonic range. \nEcholocation calls can be assigned to three main categories. Search phase calls are usually \nspecies-specific but adaptable to the environment, with longer calls characteristic for open \nspace foraging and commuting and shorter call duration typical in cluttered environments \n(Kalko and Schnitzler, 2001; Brinkløv et al., 2010). Approach calls are similar to calls \nproduced in clutter, but might shift in frequency. The call interval (also called pulse interval) is \nalso much reduced to yield quicker updates about prey or obstacle position. Finally, buzzes \nare produced during the very end of a prey interception, with much reduced call amplitude \nand call intervals. Social vocalisations are much more variable in some species and consist \nof longer multi-node sequences resembling bird song (Smotherman et al., 2016; Springall et \nal., 2019). Species can in some cases be recognised by the frequency range and \nmodulation, the duration and the intervals between calls, while specific behaviours can \nsometimes be inferred from the type of call: buzzes mostly indicating feeding, social calls \nindicating varying social behaviours such as mate attraction, territorial defence and individual \nrecognition (Knörnschild et al., 2017; Knörnschild et al., 2019; Smotherman et al., 2006).  \n \nWhile PAM is already effectively supplying vast amounts of audio data, the processing \nworkflow often lags in efficiency. Approaches range from manual detection and classification \nof bat calls (Sugai et al., 2018) to more or less automated pipelines with or without manual \nvalidation of software performance. The applicability of commercial software is challenged by \nprice and decreased transparency. Open source software also exists, but often comes \ntrained for a limited geographic range focused outside the Global South, where bat species \ndiversity is highest, and might not be easy to adapt and use without a computer science \nbackground. An exception to this is BatDetect2 (Aodha et al., 2022b), which contains \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\npre-trained models for several countries, including two in the Global South. It does, however, \nrequire some familiarity running scripts from the console, especially if fine-tuning of the \nmodel is needed. For all these software, retrainability remains an issue. If used in a \nsoundscape with new noise sources or bat species, performance is likely to drop. Retraining \nwithout coding experience is often not possible. Finally, all but one of the open source \nsolutions have focused on search phase calls, with the exception of Jameson (2024), who \ndeveloped a buzz detector: buzzfindr for Canadian species. This lack of tools to detect social \ncalls and buzzes globally hinders bat research, since foraging or social activity are often \nimportant behavioural components in a monitoring context (Dietzer et al., 2024).  \n \nTo address this, we have trained several versions of BatSpot, which is based on \nANIMAL-SPOT, a convolutional neural network designed to detect and classify animal \nvocalisations (Bergler et al., 2022). We present detection models for bat echolocation search \nphase calls, buzzes and social calls, as well as a classification model for the search phase \ncalls. We compare the performance of these models to several commercial and open source \nsolutions on a validation dataset. We provide examples on how to validate the models for a \nnew location and how to retrain the model with a reduced amount of training data. The \nmodel also allows for transfer learning, adding additional classes to the classification model. \nFinally, we provide a GUI (see Figure 1) and a README on how to train, retrain, \ntransfer-train and validate models.  \n \n \nFigure 1 Screenshot of the GUI. \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nMethods \n \nCall detector \nTo detect search phase calls in continuous recordings, we trained a binary version of \nBatSpot (hereafter called ‘call detector’). We used 12,396 noise examples and 10,766 target \nexamples (search phase calls) from Denmark. We also used additional 2,000 noise \nexamples for noise augmentation (these are added randomly in the background to increase \ngeneralisation ability). These were chosen randomly from the original noise set (with 14,396 \nnoise examples). The model was trained on clips of 20 ms duration. Longer examples were \nrandomly cut, shorter examples were zero-buffered. The frequency range was 1-95 kHz, this \ncovers the high-energy part of all calls and is enough for detection. We report the validation \nand test accuracy, which is the proportion of true positives and true negatives out of the total \nnumber of validation/test examples. The reported validation accuracy is the highest accuracy \nduring the training process, where the validation set is used to test if the model should stop \ntraining. The test accuracy is the accuracy on the test set, which is not shown to the model \nuntil after training has stopped. \n \nTo validate the model performance on full recordings we used 210 recordings from Denmark \nof variable duration. None of these had been used to generate training data. All search \nphase calls (including faint calls) were manually annotated. The call detector was run on \nwindows of 20 ms, with 10 ms overlap between windows. The threshold for considering a \ndetection true was 0.5. For each detection BatSpot gives a probability that it is correct in the \nprediction and by choosing 0.5 we allow only predictions where this probability is at least 0.5. \nSince the call detector and classifier are often used together, validation results are also \npresented together. For more detailed information, see below. \n \nCall classifier \nTo classify the detections from the call detector, we trained a multi-class version of BatSpot \n(hereafter called ‘call classifier’) including call examples of 12 species(-complexes). We used \n1,161 examples of Barbastella barbastellus, 2,840 examples of Eptesicus serotinus (now \nknown as Chenaphaeus serotinus), 1,217 examples of Myotis brandtii/mystacinus, 1,503 \nexamples of Myotis dasycneme, 2,401 examples of Myotis daubentonii, 1,325 examples of \nMyotis nattereri, 2,397 examples of Nyctalus noctula, 982 examples of Plecotus auritus, \n1,226 examples of Pipistrellus nathusii, 1,145 examples of Pipistrellus pipistrellus, 2,374 \nexamples of Pipistrellus pygmaeus and 1,517 examples of Vespertilio murinus. Examples \nwere drawn from the recordings made during the Danish national monitoring programme, \nwhere the focal individual is either observed in the field during recording or, if data quality \nallows, annotated by at least two experts. We also included 551 examples of buzzes, 899 \nexamples of approach phase calls and 439 examples of social calls. These categories were \nnot annotated to species level and considered as noise, since the call detector model was \nonly supposed to detect search phase calls. To get optimal performance the approach phase \ncall category was encoded as the noise category, serving as the default, if none of the other \ncategories is predicted with high enough probability. The model was trained on clips of 20 \nms duration. Longer examples were randomly cut, shorter examples were zero-buffered. \nThe frequency range was 10-125 kHz, which meant that we had to use a slightly modified \nversion of BatSpot, where the files were not downsampled to 192 kHz during processing, but \nrather to 250 kHz. This was done, since some species have calls extending beyond 96 kHz, \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\ninformation that can be used by the model during classification. We report the validation and \ntest accuracy, which is the proportion of correct classifications out of the total number of \nvalidation/test examples. \n \nTo validate the model on detections from full recordings, we used the same validation set as \nfor the call detector, and ran the call classifier on all the detections from the call detector. \nSince several Myotis species are notoriously difficult to discriminate from each other, we \npooled both the validation and classifications from the model into the genus level label M. \nDepending on context, the same may be true for the low frequency species E. serotinus, N. \nnoctula and V. murinus, thus these were pooled into an ENV complex. Calls assigned to B. \nbarbastellus were considered noise, since they should not occur in the area where the \nvalidation data was collected. If calls could not be assigned to a species (-complex) or \nmultiple calls were overlapping, these were excluded from validation. The call classifier was \nrun on windows of 20 ms, with 15 ms overlap between windows. The overlap was increased \nto increase the chance of the classifier getting a window with the whole call included. The \nthreshold for considering a detection true was 0.5. To quantify performance we created a \nconfusion matrix and computed performance statistics. For the detection step we computed: \n(1) recall = tp/(tp+fn), which gives the fraction of calls that were detected out of the total \nnumber of calls; (2) precision = tp/(tp+fp), which gives the fraction of correct detections out of \nall detections; and (3) F1 = 2 * (recall*precision)/(recall+precision), which gives the average \nacross the two; where tp = number of true positives, tn = number of true negatives, fp = \nnumber of false positives and fn = number of false negatives. For the classification step we \ncomputed: accuracy = correct classifications / (correct classifications + mistakes), which also \nincluded classifying the false positives of the detection model (where the correct \nclassification was noise). For the overall performance, we computed: accuracy = correct \nclassifications / (correct classifications + mistakes + misses), where the misses were the \nfalse negatives from the detection model. Since some projects only need to know if a \nspecies is present in a file or not, we also summarised the output per file by taking the \nspecies with most calls predicted and created a confusion matrix at file level, where we \nexcluded 10 out of the 210 files, as they contained calls that could not be assigned within \nPipistrellus (for the call level performance only these few calls were excluded and all other \ncalls in the file were still used). We opted not to include area-under-the-curve values for \neither analysis, as these are harder to interpret. \n \nTo compare the combined performance of the BatSpot call detector and classifier to existing \nsoftware, we ran the open source software BSG-BAT (Meramo et al., 2025), BatDetect2 \n(Aodha et al., 2022) and BAT (Fundel et al., 2023), as well as the commercial softwares \nSonochiro (version 4.1.4, Biotope, France, biotope.fr), Kaleidoscope (version 5.8.1, Wildlife \nAcoustics Inc, USA, wildlifeacoustics.com) and BTO bioacoustics pipeline (BTO Acoustic \nPipeline, n.d.). We used the same validation set as for the validation of BatSpot. In files with \none or multiple calls that could not be assigned to a species (-complex), it would not be \npossible to assign a label at the file level. For each software we used default settings and \nfollowed best practise described in the manual. For BatDetect2, we changed the threshold \nfor accepting a detection to 0.4 to balance the recall and precision. For each software, we \ncompared the primary detection and classification to the ground truth (manual annotation by \nan expert) at file level. Where the software made multiple predictions, only the prediction \nwith highest probability was used, since each of the validation files only contained a single \nspecies. If the software predicted species not present in the study area or did not classify the \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\ndetection, the file was considered to contain only noise. For each software we show the \nconfusion matrix and compute the recall, precision and F1 score.  \n \nBuzz detector \nTo detect buzzes in continuous recordings, we trained a binary version of BatSpot. We used \n2,293 noise examples and 1,288 target examples from Denmark, Germany and Panama. \nWe also used 200 additional noise examples for noise augmentation. These were chosen \nrandomly from the original noise set (of 1,488 noise examples). The model was trained on \nclips of 200 ms duration. Longer examples were randomly cut, shorter examples were \nzero-buffered. The frequency range was 1-95 kHz, which contains the part of the buzz with \nthe most energy for most species. We report the validation and test accuracy, which is the \nproportion of true positives and true negatives out of the total number of validation/test \nexamples. \n \nTo validate model performance we used 30 recordings from Denmark of variable duration \n(3-15 s), 100 recordings of 55 s from Germany and 90 recordings from Panama of variable \nduration (0.2-0.9 s). None of these had been used to generate training data. All buzzes \n(including faint and incomplete buzzes) were manually annotated. The buzz detector was \nrun on windows of 200 ms, with 100 ms overlap between windows. The threshold for \nconsidering a detection true was 0.5 for Denmark and Panama, but was set to 0.9 for \nGermany to balance recall and precision. For validation we computed: recall = tp/(tp+fn), \nprecision = tp/(tp+fp), F1 = 2 * (recall*precision)/(recall+precision), where tp = number of true \npositives, tn = number of true negatives, fp = number of false positives and fn = number of \nfalse negatives. \n \nTo compare our buzz detector to existing tools we ran the function buzzfindr from the R \npackage buzzfindr (Jameson, 2024) on the same validation recordings using the default \nsettings. We show the confusion matrix and compute the recall, precision and F1 score.  \n \nSocial call detector \nTo detect social calls in continuous recordings, we trained a binary version of BatSpot. We \nused 1,296 noise examples and 1,412 target examples from Denmark. We did not use noise \naugmentation. The model was trained on clips of 75 ms duration. Longer examples were \nrandomly cut, shorter examples were zero-buffered. The frequency range was 1-60 kHz. We \nreport the validation and test accuracy, which is the proportion of true positives and true \nnegatives out of the total number of validation/test examples. \n \nTo validate the model performance we used 19 recordings from Denmark of variable \nduration. None of these had been used to generate training data. All social calls (including \nfaint calls) were manually annotated. If a social call was not clearly a single unit (parts \nseparated by ca. 50 ms), each part was separately annotated. The social call detector was \nrun on windows of 75 ms, with 50 ms overlap between windows. The threshold for \nconsidering a detection true was 0.9. For validation we computed: recall = tp/(tp+fn), \nprecision = tp/(tp+fp), F1 = 2 * (recall*precision)/(recall+precision), where tp = number of true \npositives, tn = number of true negatives, fp = number of false positives and fn = number of \nfalse negatives. \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nRetraining \nRetraining is the use of a previously trained model to speed up and/or improve performance \nof a new model. For retraining, the weights of the old model are imported after which new \ntraining examples are used to (slightly) update these weights. To illustrate how retraining \nincreases the performance, we ran the existing call detector on a validation set of 24 files of \n5 s each from Germany. We then created a training set (using different recordings) \nconsisting of 290 noise examples and 477 target examples from Germany. We also used an \nadditional 50 noise examples for noise augmentation. These were chosen randomly from the \noriginal noise set (of 340 noise examples). With that we first trained a model from scratch \nusing the same settings and then trained a model using retraining from the Danish call \ndetector model. We restricted the number of learning rounds (epochs) to 100, but did not \nchange any other settings. To compare we present the recall, precision and F1 for all three \nmodels (baseline, retrained and trained from scratch). \n \nResults \n \nCall detector and classifier \nAfter 150 epochs (the maximum) the call detector had a validation accuracy of 0.98 and a \ntest accuracy of 0.97. The call classifier trained for 81 epochs (after no improvement for 20 \nepochs) and achieved a validation accuracy of 0.88 and a test accuracy of 0.87. For \nvalidation of the models on full recordings, we annotated 6,477 search phase calls. Of these, \n40 could not be assigned a species, 74 were overlapping, 193 were either P. nathusii or P. \npipistrellus and 490 were either P. pipistrellus or P. pygmeus, and were therefore excluded \nfrom validation. The detector performed well on the validation set, which balanced recall \n(how well the model finds calls) and precision (how many detections are correct) (see Figure \n2) and showed no strong bias towards missing certain species. The classification model also \nperformed well, with most errors being due to confusion of P. nathusii and P. pipistrellus or \nassigning false positives to ENV or Myotis, likely due to low frequency and click-like noise \nboth confusing the call detector and classifier.  \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\n \nFigure 2 Overall confusion matrix for the call detector and classifier. Rows are the manual \nannotations, columns are the results from the models for each call. Ppyg = Pipistrellus \npygmeus, Ppip = P. pipistrellus, Pnat = P. nathusii, Paur = Plecotus auritus, M = Myotis sp., \nENV = Eptesicus serotinus/Nyctalus noctula/Vespertillio murinus, -noise- = calls annotated \nas noise or classified as buzz/approach/social/Bbar, -missed- = calls not detected by the call \ndetector. \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\n \nOverall BatSpot outperformed all other tested software (see Figure 3 and 4). BatDetect2 and \nthe BTO bioacoustics pipeline came close in balancing the recall and precision as well as \naccurately classifying the files with detections (see Table 1).  \n \nFigure 3 Overall confusion matrix at file level for the call detector and classifier. Rows are \nthe manual annotations, columns are the results from the models for each call. Ppyg = \nPipistrellus pygmeus, Ppip = P. pipistrellus, Pnat = P. nathusii, Paur = Plecotus auritus, M = \nMyotis sp., ENV = Eptesicus serotinus/Nyctalus noctula/Vespertillio murinus, -noise- = calls \nannotated as noise or classified as buzz/approach/social/Bbar, -missed- = calls not detected \nby the call detector.  \n \n \n  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\n \nFigure 4 Confusion matrices for BSG-BAT, BatDetect2, BAT, Sonochiro, Kaleidoscope and \nBTO bioacoustics pipeline. Rows are the manual annotations, columns are the results from \nthe different software at file level. Ppyg = Pipistrellus pygmeus, Ppip = P. pipistrellus, Pnat = \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nP. nathusii, Paur = Plecotus auritus, M = Myotis sp., ENV = Eptesicus serotinus/Nyctalus \nnoctula/Vespertillio murinus, -noise- = files annotated as containing noise or any other bat \nspecies. \n \nTable 1 Performance of all tested software on the Danish validation set. Recall = tp/(tp+fn), \nprecision = tp/(tp+fp), F1 = 2 * (recall*precision)/(recall+precision) and accuracy = \n(tp+tn)/(tp+tn+fp+fn) for all files with detections; where tp = number of true positives, tn = \nnumber of true negatives, fp = number of false positives and fn = number of false negatives. \nThe highest value(s) in each column is/are highlighted in bold. \n Detection Classification \nSoftware Recall Precision F1 Accuracy \nBatSpot 1.00 0.95 0.97 0.95 \nBSG-BAT 0.91 0.87 0.89 0.80 \nBatDetect2 1.00 0.92 0.96 0.92 \nBAT 0.48 0.76 0.59 0.58 \nSonochiro 0.79 0.97 0.88 0.84 \nKaleidoscope 0.68 0.97 0.80 0.77 \nBTO 0.87 0.99 0.92 0.92 \n \n \nBuzz detector \nAfter 100 epochs (the maximum) the model had a validation accuracy of 0.97 and a test \naccuracy of 0.96. For the validation recordings performance was generally very high (F1: \n0.95, recall: 0.96, precision: 0.93). Performance was highest for the Panamanian recordings \n(F1: 0.97, recall: 0.95, precision: 0.99), intermediate for the Danish recordings (F1: 0.88, \nrecall: 0.85, precision: 0.92) and lowest for the German recordings (F1: 0.84, recall: 0.94, \nprecision: 0.77). Despite the increased threshold for accepting detections (0.9 instead of \n0.5), the model detected a lot of false positives for the German recordings. \n \nBuzzfindr had much lower performance on the same validation data (F1: 0.37, recall: 0.28, \nprecision: 0.54). \n \nSocial call detector \nAfter 150 epochs (the maximum) the model had a validation accuracy of 0.99 and a test \naccuracy of 0.98. For the validation recordings performance was lower but still good given \nthe limited dataset (F1: 0.91, recall: 0.93, precision: 0.98). \n \nRetraining \nThe call detector trained on only Danish data did not perform very well on the German \nvalidation set (F1: 0.48, recall: 0.88, precision: 0.33). Retraining clearly improved the \nprecision (F1: 0.79, recall: 0.79, precision: 0.78) and led to slightly better performance than \ntraining from scratch (F1: 0.73, recall: 0.83, precision: 0.64). \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\n \nDiscussion \n \nOur aim was to develop a tool that can both detect the three most common vocalisation \ntypes of bats (search phase calls, feeding buzzes and social calls) and classify search phase \ncalls, the vocalisation type most commonly used for species identification and monitoring \nstudies (Schnitzler & Kalko, 2001; Russo et al., 2018). We present pre-trained models with \nreasonable out-of-the-box performance and the option to retrain the models with minimal \ntime investment. Finally, we show that the tool can also be used in the Global South. The \nthree detection models presented achieved good performance for the situations that they \nhad been trained for. The buzz detector in particular showed that a single model could detect \nbuzzes in mixed habitats recordings in Denmark, Germany and Panama, with very different \nnoise backgrounds and species compositions.  \n \nWhen applying the search phase call detector model trained exclusively on Danish data to \nGerman recordings, we observed a clear drop in performance. This is to be expected, since \nthe noise sources can be very different, even within the same country. This is also clearly \nwhat drove the drop in performance, with precision (the proportion of detections that \ncontained search phase calls) dropping much more than recall. Of course, if the detector is \nused in a country with not only new species (as in Germany), but also new genera or even \nfamilies that produce search phase calls that do not resemble any of the training examples, \na drop in recall is also to be expected. In both cases, retraining will be needed until a dataset \nwith enough examples across the globe becomes available and a global bat detector can be \ntrained. Our hope is that the data presented with this paper can contribute to such a global \ndataset, and that until a global model is a reality, our Danish model can serve as a basis \nmodel to retrain detectors for new locations.  \n \nClassification is a more challenging task, and one where even experts with decades of \nexperience cannot always assign call sequences of good quality to a species (Jennings et al. \n2008; Rydell et al. 2017). This is simply because some species produce very similar calls \nwhen flying in similar environments, and because a given species is often able to modify \ncalls when flying in a different environment. We therefore opted to follow the standard \nprocedure of grouping calls from Myotis species into a single category and also grouping \ncalls from species producing lower frequency calls into an ENV species-complex. The \nclassification model was trained with each species separately (with the exception of M. \nbrandtii and M. mystacinus which could not be separated when generating our training data) \nto allow the model to learn as many distinguishing features as possible. The grouping was \nthen done after the prediction step, since performance within groups was low as expected. \nAlthough an overall accuracy of 0.80 seems low, it should be kept in mind that this includes \ndetection and classification of even the faintest of calls. For many projects (e.g. \nenvironmental impact assessments), the aim is to detect and classify species at the file level \n(as presence/absence in that file), and this is where performance is much higher (F1 = 0.97). \n \nThere are also projects (e.g. those aiming to describe the characteristics of calls under \nvarying conditions), in which exact detection of start and end times for individual calls is \nimportant. Additionally, when detecting buzzes or social calls, one might want a very low \nfalse negative rate, to ensure most target events are detected. This can be accomplished by \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nlowering the detection threshold and subsequently validating all detections manually to get \nrid of the false positives. Alternatively, one might want a low false positive rate, something \nparticularly relevant when many calls are produced in a sequence and post-hoc processing \ncan be done on the detections. One way to do this is to set up a decision tree, in which, for \nexample, a file is only considered to contain calls from a species if the classifier assigns at \nleast five calls to that species within a sliding window of 3 seconds. This eliminates any false \npositives that occur spaced out in time.  \n \nBatSpot offers a GUI and README for how to retrain, to facilitate obtaining a model with \ngood performance in new recording situations, which is an important step forward and allows \nresearchers to optimise the models without coding experience. An example of this is the \noffshore environment, where boat, buoy and water noise create challenges to both the \ndetection and classification step by introducing new noise types. With the increase of wind \nfarm development offshore, and the need for both pre- and post-construction monitoring, \nmodels performing well in the offshore environment are particularly important. This is one of \nthe reasons that the search phase call detector in BatSpot was trained with a large portion of \nexamples from buoys and wind turbines in the Danish North Sea. That said, BatSpot is far \nfrom the first neural network to detect and classify search phase calls, although to the best of \nour knowledge, it is the first to additionally detect buzzes and social calls. BatDetect2 is a \nvery promising alternative for search phase calls, and achieves comparable performance \nout-of-the-box. One potential reason for BatDetect2’s high performance is the inclusion of a \nself-attention layer, which enables the model to integrate information of multiple calls in a \nsequence to predict the species producing individual calls. With the integration into a simple \npython package, BatDetect2 requires minimal coding experience, and might therefore be the \nbest option if there is no time for targeted model retraining. \n \nFor buzzes and social calls, BatSpot presents a much bigger step forward. The detection of \nbuzzes is important when looking at foraging behaviour, but has previously received almost \nno attention in the literature. The R package buzzfindr (Jameson, 2024) is an important step \nforward, but struggles to detect faint buzzes and buzzes in noisy environments, something \nimportant in settings with multiple individuals flying in an environment with much insect noise \nas well (as is the case for many of the German and Panamanian recordings). The fact that a \nsingle model was able to perform well in three countries across two continents is a major \nstep forward. Social calls have received even less attention and are also particularly difficult \nto detect automatically, because this category basically includes all vocalisations not used to \ndetect the surroundings. With our current model in BatSpot, it should be possible to reliably \ndetect longer social calls, such as the trills produced by species we have evaluated in the \nPipistrellus genus. The data presented with the paper will also serve as an important \ncontribution to a future comprehensive dataset with recordings from across Europe or wider. \nWe envision that the social call detector, once further developed, will be able to support \npassive acoustic monitoring purposes by enabling automated analysis and localisation of \nimportant roosting and mating areas. \n \nFor all three models, there are several steps to keep in mind when applying them to large \ndatasets. First, a validation set is necessary to quantify the recall and precision for each \nlocation and season included in the dataset to be analysed. This does not need to be a \nspecific percentage of all annotated data, but should rather ensure that all variation in the full \ndataset is covered and performance on it can be quantified. Validating performance is not \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nspecific to BatSpot, as even established commercial software might show a strong bias \nwhen applied in, for example, the offshore environment (Smeele et al. 2026). If retraining is \nneeded, this can be done by following the steps in the README (see below). This should \ndrastically reduce training time compared to training with all the published data plus the data \ncollected in the specific project (Verwimp et al., 2025). And finally, once the model performs \nsatisfactorily, all raw data can be processed using the scripts supplied to index and process \nlarge datasets on a high performance computing cluster, or a local computer with GPU.  \n \nConclusions \nIn this article we presented BatSpot, a set of convolutional neural networks to detect search \nphase calls, buzzes and social calls of bats as well as classify the search phase calls. \nInteraction with the models is facilitated through a GUI, but the modified ANIMAL-SPOT \nsource code is still in place, allowing advanced users to run models from the terminal as \nwell. We show that the search phase call detector can be retrained with minimal effort and \nthat the buzz detector works well across three countries and two continents. Overall, \nBatSpot will enable researchers with no coding experience to tap into the flexibility and \npower of convolutional neural networks, while at the same time contributing to a growing \nopen source set of annotated recordings, which can be used to train globally applicable \nmodels. \n \nCompeting interests \nAll authors declare they have no competing interests.  \n \nBibliography \n \nAncillotto, L., Ariano, A., Nardone, V., Budinski, I., Rydell, J., & Russo, D. (2017). Effects of \nfree-ranging cattle and landscape complexity on bat foraging: implications for bat \nconservation and livestock management. Agriculture, Ecosystems & Environment, 241, \n54–61. \nAodha, O. M., Martínez Balvanera, S., Damstra, E., Cooke, M., Eichinski, P., Browning, E., \nBarataud, M., Boughey, K., Coles, R., Giacomini, G., Swiney G., M. C. M., Obrist, M. K., \nParsons, S., Sattler, T., & Jones, K. E. (2022). Towards a general approach for bat \necholocation detection and classification. Cold Spring Harbor Laboratory. \nhttps://doi.org/10.1101/2022.12.14.520490 \nArnett, E. B., Baerwald, E. F., Mathews, F., Rodrigues, L., Rodríguez-Durán, A., Rydell, J., \nVillegas-Patraca, R., & Voigt, C. C. (2015). Impacts of wind energy development on bats: A \nglobal perspective. In Bats in the Anthropocene: Conservation of Bats in a Changing World \n(pp. 295–323). Springer International Publishing. \nhttps://doi.org/10.1007/978-3-319-25220-9_11 \nBergler, C., Smeele, S. Q., Tyndel, S. A., Barnhill, A., Ortiz, S. T., Kalan, A. K., Cheng, R. X., \nBrinkløv, S., Osiecka, A. N., Tougaard, J., Jakobsen, F., Wahlberg, M., Nöth, E., Maier, A., & \nKlump, B. C. (2022). ANIMAL-SPOT enables animal-independent signal detection and \nclassification using deep learning. Scientific Reports, 12(1). \nhttps://doi.org/10.1038/s41598-022-26429-y \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nBiffi, S., Chapman, P. J., Engler, J. O., Kunin, W. E., & Ziv, G. (2024). Using automated \npassive acoustic monitoring to measure changes in bird and bat vocal activity around \nhedgerows of different ages. Biological Conservation, 296, 110722. \nhttps://doi.org/10.1016/j.biocon.2024.110722 \nBrinkløv, S., Kalko, E. K. V., & Surlykke, A. (2010). Dynamic adjustment of biosonar intensity \nto habitat clutter in the bat Macrophyllum macrophyllum (Phyllostomidae). Behavioral \nEcology and Sociobiology, 64(11), 1867–1874. https://doi.org/10.1007/s00265-010-0998-9 \nBTO acoustic pipeline. (n.d.). BTO. Retrieved December 2, 2025, from \nhttps://www.bto.org/data/tools-products/acoustic-pipeline \nBurgin, C. J., Colella, J. P., Kahn, P. L., & Upham, N. S. (2018). How many species of \nmammals are there? Journal of Mammalogy, 99(1), 1–14. \nDietzer, M. T., Keicher, L., Kohles, J. E., Hurme, E. R., Ruczyński, I., Borowik, T., Zegarek, \nM., Choiński, M., & Dechmann, D. K. N. (2024). High temporal resolution data reveal low bat \nand insect activity over managed meadows in central Europe. Scientific Reports, 14(1). \nhttps://doi.org/10.1038/s41598-024-57915-0 \nEgert-Berg, K., Hurme, E. R., Greif, S., Goldstein, A., Harten, L., Herrera M., L. G., \nFlores-Martínez, J. J., Valdés, A. T., Johnston, D. S., Eitan, O., Borissov, I., Shipley, J. R., \nMedellin, R. A., Wilkinson, G. S., Goerlitz, H. R., & Yovel, Y. (2018). Resource ephemerality \ndrives social foraging in bats. Current Biology, 28(22), 3667-3673.e5. \nhttps://doi.org/10.1016/j.cub.2018.09.064 \nForister, M. L., Pelton, E. M., & Black, S. H. (2019). Declines in insect abundance and \ndiversity: We know enough to act now. Conservation Science and Practice, 1(8). \nhttps://doi.org/10.1111/csp2.80 \nFrank, E. G. (2024). The economic impacts of ecosystem disruptions: Costs from \nsubstituting biological pest control. Science, 385(6713). \nFundel, F., Braun, D. A., & Gottwald, S. (2023). Automatic bat call classification using \ntransformer networks. Ecological Informatics, 78, 102288. \nhttps://doi.org/10.1016/j.ecoinf.2023.102288 \nGhanem, S. J., & Voigt, C. C. (2012). Increasing awareness of ecosystem services provided \nby bats. In Advances in the Study of Behavior (pp. 279–302). Elsevier. \nHoyt, J. R., Kilpatrick, A. M., & Langwig, K. E. (2021). Ecology and impacts of white-nose \nsyndrome on bats. Nature Reviews Microbiology, 19(3), 196–210. \nhttps://doi.org/10.1038/s41579-020-00493-5 \nIUCN. (2024). 2. The IUCN Red List of Threatened Species. https://www.iucnredlist.org.  \nJameson, J. W. (2024). Buzzfindr: Automating the detection of feeding buzzes in bat \necholocation recordings. PLOS ONE, 19(8), e0306063. \nhttps://doi.org/10.1371/journal.pone.0306063 \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nJennings, N., Parsons, S., & Pocock, M. J. O. (2008). Human vs. machine: Identification of \nbat species from their echolocation calls by humans and by artificial neural networks. \nCanadian Journal of Zoology, 86(5), 371–377. https://doi.org/10.1139/z08-009 \nKnörnschild, M., Blüml, S., Steidl, P., Eckenweber, M., & Nagy, M. (2017). Bat songs as \nacoustic beacons - Male territorial songs attract dispersing females. Scientific Reports, 7(1). \nhttps://doi.org/10.1038/s41598-017-14434-5 \nKnörnschild, M., Fernandez, A. A., & Nagy, M. (2019). Vocal information and the navigation \nof social decisions in bats: Is social complexity linked to vocal complexity? Functional \nEcology, 34(2), 322–331. https://doi.org/10.1111/1365-2435.13407 \nKrivoruchko, K., Koblitz, J. C., Goldshtein, A., Biljman, K., Guillén-Servent, A., & Yovel, Y. \n(2024). A social foraging trade-off in echolocating bats reveals that they benefit from some \nconspecifics but are impaired when many are around. Proceedings of the National Academy \nof Sciences, 121(30). https://doi.org/10.1073/pnas.2321724121 \nMartínez-Fonseca, J. G., Westeen, E. P., Wan, H. Y., Cushman, S. A., & Chambers, C. L. \n(2024). A global review of landscape-scale analyses in bats reveals geographic and \ntaxonomic biases and opportunities for novel research. Biological Conservation, 299, \n110829. https://doi.org/10.1016/j.biocon.2024.110829 \nMeramo, K., Somervuo, P., Rannisto, M., Lauha, P., López‐Baucells, A., Gili, F., Roslin, T., \nLilley, T. M., & Ovaskainen, O. (2025). BSG‐BATS: An open‐source data annotation portal \nand classifier for European bat vocalizations. Methods in Ecology and Evolution. \nhttps://doi.org/10.1111/2041-210x.70220 \nNevard, L. (2021). Social foraging in vampire bats is predicted by long-term cooperative \nrelationships. The Company of Biologists. https://doi.org/10.1242/prelights.29227 \nRamírez‐fráncel, L. A., García‐herrera, L. V., Losada‐prado, S., Reinoso‐flórez, G., \nSánchez‐hernández, A., Estrada‐villegas, S., Lim, B. K., & Guevara, G. (2021). Bats and \ntheir vital ecosystem services: A global review. Integrative Zoology, 17(1), 2–23. \nhttps://doi.org/10.1111/1749-4877.12552 \nRebolo, N., Lois, N., & Lambertucci, S. A. (2024). Wind energy development in latin america \nand the caribbean: Risk assessment for flying vertebrates. Elsevier BV. \nhttps://doi.org/10.2139/ssrn.4932072 \nRoemer, C., Haquart, A., López‐Baucells, A., & Besnard, A. (2025). Current frontiers in the \npassive acoustic monitoring of bats. Methods in Ecology and Evolution, 16(11), 2534–2544. \nhttps://doi.org/10.1111/2041-210x.70157 \nRusso, D., Ancillotto, L., & Jones, G. (2018). Bats are still not birds in the digital era: \nEcholocation call variation and why it matters for bat species identification. Canadian Journal \nof Zoology, 96(2), 63–78. https://doi.org/10.1139/cjz-2017-0089 \nSchnitzler, H.-U., & Kalko, E. K. V. (2001). Echolocation by insect-eating bats. BioScience, \n51(7), 557. https://doi.org/10.1641/0006-3568(2001)051[0557:ebieb]2.0.co;2 \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint \n\nSherwin, H. A., Montgomery, W. I., & Lundy, M. G. (2012). The impact and implications of \nclimate change for bats. Mammal Review, 43(3), 171–182. \nhttps://doi.org/10.1111/j.1365-2907.2012.00214.x \nSimmons, N. B., & Cirranello, A. L. (2020). Bat Species of the World: A taxonomic and \ngeographic database. Bats of the World. \nSmeele, S. Q., Grethen, K. J., Brinkløv, S. M. M., Fjederholt, E. T., & Elmeros, M. (2026). \nModelling offshore bat activity over Danish waters. Aarhus University, DCE - Danish Centre \nfor Environment and Energy. \nhttps://dce.au.dk/fileadmin/dce.au.dk/Udgivelser/Videnskabelige_rapporter_600-699/SR691.\npdf \nSmotherman, M., Knörnschild, M., Smarsh, G., & Bohn, K. (2016a). The origins and diversity \nof bat songs. Journal of Comparative Physiology A, 202(8), 535–554. \nhttps://doi.org/10.1007/s00359-016-1105-0 \nSmotherman, M., Knörnschild, M., Smarsh, G., & Bohn, K. (2016b). The origins and diversity \nof bat songs. Journal of Comparative Physiology A, 202(8), 535–554. \nhttps://doi.org/10.1007/s00359-016-1105-0 \nSpringall, B. T., Li, H., & Kalcounis-Rueppell, M. C. (2019). The in-flight social calls of \ninsectivorous bats: Species specific behaviors and contexts of social call production. \nFrontiers in Ecology and Evolution, 7. https://doi.org/10.3389/fevo.2019.00441 \nStidsholt, L., Tena, E., Foskolos, I., Nogueras, J., de la Hera, I., Sánchez-Navarro, S., \nGarcía-Mudarra, J. L., & Ibáñez, C. (2025). Greater noctule bats prey on and consume \npasserines in flight. Science, 390(6769), 178–181. https://doi.org/10.1126/science.adr2475 \nSugai, L. S. M., Silva, T. S. F., Ribeiro, J. W., Jr, & Llusia, D. (2018). Terrestrial passive \nacoustic monitoring: Review and perspectives. BioScience, 69(1), 15–25. \nhttps://doi.org/10.1093/biosci/biy147 \nToshkova, N., Kolev, M., Deleva, S., Simeonov, T., & Popov, V. (2025). Winter is (not) \ncoming: Acoustic monitoring and temperature variation across important bat hibernacula. \nBiodiversity Data Journal, 13. https://doi.org/10.3897/bdj.13.e141801 \nVerwimp, E., Hacohen, G., & Tuytelaars, T. (2025, February 28). Same accuracy, twice as \nfast: Continuous training surpasses retraining from scratch. arXiv.Org. \nhttps://arxiv.org/abs/2502.21147 \nVoigt, C. C., Kaiser, K., Look, S., Scharnweber, K., & Scholz, C. (2022). Wind turbines \nwithout curtailment produce large numbers of bat fatalities throughout their lifetime: A call \nagainst ignorance and neglect. Global Ecology and Conservation, 37, e02149. \nhttps://doi.org/10.1016/j.gecco.2022.e02149 \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted March 13, 2026. ; https://doi.org/10.64898/2026.03.11.711063doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}