Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
RESEARCH ARTICLE
BatNet: a deep learning-based tool for automated bat
species identification from camera trap images
Gabriella Krivek
1
, Alexander Gillert
2
, Martin Harder
3
,MarcusFritze
1
,KarinaFrankowski
1
,LuisaTimm
1
,
Liska Meyer-Olbersleben
1
,UweFreiherrvonLukas
2,4
, Gerald Kerth
1
&JaapvanSchaik
1
1
Zoological Institute and Museum, Applied Zoology and Nature Conservation, University of Greifswald, Greifswald, Germany
2
Fraunhofer Institute for Computer Graphics Research IGD, Rostock, Germany
3
Forschungsgruppe H€
ohle und Karst Franken e.V. (FHKF), Almoshofer Hauptstraße 51, N€
urnberg, Germany
4
Institute for Visual and Analytic Computing, University of Rostock, Rostock, Germany
Keywords
Automated monitoring, bat conservation,
camera trap, Chiroptera, deep learning,
infrared light barrier
Correspondence
Gabriella Krivek, Zoological Institute and
Museum, Applied Zoology and Nature
Conservation, University of Greifswald,
Loitzer Strasse 26, 17489 Greifswald,
Germany. Tel.: +49 (0)3834 420-4358; Fax:
+49 (0)3834 420-4252; E-mail:
krivek.g@gmail.com
Funding Information
This work was funded by a joint research
project DIG-IT! That is supported by the
European Social Fund (ESF), reference: ESF/
14-BM-A55-0014/19, and the Ministry of
Education, Science and Culture of
Mecklenburg-Vorpommern, Germany.
Editor: Marcus Rowcliffe
Associate Editor: Rahel Sollmann
Received: 12 January 2023; Revised: 21 April
2023; Accepted: 25 April 2023
doi: 10.1002/rse2.339
Abstract
Automated monitoring technologies can increase the efficiency of ecological data
collection and support data-driven conservation. Camera traps coupled with infra-
red light barriers can be used to monitor temperate-zone bat assemblages at under-
ground hibernacula, where thousands of individuals of multiple species can
aggregate in winter. However, the broad-scale adoption of such photo-monitoring
techniques is limited by the time-consuming bottleneck of manual image proces-
sing. Here, we present BatNet, an open-source, deep learning-based tool for auto-
mated identification of 13 European bat species from camera trap images. BatNet
includes a user-friendly graphical interface, where it can be retrained to identify new
bat species or to create site-specific models to improve detection accuracy at new
sites. Model accuracy was evaluated on images from both trained and untrained
sites, and in an ecological context, where community- and species-level metrics
(species diversity, relative abundance, and species-level activity patterns) were com-
pared between human experts and BatNet. At trained sites, model performance was
high across all species (F1-score: 0.98–1). At untrained sites, overall classification
accuracy remained high (96.7–98.2%), when camera placement was comparable to
the training images (<3 m from the entrance; <45°angle relative to the opening).
For atypical camera placements (>3mor>45°angle), retraining the detector model
with 500 site-specific annotations achieved an accuracy of over 95% at all sites. In
the ecological case study, all investigated metrics were nearly identical between
human experts and BatNet. Finally, we exemplify the ability to retrain BatNet to
identify a new bat species, achieving an F1-score of 0.99 while maintaining high
classification accuracy for all original species. BatNet can be implemented directly to
scale up the deployment of camera traps in Europe and enhance bat population
monitoring. Moreover, the pretrained model can serve as a baseline for transfer
learning to automatize the image-based identification of bat species worldwide.
Introduction
Effective conservation depends on the ability to quantify
biodiversity and monitor species-level population dynam-
ics in threatened ecosystems (Primack, 1995). Bats are an
integral part of nearly all terrestrial ecosystems, where
they provide essential ecosystem services and act as eco-
logical indicators of general ecosystem health (Kunz
et al., 2011). Despite their essential ecological role, bat
populations across the globe face multiple threats, such as
the loss and degradation of suitable roosting and foraging
sites, the introduction of new infectious diseases, and
global warming coupled with increasingly unpredictable
climatic conditions (Frick et al., 2020). These effects are
especially problematic for bats, which exhibit slow life
strategies, and thus, their populations may require
decades to recover from individual mortality events
(Fleischer et al., 2017). Therefore, the need for accurate
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and
distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
1
estimates of population trends and a fundamental under-
standing of how these effects are changing bat behavior
and life history has never been more pressing.
One of the primary techniques used for monitoring
temperate-zone bat populations is visually counting bats
at their winter hibernacula. Hibernation sites are attrac-
tive for monitoring as they are used by individuals of
multiple species and by individuals from multiple sum-
mer maternity colonies (Dekeukeleire et al., 2016). How-
ever, as bats are small and many prefer to hibernate in
deep crevices, there can be large discrepancies between
winter hibernation counts and the actual population sizes
(Battersby, 2008), and some species may be entirely
missed by these visual surveys (e.g., Toffoli &
Calvini, 2021).
More accurately monitoring bat activity and population
dynamics at hibernacula is possible with the combination
of infrared light barriers and custom-made camera traps
(Krivek et al., 2023). These camera traps consist of a mir-
rorless digital camera and a white flash, which provides
high image quality (Fig. 1A) and thus allows reliable
species-level identification (Fig. 1B–D). Moreover, bats do
not change their behavior in response to the fast, white
flash of such camera traps (1/5500 s, 1/16 power), making
these photo-monitoring systems suitable as a minimally
invasive method for bat monitoring (Krivek et al., 2022).
These camera traps can either be installed on the inner
side of the light barrier and be triggered by each bat
entering the hibernaculum (i.e., ‘entry’ camera), and/or
on the outer side and be triggered by each bat leaving the
hibernaculum (i.e., ‘exit’ camera). In this study, we
focused on the use of entry cameras to describe ecological
metrics, such as species diversity, relative abundance of
species and species-level activity patterns at hibernation
sites. To obtain these metrics, thus far, the species of the
bat that triggered the camera trap had to be manually
identified, which is a time-consuming task that requires
extensive experience with the subtle morphological differ-
ences between species. Given that a site with around 600
hibernating bats may yield up to 30 000 camera trap
images every year (Krivek et al., 2023), manual image
analysis represents a substantial hurdle for large-scale
monitoring projects. Although deep learning-based species
identification from camera trap images is now common-
place for many terrestrial mammals (e.g., Norouzzadeh
et al., 2018; Tabak et al., 2019) and several automated
species identification tools have been developed for bats
from acoustic recordings (see examples in Rydell
et al., 2017), such resources do not exist for identifying
bat species from camera trap images. While manual vali-
dation of some identifications (e.g., with low confidence
or of rare species) should be performed prior to ecologi-
cal inference, such automated solutions can nevertheless
considerably speed up the identification process. Here, we
present BatNet, an open-source, deep learning-based tool
for automated bat species identification from camera trap
images. This tool was developed to identify 13 bat species
or species-complexes (i.e., similar species within a genus
that cannot be reliably distinguished based on the mor-
phological characteristics visible on the camera trap
images), encompassing all species commonly observed at
hibernacula in Northwestern Europe. BatNet consists of
three main stages: a detector that localizes all bats in an
image, a segmentation network that removes the back-
ground around the detected bats and a classifier that uses
the image crops for species identification. To train the
baseline model, we used an imbalanced training dataset of
16 333 camera trap images of 13 bat species from 32
hibernation sites (range: 375–3576 images; see Table S1
for sample sizes per species). For new locations and spe-
cies, both the detector and the classifier stages can be
retrained from within the user-friendly, coding-free
graphical interface of BatNet. Here, the detector of the
baseline model was retrained to create site-specific models
for six new hibernation sites, and the classifier was
retrained to identify one additional bat species. Model
performance was evaluated in four ways: (1) accuracy on
test images of all 13 species from trained sites (N=2163)
using the baseline model; (2) accuracy on images from
six new, untrained sites (N=49 873) using the baseline
and the site-specific models; (3) in an ecological case
study, where community- and species-level ecological
metrics (species diversity, relative abundance, and species-
level activity patterns) were compared between human
and BatNet using 5-month datasets from three sites
(N=54 748), encompassing the entire hibernation-entry
phase; and (4) accuracy on test images of the original 13
species (N=2163) supplemented by images of a newly
added species (N=1143) using the retrained classifier
model. BatNet is freely available under a CC BY-NC-SA
4.0 license (https://github.com/GabiK-bat/BatNet).
Materials and Methods
Training data and model architecture
In total 18 496 images of bats were collected at the entrance
of 32 hibernacula across Germany using custom-built cam-
era traps (Fig. 1) that were triggered by infrared light bar-
riers (ChiroTEC, Lohra). For each image, two human
experts classified the bat to species level (Barbastella barbas-
tellus,Eptesicus serotinus,Myotis bechsteinii,M. dasycneme,
M. daubentonii,M. emarginatus,M. nattereri and Nyctalus
noctula) or to species complex (the whiskered bats: Myotis
alcathoe,M. brandtii,M. mystacinus; the mouse-eared bats:
Myotis blythii,M. myotis; the long-eared bats: Plecotus
2ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
auritus,P. austriacus; the pipistrelles: Pipistrellus pipistrellus,
Pi. pygmaeus; and the horseshoe bats: Rhinolophus ferrume-
quinum,R. hipposideros). The location and species identity
of each bat in all images were annotated with bounding
boxes using the LabelMe software (Torralba et al., 2010),
and these annotations were used to train an object detector
and classifier networks. Using the same annotation tool, a
random subset of 3685 images were subsequently manually
traced to crop the bat out from the background, which
were used to train a segmentation network. From the total
dataset, 90% (N=16 333) was used to train the detector
and the classifier, and 10% (N=2163) was used for testing
final model performance (see Table S1 for sample sizes per
species). All networks were trained for 30 epochs with a
learning rate of 0.05 and a stochastic gradient descent
optimizer.
BatNet is composed of three distinct stages: detection,
segmentation, and species classification (Figure S1). First,
a Faster-R-CNN object detector (Ren et al., 2015) with a
ResNet50 (He et al., 2016) Feature Pyramid Network (Lin
et al., 2017) places a bounding box around each bat
detected in an image. Localizing classifications within an
image rather than classifying the image as a whole was
preferred, as this approach decreases the noise resulting
from the image background and provides the ability to
count and identify all animals in an image (Schneider
et al., 2020). Second, the image is cropped to the bound-
ing box and a U-Net segmentation network (Ronneberger
et al., 2015) with a MobileNet V3 backbone (Howard
et al., 2019) removes the background. Because deep learn-
ing models have the tendency to learn static background
features (Miao et al., 2019), this segmentation step
Figure 1. (A) The camera trap setup used in this study for bat monitoring at hibernation sites, composed of a mirrorless digital camera and an
external white flash. (B–D) Camera trap images of bats (insets show enhanced image crops of the captured bats) entering three of the
investigated hibernacula in Germany: (B) Batzbach (Plecotus auritus), (C) Comthurey (Myotis nattereri) and (D) Eldena (Myotis daubentonii). The
entrances of these sites were monitored with infrared light barriers, which automatically triggered the camera trap when the bat entered the
hibernaculum.
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 3
G. Krivek et al. Automated Bat Species Identification
ensures that the actual bat characteristics are used for
classification in the next step and not the background fea-
tures that remain within the bounding box. Finally, the
segmented crop of the image is classified by an ensemble
of three MobileNet V3 networks (Howard et al., 2019).
This configuration was selected, because ensemble net-
works are less prone to make highly confident yet incor-
rect predictions than a single neural network (Li &
Hoiem, 2020). Each of the three networks classifies the
original and the flipped version of the image crop, and
the six resulting predictions are then averaged. This tech-
nique, called test-time augmentation, is known to
improve the performance of image classification models
(e.g., Kim et al., 2020). The final output of the classifier
is composed of the predicted identification for each
detected bat and a confidence value between 0 and 1 for
each prediction, which indicates the level of certainty in
the species identification.
Since transfer learning is an established technique to
improve neural network performance and reduce training
time (Yosinski et al., 2014), the object detector was pre-
trained on the COCO (Common Objects in Context)
dataset (Lin et al., 2014), and all other networks were pre-
trained on ImageNet (Russakovsky et al., 2015). In addi-
tion, the training dataset was augmented with random
horizontal flips of the original camera trap images, with
image crops of bats from these modified camera trap
images, and with empty images (i.e., only background
without any bats). Since outlier exposure (i.e., training
with natural images that do not contain the target
objects) is commonly used to improve detection perfor-
mance at untrained background locations (Hendrycks
et al., 2018), random images from the ImageNet dataset
(Russakovsky et al., 2015) were also included in the train-
ing dataset of the baseline model as negative examples
(i.e., images of anything else than a bat).
Using BatNet
Camera trap images can either be processed in a fully
automated way using the command-line interface (‘batch
mode’, no limit to the number of images within the pro-
cessed folder), or in a semi-automated workflow using a
browser-based graphical user interface, which supports
manual validation of the output (‘manual mode’, opti-
mized to process around 1000 locally stored images at a
time). Both approaches result in an output that includes
species labels with confidence levels for each detected bat,
the coordinates of the corresponding bounding boxes,
metadata from the images (e.g., file name, timestamp)
and flags to denote images where the confidence level of
any identifications is below a user-defined threshold, or
where multiple bats were detected in an image or where
no bats were detected (i.e., empty images). Using these
flags, users can quickly sort and filter images that require
manual review, which is further supported by the possi-
bility to zoom in and change the brightness of the images.
Although empty images as a result of false triggers are
uncommon in this photo-monitoring system due to high
light barrier accuracy (Krivek et al., 2023), the camera
traps can be set to trigger at regular intervals (i.e., time
lapse mode) when the light barrier is blocked for an
extended period (e.g., by a spiderweb or a predator sitting
in the hibernaculum entrance), which can result in large
numbers of empty images. Flagging these empty images
drastically reduces the time required for filtering and
allows users to focus on images containing the species of
interest (Beery et al., 2019).
Within the BatNet graphical user interface, both the
object detector and the species classifier can be retrained
on new images in a coding-free environment. In both
cases, new images can either be manually annotated with
bounding boxes and species identifications in the graphi-
cal user interface, or the baseline model output can be
corrected within the user interface (i.e., possibility for
adding, removing, and modifying both bounding boxes
and species labels) and used directly. All training parame-
ters (i.e., species of interest, number of epochs, learning
rate) are adjustable, and the resulting retrained model can
be selected from a drop-down menu within the user
interface. A step-by-step guide for BatNet image proces-
sing and retraining is provided on GitHub (https://
github.com/GabiK-bat/BatNet).
Evaluation on test data
As an initial evaluation, we quantified BatNet perfor-
mance on the 2163 test images that were withheld from
the training dataset but were taken at trained background
locations. To evaluate the performance of the object
detector, we compared the intersection between the pre-
dicted and the manually created bounding boxes around
each labeled bat. We considered predictions as true posi-
tive above 0.4 Intersection over Union (IoU; 0 –no over-
lap, 1 –perfect overlap) and false negative if the overlap
was below the threshold. Predicted bounding boxes with-
out any bats were considered false positive errors. To
evaluate classifier performance, identifications were con-
sidered true positive when the human and predicted clas-
sifications were the same, false negative when the species
of interest was incorrectly classified as a different species,
and false positive when a different species was incorrectly
classified as the species of interest.
The object detection and classifier performance were
quantified by three common accuracy metrics: precision
(i.e., ratio of correctly predicted positive observations to
4ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
the total predicted positive observations; high precision
minimizes false positive errors), recall (i.e., ratio of cor-
rectly predicted positive observations to all observations
in the actual class; high recall minimizes false negative
errors) and F1-score (i.e., weighted average of precision
and recall; used for evaluation when both false negative
and false positive errors are equally undesirable).
Untrained sites and model retraining
Next, we evaluated the baseline model performance on
49 873 images from six untrained sites that were spatially
and temporally independent from the training data (for
example images see Figure S2). Untrained sites were cate-
gorized based on their similarity to the training dataset
and included three typical sites, where the camera angle
and distance from the entrance were similar to the train-
ing images (i.e., camera installed <3 m from the entrance
and at a <45°angle relative to the opening), one with
atypical camera distance (i.e., >3 m from the entrance),
and two sites with atypical camera angle (i.e., >45°angle
relative to the opening). Images from the untrained sites
were classified by one human expert and annotated with
bounding boxes and species labels.
Besides using the baseline object detector model, a total
of 24 site-specific detector models were trained for the six
sites (10 epochs, learning rate 0.001) using 25, 50, 100 or
500 site-specific annotations (i.e., bounding boxes without
species labels). As for the baseline detector model, F1-
scores were calculated for each of these detector models
to evaluate their performance.
Ecological case study
We explored the utility of BatNet for describing
community- and species-level ecological metrics using a
continuous 5-month camera trap dataset comprising the
complete hibernation-entry phase (01 August–01 January)
from one trained (Eldena) and two untrained locations
(Batzbach, Comthurey). In these datasets (N=54 748
images), the human expert only identified the species of
the bat that triggered the camera trap without considering
the bats flying in the background or annotating them with
bounding boxes. This represents the typical manual identi-
fication procedure, where the primary goal is to quantify
the number of bats per species that entered the hibernacu-
lum. In terms of the automated identifications, BatNet pre-
dictions were based on the baseline model for images from
the trained site (Eldena) and from the untrained site with
typical camera angle (Batzbach). For the untrained site
with atypical camera angle (Comthurey), images were iden-
tified using a site-specific model that was trained with 500
site-specific bounding boxes of bats.
In addition to the overall accuracy as described above,
we focused on three ecological metrics: species diversity
(i.e., the list of species detected at a site), relative abun-
dance (i.e., the percentage of identifications attributed to
each species at a site) and species-level activity patterns of
bats throughout their hibernation-entry phase (i.e., the
dates at which the total number of identifications per spe-
cies within a site had reached the 5th, 25th, 50th, 75th
and 95th percentiles). For these applications, different
confidence thresholds can be applied to the output of
BatNet to optimize the balance between high accuracy
(i.e., F1-score) and the proportion of identifications that
are retained in the final output (i.e., above confidence
threshold). Instead of using the test data (i.e., images
withheld from the baseline training), we used the data
from the ecological case study to generate an optimal
confidence threshold for each ecological application,
because these were considered more informative for real-
world applications. To define the optimal thresholds for
each application, we evaluated the proportion of false
positive errors (i.e., errors retained in the final output)
versus the false negative errors and the identifications
below the selected threshold (i.e., identifications not
retained in the final output) across all confidence thresh-
olds (Fig. 2). Based on these results, species diversity at a
hibernaculum was determined using a 95% confidence
threshold, which minimizes the proportion of false posi-
tive errors while still retaining each species, including the
rare ones. To eliminate the small number of remaining
false positives, we manually reviewed all identifications of
species that constitute less than 1% of the total dataset
based on the BatNet output. To estimate the relative
abundance of each species and describe species-specific
activity patterns, we selected a 70% confidence threshold
and discarded all identifications below this threshold. At
this threshold, the proportion of false-positive errors is
strongly reduced, but the proportion of false-negative
errors and identifications that are discarded as below the
threshold has not started to exponentially increase yet
(Fig. 2).
To describe overall accuracy in the ecological case study,
we generated confusion matrices using a 70% confidence
threshold for the BatNet output. As BatNet provides pre-
dictions for all bats detected in an image, including the
ones in the background, some images yielded multiple bat
identifications. Since true species labels were missing for
the bats that were not considered to have triggered the
camera trap by the human evaluator, BatNet predictions
for these images were manually corrected so that only the
bat that triggered the camera trap was retained for the
accuracy assessment (if the associated confidence value
exceeded the 70% confidence threshold). To correct for
human error, if there was a mismatch between the human
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 5
G. Krivek et al. Automated Bat Species Identification
label and the prediction above 70% confidence threshold
(N=243 out of 54 748 images), two additional human
experts manually reviewed the identifications. Based on this
consensus scoring, the original human identification was
either considered correct (i.e., BatNet prediction was incor-
rect, 76.5%), or incorrect (23.5%), and thus, the original
human label was corrected.
For the investigated ecological metrics, all BatNet pre-
dictions above the selected confidence threshold were
considered, including cases where multiple bats per image
met these criteria. To investigate the ability of BatNet to
accurately describe species diversity from a camera trap
dataset, we compared the list of species identified by Bat-
Net using a 95% confidence threshold with the species
that were truly present at the site based on human identi-
fications. For relative abundance, we compared the per-
centage of the dataset assigned to each species by human
identification (i.e., the bat that triggered the image) and
by BatNet using a 70% confidence threshold (i.e., includ-
ing multiple identifications per image when they were
above the threshold). Finally, we compared the activity
patterns of the four most common bat species at the
investigated sites (Myotis nattereri,M. daubentonii,M.
myotis and P. auritus) between a human expert and Bat-
Net. Specifically, we quantified the differences in the
species-level activity patterns between the two datasets by
calculating the dates at which certain percentiles (5, 25,
50, 75 and 95%) of the total number of identifications
had been reached per species and per site. For each per-
centile, differences were quantified as the number of days
between the date of the percentile obtained by the human
expert and BatNet. Additionally, we used Lin’s concor-
dance correlation coefficients (CCC) to quantify the
agreement between the human expert and BatNet regard-
ing the number of identifications per species per night
throughout the hibernation-entry phase.
Classifier retraining: adding new species
We explored the feasibility of adding a new species to the
classifier, while maintaining the classification accuracy for
the original 13 species. The baseline classifier was
retrained with 58 annotations of a new species (Miniop-
terus schreibersii) and 40–50 annotations per species origi-
nally included in the baseline training. The classifier was
retrained for 10 epochs at a learning rate of 0.001. This
comparatively small number of epochs and low learning
rate were selected to lead to smaller weight updates,
which is needed to minimize forgetting of the original
species classes (i.e., catastrophic forgetting). The
Figure 2. Proportion of false positive errors (i.e., errors retained in the final output; blue dashed lines) versus the proportion of false negative
errors and the identifications below threshold (i.e., identifications not retained in the final output; red solid lines) across all confidence thresholds
(0–100%), when using the baseline model of BatNet to process camera trap images of bats collected at three hibernation sites in Germany. For
the visualization, only those species were considered that had at least 100 identifications within a site in the BatNet output, represented by the
blue and red lines. Vertical dashed black lines indicate the confidence thresholds used for describing relative abundance and activity patterns of
species (70%) and species diversity (95%), when using 5-month camera trap datasets that encompass the entire hibernation-entry phase of bats
at these sites.
6ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
performance of the retrained model was evaluated on
1143 test images of Mi. schreibersii, in addition to the
original 2163 test images of the other 13 bat species.
Results
Test dataset evaluation
Out of the 2163 BatNet identifications on test images
from trained background locations, 15 were incorrect (12
misidentifications and 3 missed detections), yielding an
overall classification accuracy of 99.3% (CI 98.9–99.6%).
Precision, recall and F1-score ranged from 0.97 to 1.00
for all 13 bat species (for confusion matrix see Figure S3).
Untrained sites
Object detection performance of the baseline model,
quantified using the F1-score, ranged from 0.95 to 1.00 at
five of six untrained locations. It was noticeably lower at
one site (0.38 in Calw; Fig. 3), where the camera trap was
installed further from the entrance than usual (>3 m).
After retraining the baseline detector using 500 site-
specific annotations (i.e., bounding boxes without species
labels) for each of the six previously untrained sites, the
F1-score of the site-specific object detection model
increased to 0.94 in Calw and to over 0.98 at the other
five previously untrained locations.
Classification accuracy of the baseline model varied
depending on the camera angle and the distance between
the camera and the entrance (Table 1; example camera
trap images: Figure S2, confusion matrices: Figure S4).
Classification accuracy was high (96.7–98.2%) at
untrained locations with typical backgrounds (i.e., similar
camera angle and distance to the training dataset). It was
markedly lower and more variable at sites with atypical
camera placement (17.8, 86.3 and 90.8%; Table 1), pre-
sumably because many bats were not detected or
Figure 3. The object detection performance of BatNet on camera trap images of bats from six untrained hibernacula from Germany using the
baseline model (i.e., no retraining) and using site-specific models after retraining the baseline detection model with a varying number of site-
specific annotations (25, 50, 100 or 500 bounding boxes without species labels). The performance was quantified by the F1-score without using a
confidence threshold. At three of the sites (Batzbach, Gemeinezeche, Silberberg), the camera angle and distance from the entrance were similar
to the training images (i.e., camera installed <3 m from the entrance and at a <45°angle relative to the opening). In Calw, the camera distance
was atypical (i.e., >3 m from the entrance), and in Comthurey and Grube Emma, the camera angle was atypical (i.e., >45°angle relative to the
opening).
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 7
G. Krivek et al. Automated Bat Species Identification
incorrectly segmented. Notably, after retraining the detec-
tor with 500 site-specific annotations for each of the six
previously untrained sites, classification accuracy
improved to over 95% at all sites (Table 1; 95.5–99.9%).
Ecological case study
Species diversity
BatNet detected all species that were identified by human
experts at all three sites (Batzbach, Comthurey, Eldena;
Table 2). Across the three evaluated datasets
(N=54 748), manual review for the species that consti-
tuted less than 1% of the total dataset was required for
62 images (0.1% of total), resulting in the confirmation
of three true positive species (N=60) and the detection
of one false positive species (N=2).
Relative species abundance
To describe relative species abundance, we used a 70%
confidence threshold that maintained high precision and
recall for all species (Fig. 4) and retained over 90% of the
dataset at all sites (Eldena 90.1%, Batzbach 93.7%,
Comthurey 92.1% of images above the confidence thresh-
old). The difference in the relative abundance of all spe-
cies was within 1.1% at all three sites when comparing
BatNet predictions with 70% confidence threshold to
human identifications (Table 3).
Species-specific activity patterns
Species-specific activity patterns of the four investigated
species (M. daubentonii,M. myotis,M. nattereri and P.
auritus) across a 5-month period were nearly identical
between the human and BatNet identifications (see
Fig. 5A for one example per species, all other combina-
tions in Figure S5A). When the activity patterns were
compared based on the percentiles (5, 25, 50, 75 and
95%) obtained from the human and BatNet outputs, the
difference between the two methods was always less than
3 days across all percentiles per species and per site. Only
one exception occurred, when a 6-day discrepancy was
observed between the 95th percentile obtained by human
vs. BatNet (Myotis daubentonii in Comthurey;
Figure S5A).
The overall sample sizes between the human and Bat-
Net datasets differed due to classifications being discarded
below threshold (reduces the BatNet sample size), and the
classification of multiple bats per image where humans
only scored a single bat per image (increases the BatNet
Table 1. BatNet classification accuracy of bat species from camera trap images with 95% confidence interval at six untrained background loca-
tions from Germany, using the baseline model and the site-specific detector models retrained with 500 site-specific annotations (r500).
Site category Site N
images
Accuracy (95% CI) baseline Accuracy (95% CI) r500
Typical Batzbach 39 430 98.2 (98.1–98.3) 98.1 (97.9–98.2)
Gemeinezeche 997 97.6 (96.4–98.5) 99.9 (99.4–100)
Silberberg 1000 96.7 (95.4–97.7) 99.8 (99.3–100)
Atypical angle Comthurey 6472 90.8 (90.1–91.5) 97.3 (96.9–97.7)
Grube Emma 979 86.3 (84–88.4) 97.5 (96.3–98.3)
Atypical distance Calw 995 17.8 (15.5–20.3) 95.5 (94–96.7)
Nindicates the number of images used for evaluation. Hibernation sites were categorized based on their similarity to the training dataset in terms
of the camera angle and distance from the entrance. In a typical monitoring setup, the camera was installed <3 m from the entrance and at a
<45°angle relative to the opening. The setup was considered atypical when the camera was installed more than 3 m away from the entrance or
it was positioned at a >45°angle relative to the opening.
Table 2. Bat species diversity (i.e., species present at the site) based
on BatNet predictions of species identity with 95% confidence thresh-
old and human expert species identifications at three hibernation sites
in Germany.
Site Species N
BatNet
N
human
Batzbach Myotis nattereri 13 304 (43.2%) 19 416
Myotis bechsteinii 11 077 (36%) 11 901
Plecotus sp. 2202 (7.16%) 2191
Myotis daubentonii 1827 (5.94%) 2666
Myotis myotis 1470 (4.78%) 1653
Myotis brandtii 879 (2.86%) 1363
Myotis dasycneme 2 (0.01%) 0
Comthurey Myotis nattereri 2836 (45.7%) 3019
Myotis myotis 2239 (36.1%) 2263
Myotis daubentonii 1024 (16.5%) 1071
Barbastella barbastellus 73 (1.18%) 76
Plecotus sp. 36 (0.58%) 37
Eldena Myotis nattereri 5542 (72.4%) 6403
Myotis daubentonii 1743 (22.8%) 2192
Plecotus sp. 345 (4.51%) 375
Myotis myotis 19 (0.25%) 71
Myotis brandtii 5 (0.07%) 51
The number of identifications (N) and the proportion of all identifica-
tions within the site that it represents (%) are provided for each spe-
cies identified by BatNet. Bold text indicates that the total proportion
of predicted BatNet identifications for that species was below the 1%
threshold, which was used to recommend manual review of these
identifications.
8ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
sample size). Despite these differences, we observed high
concordance between human and BatNet classifications
per species, per day (range: 0.989–0.999; Fig. 5B and
Figure S5B).
New species
After retraining the baseline model with 58 annotations
of Mi. schreibersii,BatNet achieved an F1-score of 0.99
for the new species (Fig. 6). The performance for the
original 13 species remained high (F1-score range: 0.94–
0.99). Overall classification accuracy of the model was
98% (CI 97.3–98.3%). When applying a 70% confidence
threshold, out of the 1413 Mi. schreibersii identifications
196 were below the threshold and only 1 identification
was incorrect (F1-score 1.00; Fig. S6).
Discussion
BatNet is a deep learning-based tool for automated iden-
tification of 13 Northwestern European bat species, that
can be retrained to adjust to new sites and to include
new species within a coding-free environment. On test
images from trained locations, the baseline model
achieved high species-level classification accuracy across
all 13 bat species (F1-score range: 0.98–1.00). This is
likely a result of the localization and segmentation steps
implemented before species classification, which were not
used in other image-based identification studies that
found highly variable model performance for different
species (e.g., V
elez et al., 2023; Whytock et al., 2021).
Overall classification accuracy of the baseline model
remained remarkably high at untrained sites (96.7–
98.2%), where the camera angle and distance from the
entrance were comparable to the training images. At
untrained sites with an atypical camera setup, site-specific
models reached an overall classification accuracy above
95% after retraining with 500 annotations. These results
are particularly important, as classification accuracy mea-
sures derived from trained locations are known to
decrease significantly when applied on images from new
locations (Schneider et al., 2020), despite models being
trained with broad and diverse image datasets. The possi-
bility to retrain the object detector and create site-specific
models with minimal manual annotation effort allows
BatNet to overcome detection difficulties related to new
backgrounds and camera setups. Beyond overall accuracy,
we showed that BatNet yields nearly identical results to
manual identification when used to quantify ecologically
relevant community- and species-level metrics, such as
species diversity, relative abundance, and species-specific
activity patterns. Finally, retraining the baseline model
with an additional, morphologically similar, new bat
species resulted in high classification accuracy, both for
the new species (F1-score: 0.99), and for all other 13 spe-
cies (F1-score: 0.94–0.99). Consequently, BatNet repre-
sents an accurate and highly adaptable platform for
automation of camera trap-based bat monitoring.
Improving the speed and scalability of camera trap-
based monitoring of bats has large implications for bat
conservation, given the improvement this method consti-
tutes over winter hibernation counts for monitoring bat
population dynamics (Krivek et al., 2023). Importantly,
camera traps attached to infrared light barriers can be
used to accurately describe species diversity at a hibernac-
ulum, since they are able to detect all species entering the
site, including those that are often vastly undercounted or
not detected at all during visual surveys (e.g., crevice-
roosting species; Toffoli & Calvini, 2021). Furthermore,
the continuous nature of camera trap-based monitoring
allows us to describe the activity patterns of different spe-
cies. Here this was exemplified using percentiles, where
the 5th and 95th percentiles can serve as a reliable mea-
sure of the start and end of the species-specific activity
during the hibernation-entry phase, and the combination
of the 25th, 50th and 75th percentiles can indicate the
peak activity of different species. These measures can be
used then to compare activity patterns between species,
sites and years in a standardized way. Exploring these
fine-scale changes in bat activity can help describe how
species differ in their hibernation phenology and in terms
of their response to changing weather conditions (cf.
Meier et al., 2022) and contribute to data-driven conser-
vation actions. Finally, the installation of camera traps
with infrared light barriers could be a promising new sur-
vey method to minimize direct contact with bats and
thus, prevent human disturbance and possible introduc-
tion of pathogens to new sites (e.g., WNS, Covid-19; Ble-
hert et al., 2009; Kingston et al., 2021).
Although not investigated here, dual camera trap setups
(i.e., both entry and exit camera) have the potential to
also quantify the absolute abundances of bat species at
hibernacula, which remains difficult for many species
based on traditional monitoring methods (Van der Meij
et al., 2015). By adding up the net number of entries (i.e.,
identifications in the entry camera) and exits (i.e., identi-
fications in the exit camera) per species throughout the
hibernation entry or emergence phases, species-level pop-
ulation sizes could be estimated –an approach similar to
estimating population sizes of mixed species assemblages
using light barrier data (Krivek et al., 2023). For such
applications, BatNet should be implemented in a semi-
automated workflow, where identifications below the con-
fidence threshold of all species are manually reviewed in
the graphical user interface. Additionally, in images with
multiple bats, users must select the identification of the
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 9
G. Krivek et al. Automated Bat Species Identification
10 ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
bat that triggered the camera trap and discard the identi-
fications of bats in the background.
The primary limitation to the implementation of this
photo-monitoring method is that the light barriers that
are used to trigger the camera trap can only monitor
entrance sizes of up to 35 9300 cm. However, in Ger-
many and many other European countries, the entrances
of many large complex mines and caves, where gains in
monitoring resolution are expected to be greatest, have
already been reduced in size to limit human disturbance
and access (Krivek et al., 2023). Thus, although modifica-
tions to the entrance should always be performed with
caution (e.g., Pugh & Altringham, 2005), the method may
be nevertheless widely applicable to monitor temperate
zone bats that predominantly make use of underground
sites as hibernacula. Finally, it should be noted that if
absolute population estimates are not needed, the system
could also be installed to only cover a portion of the total
entrance, in which case ecological metrics could still be
estimated under the assumption that the bats flying
through the monitored area constitute a random sample
of the total assemblage.
Comparison with other automated species
identification approaches
The accuracy of BatNet, both at trained and untrained
sites, is remarkably high in comparison to other deep
learning solutions for automated, image-based mammal
species identification (e.g., Norouzzadeh et al., 2018;
Tabak et al., 2019). In large part, this may be explained
by several key differences between classic wildlife camera
trap setups and the camera traps triggered by infrared
light barriers here used for bat monitoring. First, these
custom-made camera traps are installed at the entrance of
hibernation sites that are nearly exclusively used by bats.
Therefore, only a relatively narrow species range had to
be considered for training the networks. Second, since
these camera traps are triggered by bats flying through an
infrared light barrier, their distance from the camera
when the image is taken remains highly consistent. Thus,
the camera can be manually focused at a fixed depth to
ensure that most bats appear sharp on the images. Third,
the environment is often comparatively simple and artifi-
cial, and the bats are only rarely partially occluded, which
contrasts sharply with the complex, vegetation-rich back-
drop of most camera trap studies. This allows for rela-
tively simple segmentation and isolation of the target
from the background. Finally, the use of white flash with
standardized settings provides a fixed amount of white
light in an otherwise completely dark environment. This
results in a better and more standardized image quality
than afforded by infrared flashes and variable lighting
conditions in most traditional wildlife camera setups. The
resulting high image quality allows identification of differ-
ent bat species with high certainty, even though the
Table 3. Relative bat species abundance per site based on BatNet
predictions of species identity with 70% confidence threshold and
human species identifications at three hibernation sites in Germany.
Site Species BatNet % Human %
Batzbach
N
images
=39 190
Myotis nattereri 48.40 49.50
Myotis bechsteinii 31.40 30.40
Myotis daubentonii 6.67 6.80
Plecotus sp. 5.86 5.59
Myotis myotis 4.32 4.22
Myotis brandtii 3.28 3.48
Myotis dasycneme 0.03 0.00
Barbastella barbastellus <0.01 0.00
Myotis emarginatus <0.01 0.00
Pipistrellus sp. <0.01 0.00
Rhinolophus sp. <0.01 0.00
Comthurey
N
images
=6466
Myotis nattereri 47.30 46.70
Myotis myotis 34.70 35.00
Myotis daubentonii 16.20 16.60
Barbastella barbastellus 1.24 1.18
Plecotus sp. 0.55 0.57
Eldena
N
images
=9092
Myotis nattereri 70.90 70.40
Myotis daubentonii 24.30 24.10
Plecotus sp. 3.99 4.12
Myotis myotis 0.64 0.78
Myotis brandtii 0.16 0.56
Myotis emarginatus 0.01 0.00
Pipistrellus sp. 0.01 0.00
The number of images evaluated (N) is indicated for each site.
Figure 4. Confusion matrix of human species identifications and BatNet predictions of species identity with 70% confidence threshold for
camera trap images of bats collected at three hibernation sites in Germany. The confusion matrix shows the distribution of classification error
within a species, where the diagonal represents the number of accurate classifications and all other cells in the matrix describe the number of
errors (i.e., missed detections or misclassifications). The color of the cells reflects the number of classifications within each category, with dark
purple cells indicating high numbers and light blue cells indicating low numbers. Identifications below the confidence threshold (70%) were
summarized according to their true species label (‘below threshold’). For images when multiple bats were detected by BatNet, but humans only
identified the bat that triggered the camera trap, additional BatNet identifications were summarized according to their predicted species label
(‘multiple bats’). Precision refers to the ratio of correctly predicted positive observations to the total predicted positive observations. Recall
indicates the ratio of correctly predicted positive observations to all observations in the actual class. F1-score is the weighted average of precision
and recall.
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 11
G. Krivek et al. Automated Bat Species Identification
Figure 5. (A) Activity patterns of four bat species (Myotis daubentonii,M. myotis,M. nattereri and Plecotus auritus) throughout the hibernation-
entry phase (01 August–01 January), based on species identifications from camera trap images by human experts (orange) and BatNet predictions
with 70% confidence threshold (blue). Camera trap images of bats were collected at three hibernation sites in Germany (Batzbach, Comthurey,
Eldena). To quantify the differences between the activity patterns obtained by human experts vs. BatNet, percentiles were used across the 5-
month datasets (5% and 95% indicated with vertical dashed gray lines, 25% and 75% indicated with vertical solid gray lines, and 50% indicated
with vertical solid black lines). The sample size (N) indicates the total number of identifications across the season. (B) Concordance plots indicate
the agreement between the number of human and BatNet identifications per bat species per night, quantified by the Lin’s CCC (range: 0–1).
These coefficients indicate how far the observed data deviate from the line of perfect concordance (black solid line).
12 ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
morphological differences between bat species are far
more subtle than between many other mammals. How-
ever, the image quality also depends on the camera trap
angle and distance from the entrance, as demonstrated
here for several sites with atypical camera placements
(Calw, Comthurey, Grube Emma). Therefore, for optimal
performance of BatNet, camera traps should be placed 2–
3 m away from the entrance and ideally in a 45°angle or
less to ensure the best possible image conditions for reli-
able species identification.
The performance of BatNet was further improved by
implementing techniques that have not been commonly
used in other automated, image-based species identifica-
tion pipelines (e.g., Norouzzadeh et al., 2018; Tabak
et al., 2019). First, deep learning models can learn the
background features of specific camera trap stations
instead of the focal animals (Miao et al., 2019), which
introduces bias. To ensure that the classifier focuses on
the characteristics of bats instead of the common back-
ground features, we trained a U-Net segmentation
Figure 6. Confusion matrix of human species identifications and BatNet predictions of species identity after retraining the baseline model to be
able to identify a new European bat species, Miniopterus schreibersii, in addition to the 13 bat species included in the original training data. The
confusion matrix shows the distribution of classification error within a species, where the diagonal represents the number of accurate
classifications and all other cells in the matrix describe the number of errors (i.e., missed detections or misclassifications). The color of the cells
reflects the number of classifications within each category, with dark purple cells indicating high numbers and light blue cells indicating low
numbers. Precision refers to the ratio of correctly predicted positive observations to the total predicted positive observations. Recall indicates the
ratio of correctly predicted positive observations to all observations in the actual class. F1-score is the weighted average of precision and recall.
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 13
G. Krivek et al. Automated Bat Species Identification
network to automatize background removal. While such
approaches may be more difficult to implement for data-
sets with more complex backgrounds, they may neverthe-
less be worthwhile. Second, single neural networks are
more prone to make highly confident yet incorrect pre-
dictions (Li & Hoiem, 2020). Here, we used an ensemble
of three neural networks for classification, where each
network classified the original and the flipped version of
the image (i.e., test-time augmentation). This resulted in
more informative confidence levels that could be used for
discarding low-confidence identifications or filtering them
out for manual review. Exploring the adoption of these
techniques in other deep learning-based species identifica-
tion approaches may similarly improve their performance.
For bats, deep learning-based species identification of
passive acoustic recordings have become increasingly pop-
ular, with several automated classifiers of echolocation
calls being developed (Mac Aodha et al., 2018; Rydell
et al., 2017; Tabak et al., 2022). Such approaches can sim-
ilarly be applied to characterize bat assemblages at under-
ground sites. However, several taxa, most notably the
genus Myotis, remain difficult to identify automatically
(Rydell et al., 2017) due to the high variability in call fea-
tures within species. This is exacerbated when multiple
individuals of several species are calling simultaneously
(Bergmann et al., 2022). Moreover, these issues similarly
affect the manual validation of acoustic recordings,
whereas validation from camera trap images is readily fea-
sible. Despite these shortcomings, acoustic surveys repre-
sent an important method for wide-spread surveillance
and scouting and remain one of the only methods for
monitoring hibernation sites where light barriers cannot
be readily installed and that cannot be visually counted
(e.g., complex sites with many large entrances, rock crev-
ices and piles; Blomberg et al., 2021).
Application in bat monitoring and
conservation
Automated monitoring of hibernacula combined with the
implementation of BatNet has the potential to improve bat
population monitoring worldwide. In Northwestern
Europe, the ability to retrain BatNet for new locations
allows it to be directly applied to vastly scale up camera
trap-based bat monitoring while maintaining high accu-
racy. In other regions, the pretrained model of BatNet can
be used as a baseline for transfer learning to automatize
identification of a broad range of bat species, beyond our
target species list. In adjacent regions this may only require
minor modification of the species list, to add species such
as illustrated here for Mi. schreibersii. In other areas, using
the pretrained model as a baseline is expected to produce
more accurate and stable results with less computational
expense than pretraining on conventional image datasets,
because of the general features the baseline model learned
from a diverse, yet bat-specific camera trap dataset.
Prior to ecological inference for new datasets, model
performance should always be carefully evaluated by a
human using a subsample of manually identified images
to detect any new or hidden biases (Norouzzadeh
et al., 2021; Schneider et al., 2020). In this context, Bat-
Net and its graphical user interface improve the efficiency
of camera trap analysis in several ways. First, it allows
sorting and filtering images based on their flags (i.e.,
empty, with multiple bats, below confidence threshold).
Second, it is possible to manually review the final output
within a user-friendly graphical interface (i.e., add,
remove, or modify bounding boxes around bats and their
species labels). Finally, the user interface also supports the
coding-free retraining of the baseline detector model for
new sites and of the classifier model for new bat species.
Overall, these aspects can help ecologists establish more
efficient workflows for processing large camera trap data-
sets (V
elez et al., 2023). Given the numerous stressors
affecting global bat populations (Frick et al., 2020) and
the legal obligation to monitor bat populations world-
wide, a greater flow of monitoring data is essential to
support data-driven wildlife management and conserva-
tion decisions. BatNet drastically improves our ability to
achieve these objectives.
Acknowledgements
We would like to thank the entire ChiroTEC team for
providing an identified set of camera trap images and
Karl Kugelschafter for his valuable insights, to Alexander
Seliger and Marvin Marzenberger for helping with the
training data preparation, to Jonas Denck for advice
regarding the development of BatNet, and to Thomas Lil-
ley and two anonymous reviewers for their helpful com-
ments on a previous version of this manuscript. This
work was funded by a joint research project DIG-IT! That
is supported by the European Social Fund (ESF), refer-
ence: ESF/14-BM-A55-0014/19, and the Ministry of Edu-
cation, Science and Culture of Mecklenburg-
Vorpommern, Germany. G. Kr. is an associate member of
the DFG Research training Group ‘Biological Responses
to Novel and Changing Environments’; RTG 2010. Open
Access funding enabled and organized by Projekt DEAL.
Data Availability Statement
BatNet is freely available under a CC BY-NC-SA 4.0
license at https://github.com/GabiK-bat/BatNet, along
with data and scripts used for evaluation, under a CC
BY-NC-ND 4.0 license.
14 ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
References
Battersby, J. (2008) Surveillance and monitoring methods for
European bats. Guidelines Produced by the Agreement on
the Conservation of Populations of European Bats
(EUROBATS). p. 85.
Beery, S., Morris, D., Yang, S., Simon, M., Norouzzadeh, A. &
Joshi, N. (2019) Efficient pipeline for automating species ID
in new camera trap projects. Biodiversity Information Science
and Standards,3, e37222.
Bergmann, A., Burchardt, L.S., Wimmer, B., Kugelschafter, K.,
Gloza-Rausch, F. & Kn€
ornschild, M. (2022) The soundscape
of swarming: proof of concept for a noninvasive acoustic
species identification of swarming Myotis bats. Ecology and
Evolution,12(11), e9439.
Blehert, D.S., Hicks, A.C., Behr, M., Meteyer, C.U., Berlowski-Zier,
B.M., Buckles, E.L. et al. (2009) Bat white-nose syndrome: an
emerging fungal pathogen? Science,323(5911), 227.
Blomberg, A.S., Vasko, V., Meierhofer, M.B., Johnson, J.S.,
Eeva, T. & Lilley, T.M. (2021) Winter activity of boreal bats.
Mammalian Biology,101, 609–618.
Dekeukeleire, D., Janssen, R., Haarsma, A.-J., Bosch, T. & Van
Schaik, J. (2016) Swarming behaviour, catchment area and
seasonal movement patterns of the Bechstein’s bats:
implications for conservation. Acta Chiropterologica,18(2),
349–358.
Fleischer, T., Gampe, J., Scheuerlein, A. & Kerth, G. (2017) Rare
catastrophic events drive population dynamics in a bat species
with negligible senescence. Scientific Reports,7(1), 1–9.
Frick, W.F., Kingston, T. & Flanders, J. (2020) A review of the
major threats and challenges to global bat conservation.
Annals of the New York Academy of Sciences,1469(1), 5–25.
He, K., Zhang, X., Ren, S. & Sun, J. (2016) Deep residual
learning for image recognition. Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pp.
770–778.
Hendrycks, D., Mazeika, M. & Dietterich, T. (2018) Deep
anomaly detection with outlier exposure. arXiv [Preprint]
Arxiv:1812.04606.
Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B., Tan,
M. et al. (2019) Searching for MobileNetV3. Proceedings of
the IEEE International Conference on Computer Vision, pp.
1314–1324.
Kim, I., Kim, Y. & Kim, S. (2020) Learning loss for test-time
augmentation. Advances in Neural Information Processing
Systems,33, 4163–4174.
Kingston, T., Frick, W., Kading, R., Leopardi, S., Medellin, R.,
Mendenhall, I.H. et al. (2021) IUCN SSC Bat Specialist
Group (BSG) recommended strategy for researchers to
reduce the risk of transmission of SARS-CoV-2 from
humans to bats. Version 2.0, AMP: Assess, Modify, Protect.
Krivek, G., Mahecha, E.P.N., Meier, F., Kerth, G. & van
Schaik, J. (2023) Counting in the dark: estimating
population size and trends of bat assemblages at hibernacula
using infrared light barriers. Animal Conservation. Available
from: https://doi.org/10.1111/acv.12856
Krivek, G., Schulze, B., Poloskei, P.Z., Frankowski, K.,
Mathgen, X., Douwes, A. et al. (2022) Camera traps with
white flash are a minimally invasive method for long-term
bat monitoring. Remote Sensing in Ecology and Conservation,
8(3), 284–296.
Kunz, T.H., Braun de Torrez, E., Bauer, D., Lobova, T. &
Fleming, T.H. (2011) Ecosystem services provided by bats.
Annals of the New York Academy of Sciences,1223(1), 1–38.
Li, Z. & Hoiem, D. (2020) Improving confidence estimates for
unfamiliar examples. IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pp. 2686–2695.
Lin, T.-Y., Doll
ar, P., Girshick, R., He, K., Hariharan, B. &
Belongie, S. (2017) Feature pyramid networks for object
detection. IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pp. 2117–2125.
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,
Ramanan, D. et al. (2014) Microsoft COCO: common
objects in context. European Conference on Computer Vision,
8693, 740–755.
Mac Aodha, O., Gibb, R., Barlow, K.E., Browning, E., Firman,
M., Freeman, R. et al. (2018) Bat detective—deep learning
tools for bat acoustic signal detection. PLoS Computational
Biology,14(3), e1005995.
Meier, F., Grosche, L., Reusch, C., Runkel, V., van Schaik, J. &
Kerth, G. (2022) Long-term individualized monitoring of
sympatric bat species reveals distinct species-and
demographic differences in hibernation phenology. BMC
Ecology and Evolution,22(1), 1–12.
Miao, Z., Gaynor, K.M., Wang, J., Liu, Z., Muellerklein, O.,
Norouzzadeh, M.S. et al. (2019) Insights and approaches using
deep learning to classify wildlife. Scientific Reports,9(1), 1–9.
Norouzzadeh, M.S., Morris, D., Beery, S., Joshi, N., Jojic, N. &
Clune, J. (2021) A deep active learning system for species
identification and counting in camera trap images. Methods
in Ecology and Evolution,12(1), 150–161.
Norouzzadeh, M.S., Nguyen, A., Kosmala, M., Swanson, A.,
Palmer, M.S., Packer, C. et al. (2018) Automatically
identifying, counting, and describing wild animals in
camera-trap images with deep learning. Proceedings of the
National Academy of Sciences of the United States of America,
115(25), E5716–E5725.
Primack, R.B. (1995) Essentials of conservation biology, Vol. 23.
Sunderland: Sinauer Associates.
Pugh, M. & Altringham, J.D. (2005) The effect of gates on
cave entry by swarming bats. Acta Chiropterologica,7(2),
293–299.
Ren, S., He, K., Girshick, R. & Sun, J. (2015) Faster R-CNN:
towards real-time object detection with region proposal
networks. Advances in Neural Information Processing Systems,
28,1–9.
Ronneberger, O., Fischer, P. & Brox, T. (2015) U-Net:
convolutional networks for biomedical image segmentation.
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 15
G. Krivek et al. Automated Bat Species Identification
International Conference on Medical Image Computing and
Computer-Assisted Intervention,9351, 234–241.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S.
et al. (2015) ImageNet large scale visual recognition challenge.
International Journal of Computer Vision,115(3), 211–252.
Rydell, J., Nyman, S., Ekl€
of, J., Jones, G. & Russo, D. (2017)
Testing the performances of automated identification of bat
echolocation calls: a request for prudence. Ecological
Indicators,78, 416–420.
Schneider, S., Greenberg, S., Taylor, G.W. & Kremer, S.C.
(2020) Three critical factors affecting automated image
species recognition performance for camera traps. Ecology
and Evolution,10(7), 3503–3517.
Tabak, M.A., Murray, K.L., Reed, A.M., Lombardi, J.A. & Bay,
K.J. (2022) Automated classification of bat echolocation call
recordings with artificial intelligence. Ecological Informatics,
68, 101526.
Tabak, M.A., Norouzzadeh, M.S., Wolfson, D.W., Sweeney,
S.J., VerCauteren, K.C., Snow, N.P. et al. (2019) Machine
learning to classify animal species in camera trap images:
applications in ecology. Methods in Ecology and Evolution,
10(4), 585–590.
Toffoli, R. & Calvini, M. (2021) Long term trends of
hibernating bats in North-Western Italy. Biologia,76(2),
633–643.
Torralba, A., Russell, B.C. & Yuen, J. (2010) LabelMe: online
image annotation and applications. Proceedings of the IEEE,
98(8), 1467–1484.
Van der Meij, T., Van Strien, A., Haysom, K., Dekker, J., Russ, J.,
Biala, K. et al. (2015) Return of the bats? A prototype indicator
of trends in European bat populations in underground
hibernacula. Mammalian Biology,80(3), 170–177.
V
elez, J., McShea, W., Shamon, H., Castiblanco-Camacho, P.J.,
Tabak, M.A., Chalmers, C. et al. (2023) An evaluation of
platforms for processing camera-trap data using artificial
intelligence. Methods in Ecology and Evolution,14(2), 459–477.
Whytock, R.C.,
Swie_
zewski, J., Zwerts, J.A., Bara-Słupski, T.,
Koumba Pambo, A.F., Rogala, M. et al. (2021) Robust ecological
analysis of camera trap data labelled by a machine learning
model. Methods in Ecology and Evolution,12(6), 1080–1092.
Yosinski, J., Clune, J., Bengio, Y. & Lipson, H. (2014) How
transferable are features in deep neural networks? Advances
in Neural Information Processing Systems,27,1–9.
Supporting Information
Additional supporting information may be found online
in the Supporting Information section at the end of the
article.
Table S1. Number of camera trap images per bat species
used for training BatNet and testing the baseline model
performance.
Figure S1. Schematic overview of BatNet, a deep learning-
based tool that automatically identifies bat species from
camera trap images in three steps: bat detection (object
detector), background removal (segmentation network) and
species classification (ensemble of classifiers). The final out-
put includes a species prediction with a confidence level.
Optionally, low-confidence predictions can be manually
reviewed in the graphical user interface by human experts.
Figure S2. Example camera trap images from six untrained
locations that were categorized based on their similarity to
the training dataset, including three typical hibernation
sites (camera installed <3 m from the entrance and at a
<45°angle relative to the opening; A –Batzbach, B –
Gemeinezeche, C –Silberberg), two sites with atypical cam-
era angle (>45°angle relative to the opening; D –
Comthurey, E –Grube Emma) and one with atypical cam-
era distance (>3 m from the entrance; F –Calw).
Figure S3. Confusion matrix of human identifications
and BatNet predictions (without confidence threshold)
for test images from trained background locations. The
confusion matrix shows the distribution of classification
error within a species, where the diagonal represents the
number of accurate classifications and all other cells in
the matrix describe the number of errors (i.e., missed
detections or misclassifications). The color of the cells
reflects the number of classifications within each category,
with dark purple cells indicating high numbers and light
blue cells indicating low numbers. Precision refers to the
ratio of correctly predicted positive observations to the
total predicted positive observations. Recall indicates the
ratio of correctly predicted positive observations to all
observations in the actual class. F1-score is the weighted
average of precision and recall.
Figure S4. Confusion matrix of human identifications
and BatNet predictions (without confidence threshold)
for camera trap images from six untrained background
locations using the baseline model and the site-specific
models retrained with 500 local annotations (r500). The
confusion matrix shows the distribution of classification
error within a species, where the diagonal represents the
number of accurate classifications and all other cells in
the matrix describe the number of errors (i.e., missed
detections or misclassifications). The color of the cells
reflects the number of classifications within each category,
with dark purple cells indicating high numbers and light
blue cells indicating low numbers. Precision refers to the
ratio of correctly predicted positive observations to the
total predicted positive observations. Recall indicates the
ratio of correctly predicted positive observations to all
observations in the actual class. F1-score is the weighted
average of precision and recall.
Figure S5. (A) Activity patterns of Myotis daubentonii,
Myotis myotis,Myotis nattereri and Plecotus auritus
16 ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Automated Bat Species Identification G. Krivek et al.
throughout the hibernation-entry phase (01 August–01
January), based on species identifications from camera
trap images by human experts (orange) and BatNet pre-
dictions with 70% confidence threshold (blue). Camera
trap images of bats were collected at three hibernation
sites in Germany (Batzbach, Comthurey, Eldena). To
quantify the differences between the activity patterns
obtained by humans versus BatNet, percentiles were used
across the 5-month datasets (5 and 95% indicated with
vertical dashed gray lines, 25 and 75% indicated with ver-
tical solid gray lines, and 50% indicated with vertical solid
black lines). The sample size (N) indicates the total num-
ber of identifications across the season. (B) Concordance
plots indicate the agreement between the number of
human and BatNet identifications per bat species per
night, quantified by the Lin’s concordance correlation
coefficient (CCC, range: 0–1). These coefficients indicate
how far the observed data deviate from the line of perfect
concordance (black solid line).
Figure S6. Confusion matrix of human identifications
and BatNet predictions with 70% confidence threshold
after retraining the baseline model to be able to identify
a new European bat species, Miniopterus schreibersii,in
addition to the 13 bat species included in the original
training data. The confusion matrix shows the distribu-
tion of classification error within a species, where the
diagonal represents the number of accurate classifications
and all other cells in the matrix describe the number of
errors (i.e., missed detections or misclassifications). The
color of the cells reflects the number of classifications
within each category, with dark purple cells indicating
high numbers and light blue cells indicating low num-
bers. Precision refers to the ratio of correctly predicted
positive observations to the total predicted positive
observations. Recall indicates the ratio of correctly pre-
dicted positive observations to all observations in the
actual class. F1-score is the weighted average of precision
and recall.
ª2023 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 17
G. Krivek et al. Automated Bat Species Identification
