-
https://cpw.cvlcollections.org/files/original/b96a60b885d924667cba67671c0bbad0.pdf
63ec950ded61f580d841f6c6eda0d32b
PDF Text
Text
The research in this publication was partially or fully funded by Colorado Parks and Wildlife.
Dan Prenzlow, Director, Colorado Parks and Wildlife • Parks and Wildlife Commission: Marvin McDaniel, Chair • Carrie Besnette Hauser, Vice-Chair
Marie Haskett, Secretary • Taishya Adams • Betsy Blecha • Charles Garcia • Dallas May • Duke Phillips, IV • Luke B. Schafer • James Jay Tutchton • Eden Vardy
�Received: 7 September 2018
|
Accepted: 4 October 2018
DOI: 10.1111/2041-210X.13120
A P P L I C AT I O N
Machine learning to classify animal species in camera trap
images: Applications in ecology
Michael A. Tabak1,2
| Mohammad S. Norouzzadeh3 | David W. Wolfson1 |
Steven J. Sweeney1 | Kurt C. Vercauteren4 | Nathan P. Snow4
| Joseph M. Halseth4 |
Paul A. Di Salvo1 | Jesse S. Lewis5 | Michael D. White6 | Ben Teton6 |
James C. Beasley7 | Peter E. Schlichting7 | Raoul K. Boughton8 | Bethany Wight8 |
Eric S. Newkirk9 | Jacob S. Ivan9 | Eric A. Odell9 | Ryan K. Brook10 |
Paul M. Lukacs11 | Anna K. Moeller11 | Elizabeth G. Mandeville2,12 | Jeff Clune3 |
Ryan S. Miller1
1
Center for Epidemiology and Animal Health, United States Department of Agriculture, Fort Collins, Colorado; 2Department of Zoology and
Physiology, University of Wyoming, Laramie, Wyoming; 3Computer Science Department, University of Wyoming, Laramie, Wyoming; 4National Wildlife
Research Center, United States Department of Agriculture, Fort Collins, Colorado; 5College of Integrative Sciences and Arts, Arizona State University, Mesa,
Arizona; 6Tejon Ranch Conservancy, Lebec, California; 7Savannah River Ecology Laboratory, Warnell School of Forestry and Natural Resources, University
of Georgia, Aiken, South Carolina; 8Range Cattle Research and Education Center, Wildlife Ecology and Conservation, University of Florida, Ona, Florida;
9
Colorado Parks and Wildlife, Fort Collins, Colorado; 10Department of Animal and Poultry Science, University of Saskatchewan, Saskatoon, SK, Canada;
11
Wildlife Biology Program, Department of Ecosystem and Conservation Sciences, W.A. Franke College of Forestry and Conservation, University of Montana,
Missoula, Montana and 12Department of Botany, University of Wyoming, Laramie, Wyoming
Correspondence
Michael A. Tabak
Email: tabakma@gmail.com
and
Ryan S. Miller
Email: ryan.s.miller@aphis.usda.gov
Funding information
U.S. Department of Energy, Grant/Award
Number: DE-EM0004391; USDA Animal
and Plant Health Inspection Service,
National Wildlife Research Center and
Center for Epidemiology and Animal Health;
Colorado Parks and Wildlife; Canadian
Natural Science and Engineering Research
Council; University of Saskatchewan; Idaho
Department of Game and Fish
Handling Editor: Theoni Photopoulou
Abstract
1. Motion-activated cameras (“camera traps”) are increasingly used in ecological and
management studies for remotely observing wildlife and are amongst the most
powerful tools for wildlife research. However, studies involving camera traps result in millions of images that need to be analysed, typically by visually observing
each image, in order to extract data that can be used in ecological analyses.
2. We trained machine learning models using convolutional neural networks with the
ResNet-18 architecture and 3,367,383 images to automatically classify wildlife species
from camera trap images obtained from five states across the United States. We tested
our model on an independent subset of images not seen during training from the
United States and on an out-of-sample (or “out-of-distribution” in the machine learning
literature) dataset of ungulate images from Canada. We also tested the ability of our
model to distinguish empty images from those with animals in another out-of-sample
dataset from Tanzania, containing a faunal community that was novel to the model.
3. The trained model classified approximately 2,000 images per minute on a laptop
computer with 16 gigabytes of RAM. The trained model achieved 98% accuracy at
identifying species in the United States, the highest accuracy of such a model to
date. Out-of-sample validation from Canada achieved 82% accuracy and correctly
identified 94% of images containing an animal in the dataset from Tanzania. We
provide an
Methods Ecol Evol. 2019;10:585–590.
r
package (Machine Learning for Wildlife Image Classification) that
wileyonlinelibrary.com/journal/mee3�
© 2018 The Authors. Methods in Ecology and
Evolution © 2018 British Ecological Society
|
585
�586
|
Methods in Ecology and Evolu on
TABAK et al.
allows the users to (a) use the trained model presented here and (b) train their own
model using classified images of wildlife from their studies.
4. The use of machine learning to rapidly and accurately classify wildlife in camera
trap images can facilitate non-invasive sampling designs in ecological studies by
reducing the burden of manually analysing images. Our
r
package makes these
methods accessible to ecologists.
KEYWORDS
artificial intelligence, camera trap, convolutional neural network, deep neural networks, image
classification, machine learning, r package, remote sensing
1 | I NTRO D U C TI O N
by a third observer (Appendix S1). If any part of an animal (e.g., leg
Camera traps are increasingly used to remotely observe wildlife
as an image of the species. If an image did not contain any animals, it
over large geographical areas with minimal human involvement and
was classified as empty. The images from Canada were not used for
or ear) was identified as being present in an image, this was included
have made considerable contributions to ecology (Howe, Buckland,
training, but were used as an out-of-sample dataset for validation.
Després-Einspenner, & Kühl, 2017; O’Connell, Nichols, & Karanth,
This resulted in a total of 3,741,656 classified images that included
2011; Rovero, Zimmermann, Bersi, & Meek, 2013). A common lim-
27 species or groups (see Table 1) across the study locations. We
itation is these methods lead to a large accumulation of images
present these images and their classifications for other scientists
which must be first classified in order to be used in ecological stud-
to use for model development as the North American Camera Trap
ies (Niedballa, Sollmann, Courtiol, & Wilting, 2016; Swanson et al.,
Images (NACTI) dataset. To increase processing speed, images were
2015). The burden of manually viewing and classifying images often
resized to 256 × 256 pixels following the methods and using the
constrains studies by reducing the sampling intensity (e.g., number
Python script of Norouzzadeh et al. (2018). To have a more robust
of cameras deployed), limiting the geographical extent and duration
model, we randomly applied different label-preserving transfor-
of studies. Recently, machine learning has emerged as a potential
mations (cropping, horizontal flipping, and brightness and contrast
solution for automatically classifying images from camera traps
modifications), called data augmentation (Krizhevsky, Sutskever, &
(Chen, Han, He, Kays, & Forrester, 2014; Gomez Villa, Salazar, &
Hinton, 2012).
Vargas, 2017; Norouzzadeh et al., 2018; Swinnen, Reijniers, Breno,
& Leirs, 2014; Yu et al., 2013).
We randomly selected 90% of the classified images for each species or group to train the model and 10% of the images to test it.
We sought to develop a machine learning approach that can be
However, we wanted to evaluate the model’s performance for each
applied across study sites and provide software that ecologists can
species present at each study site, so we used conditional sampling
use for identification of wildlife in their own camera trap images.
in which we altered training–testing allocation for the rare situations
Using over three million identified images of wildlife from camera
(four total instances) where there were few classified images of a
traps from five locations across the United States, we trained and
species at a site. Specifically, with 1–9 classified images for a species
tested deep learning models that automatically classify wildlife.
at a site (two instances), we used all of these images for testing and
package (Machine Learning for Wildlife Image
none for training (the model was trained using only images of these
Classification [MLWIC]) that allows researchers to classify camera
species from other sites); for site-species pairs with 10–30 images
trap images from North America or train their own machine learning
(two instances), 50% were used for training and testing; and for >30
models to classify images.
images per site for each species, 90% were allocated to training and
We provide an
r
10% to testing (Appendices S3–S7 show the number of training and
2 | M ATE R I A L S A N D M E TH O DS
2.1 | Camera trap images
Species in camera trap images from five locations across the United
test images for each species at each site). This resulted in 3,367,383
images used to train the model and 374,273 images used for testing.
2.2 | Machine learning process
States (California, Colorado, Florida, South Carolina and Texas) and
As machine learning methods are new to many ecologists, we pro-
one location from Canada (Saskatchewan) were identified manually
vide a brief introduction in a supplement (Appendix S2). Following
by researchers (see Appendix S1 for a description of each field loca-
Norouzzadeh et al., we trained a deep convolutional neural net-
tion). Images were either classified by a single wildlife expert or eval-
work (ResNet-18) architecture (He, Zhang, Ren, & Sun, 2016)
uated independently by two researchers; any conflicts were decided
using the TensorFlow framework (Adabi et al., 2016) using Mount
�Didelphis virginiana
Equus spp.
Homo sapiens
Leporidae
Lynx rufus
Mephitis mephitis
Rodentia
Odocoileus hemionus
Odocoileus virginianus
Procyon lotor
Opossum
Horse
Human
Rabbits
Bobcat
Striped skunk
Rodent
Mule deer
White-t ailed deer
Raccoon
414,119
3,367,365
Empty
Total
23,413
61,063
Aves
Vehicle
79,628
10,749
287,017
59,072
13,272
42,948
87,900
87,700
3,279
10,331
22,889
17,768
88,667
2,517
1,804
3,919
8,926
4,037
1,991
185,390
20,851
2,039
1,817,109
8,967
Number of training
images
Bird
Ursus americanus
Meleagris gallopavo
Turkey
Black bear
Dasypus novemcinctus
Armadillo
Vulpes vulpes and Urocyon
Cinereoargenteus
Corvidae
Corvid
Sus scrofa
Mustelidae
Mustelidae
Fox
Cervus canadensis
Elk
Wild pig
Canidae
Canidae
Puma concolor
Callipepla californica
Quail
Sciurus spp.
Bos taurus
Cattle
Mountain lion
Alces alces
Moose
Squirrel
Scientific name
Model performance for each species or group
Species or group
name
TA B L E 1
364,660
46,016
6,787
2,602
8,850
1,204
31,893
6,566
1,484
4,781
1,360
8,543
366
1,154
2,554
1,977
9,854
281
210
447
993
452
223
20,606
2,321
236
201,903
997
Number of test
images
0.98
0.96
0.94
0.93
0.95
0.91
0.98
0.97
0.92
0.90
0.94
0.98
0.79
0.95
0.91
0.95
0.96
0.94
0.79
0.90
0.89
0.84
0.77
0.99
0.89
0.91
0.99
0.98
Recall
1.00
1.00
1.00
1.00
1.00
0.99
1.00
1.00
0.98
1.00
1.00
1.00
0.98
0.98
0.99
1.00
1.00
0.99
0.96
1.00
0.99
1.00
0.99
1.00
0.99
0.96
1.00
1.00
Top-5 recall
0.98
0.94
0.95
0.95
0.98
0.94
0.98
0.95
0.97
0.89
0.95
0.98
0.88
0.96
0.94
0.96
0.97
0.94
0.88
0.90
0.93
0.80
0.87
0.99
0.93
0.93
0.99
0.98
Precision
0.06
0.05
0.05
0.02
0.06
0.02
0.05
0.03
0.11
0.05
0.02
0.12
0.04
0.06
0.04
0.03
0.06
0.12
0.10
0.07
0.20
0.13
0.01
0.07
0.07
0.01
0.02
False-positive rate
0.04
0.06
0.07
0.05
0.09
0.02
0.03
0.08
0.10
0.06
0.02
0.21
0.05
0.09
0.05
0.04
0.06
0.21
0.10
0.11
0.16
0.23
0.01
0.11
0.09
0.01
0.02
False-negative rate
TABAK et al.
Methods in Ecology and Evolu on
|
587
�|
588
Methods in Ecology and Evolu on
TABAK et al.
Moran, a high performance computing cluster (Advanced Research
on images with a completely different species community (from
Computing Center, 2012). We used the ReLU activation function,
Tanzania) to determine the model’s ability to correctly classify im-
55 epochs, a backpropagation algorithm of Stochastic Gradient
ages as having an animal or being empty when encountering new
Descent with Momentum (Goodfellow, Bengio, & Courville, 2016),
species that it has not been trained to recognize. This was done
and the learning rate (η) and weight decay varied by epoch number
using 3.2 million classified images from the Snapshot Serengeti
as described in Appendix S8.
dataset (Swanson et al., 2015).
In Appendix S2, we describe the calculation of metrics including
accuracy, recall, precision and false-positive and false-negative error
rates. Briefly, recall and precision are measures of the model’s per-
3 | R E S U LT S
formance at correctly identifying each species. We fit generalized
additive models (GAMs) to the relationship between recall and the
Our model performed well, achieving 97.6% accuracy of identi-
logarithm (base 10) of the number of images used to train the model;
fying the correct species with the top guess. The top-5 accuracy
see Appendix S9 for a description of this model. We also calculated
was >99.9%. Figure 1 provides examples of image classification by
the recall and rates of error specific to each of the five datasets from
the model. The model confidence in the correct answer varied,
which images were acquired.
but was mostly >95%; see Figure 2 for confidences for each image
for three example species. In Appendix S10, we present a confu-
2.3 | Model validation
sion matrix comparing the classifications by the model with those
from manual classification. Supporting a similar finding for camera
To evaluate how the model would perform for a completely new
trap images in Norouzzadeh et al. (2018), and a general trend in
study site in North America, we used a dataset of 5,900 classi-
deep learning (Goodfellow et al., 2016), species and groups that
fied images of ungulates (moose, cattle, elk and wild pigs) from
had more images available for training were classified more ac-
Saskatchewan, Canada, by running the trained model on these
curately (Figure 3, Table 1). GAMs relating the number of training
images. We also evaluated the ability of the model to operate
images with recall predicted 95% recall could be achieved when
(a)
Correct classification by model
Model Guess
Confidence (%)
Wild pig
96.11
Cattle
2.38
Empty
1.49
White-tailed deer <0.1
Moose
<0.1
Answer from human classifiers: Wild pig
(b) Incorrect
classification by model
Model Guess
Wild pig
Cattle
Moose
Black bear
Bobcat
Confidence (%)
48.82
31.27
16.93
2.51
0.51
Answer from human classifiers: Cattle
F I G U R E 1 Examples of images that could be difficult to classify. The model correctly identifies a wild pig (a) by seeing only its
hindquarters and tail (right side of image). The model incorrectly classifies a cattle as a wild pig (b), as only an ear is visible in the image; note
that the model has relatively low confidence in the top guess for this image. Nevertheless, cattle are within the top-5 guesses for this image,
so while it is incorrect, it counts towards the top-5 recall for cattle
�Methods in Ecology and Evolu on
TABAK et al.
20,000
while 94.3% of images containing an animal were classified as containing an animal. Our trained model was capable of classifying approximately 2,000 images per minute on a Macintosh laptop with 16
0
Frequency
gigabytes of RAM.
(b) White−tailed deer
200
4 | D I S CU S S I O N
0
150
589
dataset, we found that 85.1% were classified correctly as empty,
(a) Wild pig
10,000
400
|
(c) Corvidae
To our knowledge, our model achieved the highest accuracy (97.6%)
100
to date in using machine learning to classify wildlife in camera trap
50
images (a recent paper achieved 95% accuracy; Norouzzadeh et al.,
0
0.0
0.2
0.4
0.6
0.8
1.0
Model confidence in this species or group when it is in the image
F I G U R E 2 Histograms represent the confidence assigned
by all of the top-5 guesses by the model for each of these three
example species when it was present in an image. The dashed
line represents 95% confidence; the majority of model-assigned
confidences were greater than this value
2018). This model performed almost as well during the night as during the day (accuracy = 97% and 98%, respectively). We provide
this model as an
r
package (MLWIC), which is especially useful for
researchers studying the species and groups available in this package (Table 1) in North America, as it performed well (82% accuracy)
in classifying ungulates in an out-of-sample test of images from
Canada. The model can also be valuable for researchers studying
other species by removing images without any animals from the
dataset before beginning manual classification, as we achieved high
accuracy in separating empty images from those containing animals
1.00
in a dataset from Tanzania. This r package can also be a valuable tool
Recall
0.95
for any researchers that have classified images, as they can use the
package to train their own model that can then classify any subse-
0.90
quent images collected.
The ability to rapidly identify millions of images from camera
0.85
traps can fundamentally change the way ecologists design and im0.80
plement wildlife studies. The burden of classifying images from camera traps has led ecologists to limit the duration and size of camera
0.75
10
3
10
4
5
10
10
6
Number of training images
F I G U R E 3 Model recall (the ability of the model to recognize
species) increased with the size of the training dataset for that
species. Points represent each species or group of species. The line
represents the result of generalized additive models relating the
two variables (see Appendix S9 for details)
trap studies (Kelly et al., 2008; Scott et al., 2018). By removing this
burden, camera traps can be applied in more studies including monitoring invasive or sensitive species, long-term ecological research
and small-scale occupancy studies.
AC K N OW L E D G E M E N T S
We thank the hundreds of volunteers and employees who manually
classified images and deployed camera traps. We thank Dan Walsh
approximately 54,000 training images were available for a spe-
for facilitating cooperation amongst groups. Camera trap projects
cies or group. However, for several species and groups, 95% recall
were funded by the U.S. Department of Energy under award # DE-
was achieved with fewer than 50,000 images (Figure 3). We found
EM0004391 to the University of Georgia Research Foundation;
there was not a large effect of daytime versus night-t ime on accu-
USDA Animal and Plant Health Inspection Service, National
racy in the model as daytime accuracy was 98.2% and night-t ime
Wildlife Research Center and Center for Epidemiology and Animal
accuracy was 96.6%. The top-5 accuracies for both times of day
Health; Colorado Parks and Wildlife; Canadian Natural Science and
were ≥99.9%. When we subsetted the testing dataset by study
Engineering Research Council; University of Saskatchewan; and
site, we found that site-specific accuracies ranged from 90% to
Idaho Department of Game and Fish.
99% (Appendices S3–S7).
When we conducted out-of-sample validation by using our
model to evaluate images of ungulates from Canada, we achieved
AU T H O R S ’ C O N T R I B U T I O N S
an overall accuracy of 81.8% with a top-5 accuracy of 90.9%. When
M.A.T., R.S.M., K.C.V., N.P.S., S.J.S. and D.W.W. conceived of the
we tested the ability of our model to accurately predict the presence
project; D.W.W., J.S.L., M.A.T., R.K.B., B.W., P.A.D., J.C.B., M.D.W.,
or absence of an animal in the image using the Serengeti Snapshot
B.T., P.E.S., N.P.S., K.C.V., J.M.H., E.S.N., J.S.I., E.A.O., R.K.B., P.M.L.
�590
|
Methods in Ecology and Evolu on
TABAK et al.
and A.K.M. oversaw collection and manual classification of wildlife
in camera trap images from the study sites; M.S.N. and J.C. developed and programmed the machine learning models; M.A.T. led the
analyses and writing of the r package; E.G.M. assisted with r package
development and computing; M.A.T. and R.S.M. led the writing. All
authors contributed critically to drafts and gave final approval for
submission.
DATA ACCESSIBILITY
The trained model is available in the
GitHub
r
package MLWIC from
(https://github.com/mikeyEcology/MLWIC;
https://doi.
org/10.5281/zenodo.1445736). We provide the >3.7 million classified images as the North American Camera Trap Images (NACTI)
dataset in the Labeled Information Library of Alexandria: Biology &
Conservation (LILA:BC) digital repository (available online at http://
lila.science/datasets/nacti).
ORCID
Michael A. Tabak
Nathan P. Snow
http://orcid.org/0000-0002-2986-7885
http://orcid.org/0000-0002-5171-6493
REFERENCES
Adabi, M., Barhab, P., Chen, J., Chen, Z., Davis, A., Dean, J., … Zheng, X.
(2016). TensorFlow: a system for large-scale machine learning (Vol.
16, pp. 265–283). Presented at the 12th USENIX Symposium
on Operating Systems Design and Implementation, USENIX
Association.
Advanced Research Computing Center. (2012). Mount Moran: IBM system X cluster. Laramie, WY: University of Wyoming. Retrieved from
https://arcc.uwyo.edu/guides/mount-moran
Chen, G., Han, T. X., He, Z., Kays, R., & Forrester, T. (2014). Deep convolutional neural network based species recognition for wild animal
monitoring (pp. 858–862). IEEE International Conference on Image
Processing (ICIP). https://doi.org/10.1109/icip.2014.7025172
Gomez Villa, A., Salazar, A., & Vargas, F. (2017). Towards automatic wild animal monitoring: Identification of animal species in
camera-t rap images using very deep convolutional neural networks. Ecological Informatics, 41, 24–32. https://doi.org/10.1016/j.
ecoinf.2017.07.004
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning (1st ed.).
Cambridge, MA: MIT Press.
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). IEEE. https://doi.
org/10.1109/cvpr.2016.90
Howe, E. J., Buckland, S. T., Després-Einspenner, M.-L., & Kühl,
H. S. (2017). Distance sampling with camera traps. Methods
in Ecology and Evolution, 8(11), 1558–1565. https://doi.
org/10.1111/2041-210X.12790
Kelly, M. J., Noss, A. J., Di Bitetti, M. S., Maffei, L., Arispe, R. L.,
Paviolo, A., … Di Blanco, Y. E. (2008). Estimating puma densities
from camera trapping across three study sites: Bolivia, Argentina,
and Belize. Journal of Mammalogy, 89(2), 408–418. https://doi.
org/10.1644/06-MAMM-A-424R.1
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural
information processing systems (pp. 1097–1105).Retrieved from https://
papers.nips.cc/book/advances-in-neural-information-processingsystems-25-2012.
Niedballa, J., Sollmann, R., Courtiol, A., & Wilting, A. (2016). camtrapR: An R package for efficient camera trap data management.
Methods in Ecology and Evolution, 7(12), 1457–1462. https://doi.
org/10.1111/2041-210X.12600
Norouzzadeh, M. S., Nguyen, A., Kosmala, M., Swanson, A., Palmer, M.
S., Packer, C., & Clune, J. (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. https://doi.org/10.1073/
pnas.1719367115
O’Connell, A. F., Nichols, J. D., & Karanth, K. U. (Eds.) (2011). Camera traps
in animal ecology: Methods and analyses. Tokyo, Japan; New York, NY:
Springer.
Rovero, F., Zimmermann, F., Bersi, D., & Meek, P. (2013). “Which camera
trap type and how many do I need?” A review of camera features and
study designs for a range of wildlife research applications. Hystrix,
the Italian Journal of Mammalogy, 24(2), 1–9.
Scott, A. B., Phalen, D., Hernandez-Jover, M., Singh, M., Groves, P., & Toribio,
J.-A. L. M. L. (2018). Wildlife presence and interactions with chickens on
Australian commercial chicken farms assessed by camera traps. Avian
Diseases, 62(1), 65–72. https://doi.org/10.1637/11761-101917-Reg. 1
Swanson, A., Kosmala, M., Lintott, C., Simpson, R., Smith, A., & Packer, C.
(2015). Snapshot Serengeti, high-frequency annotated camera trap
images of 40 mammalian species in an African savanna. Scientific
Data, 2, 150026. https://doi.org/10.1038/sdata.2015.26
Swinnen, K. R. R., Reijniers, J., Breno, M., & Leirs, H. (2014). A novel
method to reduce time investment when processing videos from camera trap studies. PLoS ONE, 9(6), e98881. https://doi.org/10.1371/
journal.pone.0098881
Yu, X., Wang, J., Kays, R., Jansen, P. A., Wang, T., & Huang, T. (2013).
Automated identification of animal species in camera trap images.
EURASIP Journal on Image and Video Processing, 2013(1), 52. https://
doi.org/10.1186/1687-5281-2013-52
S U P P O R T I N G I N FO R M AT I O N
Additional supporting information may be found online in the
Supporting Information section at the end of the article.
How to cite this article: Tabak MA, Norouzzadeh MS, Wolfson
DW, et al. Machine learning to classify animal species in
camera trap images: Applications in ecology. Methods Ecol Evol.
2019;10:585–590.
https://doi.org/10.1111/2041-210X.13120
�
Dublin Core
The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/.
Title
A name given to the resource
Journal Articles
Description
An account of the resource
CPW peer-reviewed journal publications
Text
A resource consisting primarily of words for reading. Examples include books, letters, dissertations, poems, newspapers, articles, archives of mailing lists. Note that facsimiles or images of texts are still of the genre Text.
Dublin Core
The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/.
Title
A name given to the resource
Machine learning to classify animal species in camera trap images: applications in ecology
Description
An account of the resource
<ol>
<li>Motion-activated cameras (“camera traps”) are increasingly used in ecological and management studies for remotely observing wildlife and are amongst the most powerful tools for wildlife research. However, studies involving camera traps result in millions of images that need to be analysed, typically by visually observing each image, in order to extract data that can be used in ecological analyses.</li>
<li>We trained machine learning models using convolutional neural networks with the ResNet-18 architecture and 3,367,383 images to automatically classify wildlife species from camera trap images obtained from five states across the United States. We tested our model on an independent subset of images not seen during training from the United States and on an out-of-sample (or “out-of-distribution” in the machine learning literature) dataset of ungulate images from Canada. We also tested the ability of our model to distinguish empty images from those with animals in another out-of-sample dataset from Tanzania, containing a faunal community that was novel to the model.</li>
<li>The trained model classified approximately 2,000 images per minute on a laptop computer with 16 gigabytes of RAM. The trained model achieved 98% accuracy at identifying species in the United States, the highest accuracy of such a model to date. Out-of-sample validation from Canada achieved 82% accuracy and correctly identified 94% of images containing an animal in the dataset from Tanzania. We provide an<span> </span><span class="smallCaps">r</span><span> </span>package (Machine Learning for Wildlife Image Classification) that allows the users to (a) use the trained model presented here and (b) train their own model using classified images of wildlife from their studies.</li>
<li>The use of machine learning to rapidly and accurately classify wildlife in camera trap images can facilitate non-invasive sampling designs in ecological studies by reducing the burden of manually analysing images. Our<span> </span><span class="smallCaps">r</span><span> </span>package makes these methods accessible to ecologists.</li>
</ol>
Bibliographic Citation
A bibliographic reference for the resource. Recommended practice is to include sufficient bibliographic detail to identify the resource as unambiguously as possible.
Tabak, M. A., M. S. Norouzzadeh, D. W. Wolfson, S. J. Sweeney, K. C. Vercauteren, N. P. Snow, J. M. Halseth, P. A. Di Salvo, J. S. Lewis, M. D. White, B. Teton, J. C. Beasley, P. E. Schlichting, R. K. Boughton, B. Wight, E. S. Newkirk, J. S. Ivan, E. A. Odell, R. K. Brook, P. M. Lukacs, A. K. Moeller, E. G. Mandeville, J. Clune, and R. S. Miller. 2018. Machine learning to classify animal species in camera trap images: applications in ecology. Methods in Ecology and Evolution 10:585–590. <a class="epub-doi" href="https://doi.org/10.1111/2041-210X.13120" target="_blank" rel="noreferrer noopener">https://doi.org/10.1111/2041-210X.13120</a>
Creator
An entity primarily responsible for making the resource
Tabak, Michael A.
Norouzzadeh, Mohammad S.
Wolfson, David W.
Sweeney, Steven J.
VerCauteren, Kurt C.
Snow, Nathan P.
Halseth, Joseph M.
Di Salvo, Paul A.
Lewis, Jesse S.
White, Michael D.
Teton, Ben
James C. Beasley,
Schlichting, Peter E.
Boughton, Raoul K.
Wight, Bethany
Newkirk, Eric S.
Ivan, Jacob S.
Odell, Eric A.
Brook, Ryan K.
Lukacs, Paul M.
Moeller, Anna K.
Mandeville, Elizabeth G.
Clune, Jeff
Miller, Ryan S.
Subject
The topic of the resource
Artificial intelligence
Camera trap
Convolutional neural network
Deep neural networks
Image classification
Machine learning
r package
Remote sensing
Extent
The size or duration of the resource.
6 pages
Date Created
Date of creation of the resource.
2018-11-26
Rights
Information about rights held in and over the resource
<a href="http://rightsstatements.org/vocab/InC-NC/1.0/" target="_blank" rel="noreferrer noopener">In Copyright - Non-Commercial Use Permitted</a>
Format
The file format, physical medium, or dimensions of the resource
application/pdf
Language
A language of the resource
English
Is Part Of
A related resource in which the described resource is physically or logically included.
Methods in Ecology and Evolution
Type
The nature or genre of the resource
Article