CoChat
Sign Up
Cited by 110Wildlife Ecology and ConservationAdvanced Neural Network ApplicationsSmart Agriculture and AI

A comparison of deep learning and citizen science techniques for counting wildlife in aerial survey images

Open access
Torney, Colin J., Lloyd‐Jones, David J., Chevallier, Mark, Moyer, David C., Maliti, Honori T., Mwita, Machoke, Kohi, Edward M., Hopcraft, Grant C.·2019·Methods in Ecology and Evolution
DOI 10.1111/2041-210x.13165
Open this paper in CoChat — free accountOpen PDF

Abstract

Fast and accurate estimates of wildlife abundance are an essential component of efforts to conserve ecosystems in the face of rapid environmental change. A widely used method for estimating species abundance involves flying aerial transects, taking photographs, counting animals within the images and then inferring total population size based on a statistical estimate of species density in the region. The intermediate task of manually counting the aerial images is highly labour intensive and is often the limiting step in making a population estimate. Here, we assess the use of two novel approaches to perform this task by deploying both citizen scientists and deep learning to count aerial images of the 2015 survey of wildebeest ( Connochaetes taurinus ) in Serengeti National Park, Tanzania. Through the use of the online platform Zooniverse, we collected multiple non‐expert counts by citizen scientists and used three different aggregation methods to obtain a single count for the survey images. We also counted the images by developing a bespoke deep learning method via the use of a convolutional neural network. The results of both approaches were then compared. After filtering of the citizen science counts, both approaches provided highly accurate total estimates. The deep learning method was far faster and appears to be a more reliable and predictable approach; however, we note that citizen science volunteers played an important role when creating training data for the algorithm. Notably, our results show that accurate, species‐specific, automated counting of aerial wildlife images is now possible.

AI snapshot

Written from the paper’s full text.

TL;DR

This study compared citizen-science counts (via Zooniverse) and a bespoke deep-learning object detector (YOLOv3) against a single expert count of 1,000 aerial survey images of wildebeest from the 2015 Serengeti census. Both approaches produced highly accurate total estimates after filtering, but the deep-learning method was far faster and showed no systematic bias, while citizen scientists systematically undercounted.

Key findings

  • The YOLO algorithm showed no systematic bias: it recorded 20,631 wildebeest versus the expert count of 20,489 (0.69% overcount), with a mean absolute error of 1.70 and RMS error of 5.94 per image across 1,000 images.
  • Raw citizen-science aggregates systematically undercounted: mean of 15 counts gave −11.44% (MAE 2.8, RMS 8.11) and median gave −8.85% (MAE 1.83, RMS 5.85).
  • Filtering out the 5 lowest counts per tile and averaging the remaining 10 produced a highly accurate count (0.55% overcount, MAE 1.58, RMS 4.05), comparable to YOLO.
  • Speed: counting 1,000 test images took 2 hours on an NVIDIA Quadro GP100 GPU, versus the manual process that typically takes 3–4 skilled counters approximately 3–6 weeks; training took 34 hours.
  • Citizen scientists remained important for creating the training data used by the algorithm, so the authors view the two approaches as complementary.

Methods

  • Aerial survey: 2015 Serengeti wildebeest count flown 30 April and 2 May, 10.3 hours of photographic sampling along east–west transects covering 2,040 km at ~700 ft (213 m) altitude and ~100 knots; 1,584 georeferenced images at 7360×4912 pixels.
  • Citizen science: images split into 12 tiles and uploaded to the Zooniverse "Serengeti Wildebeest Count" project; 9,870 images were counted by 2,212 volunteers between 10 and 31 May 2017, with each image counted by 15 volunteers. Three aggregation methods were compared (mean, median, filtered mean removing the 5 lowest counts).
  • Deep learning: YOLOv3 implemented in Keras/TensorFlow, trained on 500 randomly selected images tiled into 864×864 sub-images (20,000 training images), with modifications including reducing anchor boxes from 9 to 3, removing multiscale boxes, and increasing the no-object loss weight from 0.5 to 2; evaluated on 1,000 randomly selected non-training images against a single expert count (DJL) treated as the gold standard.

Limitations

  • Stated: the expert count is taken as the gold standard but "could in principle deviate from the unknown true count."
  • Stated: the optimal filtering of citizen-science data (removing the 5 lowest counts) was determined on the same images used to assess the method, so there was no division into train and test images; the authors also note there is no guarantee the filtering approach is transferable and it may be affected by guidelines, image resolution/size, or the volunteer pool.
  • Stated: whether the citizen-science undercounting bias is persistent or predictable "can only be revealed by repeated citizen science counts of the survey and comparison to expert counts"; a rigorous validation process would be needed before citizen-science counts could be the sole basis for a population estimate.
  • Evident from the text: the deep-learning algorithm depended on crowd-sourced data to build its training set, and the authors note the greatest current challenge for bespoke applications is obtaining sufficiently large training datasets.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1111/2041-210x.13165

Cites

56 of 56 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 20 cited sources checked · 0 supported · 0 flagged

  • Findingown data
    § Abstract · p. 2

    After filtering of the citizen science counts, both the citizen science and deep learning approaches provided highly accurate total estimates of wildebeest numbers in the 2015 Serengeti aerial survey images.

    “After filtering of the citizen science counts, both approaches provided highly accurate total estimates.”
    Quote in the paperSubstantial claim
  • Findingown data
    § Abstract · p. 2

    The deep learning method was far faster than the citizen science approach for counting the aerial survey images.

    “learning method was far faster and appears to be a more reliable and predictable approach”
    Quote in the paperSubstantial claim
  • Findingown data
    § 2.2. Zooniverse image counts · p. 8

    Volunteers would struggle to count entire aerial images due to their large size and high resolution.

    “Following initial trials on the website in August 2015 it was determined that volunteers would struggle to count entire aerial images due to their large size and high resolution.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 3. Results · p. 10

    All methods result in low per image error rate.

    “From these results we see that all methods result in low per image error rate.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The single expert count is the true number of wildebeest in each image.
  • Findingown data
    § 3. Results · p. 11

    A more accurate estimate is obtained by taking the mean of the 10 highest counts.

    “Instead a more accurate 217 estimate is obtained by taking the mean of the 10 highest counts.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The single expert count is the true number of wildebeest in each image.
  • Classified as probably not a substantial claim. 28
  • May combine two claims. 21
  • The quote only partly states this claim. 1

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given
    § Abstract · p. 2

    Citizen science volunteers played an important role when creating training data for the deep learning algorithm.

    “citizen science volunteers played an important role when creating training data for the algorithm.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 1. Introduction · p. 4

    Wildebeest population estimates in the Serengeti have been infrequent and have had large errors.

    “Infrequent counts with large errors highlight the need for new approaches.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given
    § 2.3. Implementation of object detection algorithm · p. 9

    In this application, objects are only present at a single scale.

    “YOLOv3 is a multiscale detector whereas for our application objects are only present at a single scale”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3. Results · p. 10

    Zooniverse volunteers showed a systematic tendency to undercount the images, with approximately 11% and 9% undercount for the total dataset for the mean and median respectively.

    “The zooniverse volunteers showed a systematic tendency to undercount the images, hence there was approximately an 11% and 9% undercount for the total dataset for both the mean and the median.”
    • Also stated in: § 4. Discussion · p. 11 · plain
      “For the citizen science counts we observe a systematic bias in the errors the counters made.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The single expert count is the true number of wildebeest in each image.

  • May combine two claims.

    Findingnone given
    § 4. Discussion · p. 12

    If the minimum or maximum count per image were simply taken, the total count could have been either half or almost double the true count.

    “Note, that if we simply took the minimum or maximum count per image the total count could 245 have been either half or almost double the true count.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The single expert count is the true number of wildebeest in each image.

  • May combine two claims.

    Findingreasoning
    § 4. Discussion · p. 12

    The 1000 images can be processed in under 2 hours, meaning every future census could be counted within 24 hours.

    “The 1000 images can be processed in under 2 hours, meaning every future census could be counted within 24 hours.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § Abstract · p. 2

    The deep learning method appears to be a more reliable and predictable approach than the citizen science approach for counting the aerial survey images.

    “learning method was far faster and appears to be a more reliable and predictable approach”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 1. Introduction · p. 3

    The challenge of detecting and responding to changes in animal abundance is especially acute in the case of migratory species.

    “migratory species (Harris et al., 2009; Singh and Milner-Gulland, 2011).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Harris et al., 2009

      Not checked in the free run.

    • cites Singh and Milner-Gulland, 2011

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Premisenone given
    § 1. Introduction · p. 5

    Estimating population sizes of migratory species is a vital but logistically challenging task.

    “species is a vital, but logistically challenging task.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 1. Introduction · p. 4

    Since the 1970s, wildebeest population estimates in the Serengeti have been made using the Jolly II method with data collected using vertical aerial photography and manual image counts.

    “Since the 1970s estimates have been made using the Jolly II method (Jolly, 1969) with data collected using vertical aerial photography and manual image counts.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Jolly, 1969

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 1. Introduction · p. 5

    Localized environmental disturbances in large geographic areas are often hard to detect.

    “Localized environmental disturbances in large geographic areas are 58 often hard to detect, while the fact that migrations span national and regional borders, means implementing protection 59 strategies typically involves substantial time to coordinate (Lovejoy et al., 1987; Wilcove and Wikelski, 2008).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Lovejoy et al., 1987

      Not checked in the free run.

    • cites Wilcove and Wikelski, 2008

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 2.4. Expert count · p. 9

    The study's results compare the two novel counting methods against a count by a single experienced expert, which could deviate from the unknown true count.

    “Hence our results are a comparison between the two novel methods employed and a count by a single experienced expert, which could in principle deviate from the unknown true count.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The expert count is taken to be the gold standard for comparison, despite potential bias.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims. The quote only partly states this claim.

    Interpretationreasoning
    § 1.2. Automated computer vision · p. 7

    The image resolution was lower in 2009 (4288x2848) and therefore a greater error should be expected when comparing results from Torney et al., 2016.

    “It should be noted that the image resolution was lower in 2009 (4288x2848) and therefore a greater error should be expected.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 1. Introduction · p. 5

    Because migrations span national and regional borders, implementing protection strategies typically involves substantial time to coordinate.

    “Localized environmental disturbances in large geographic areas are 58 often hard to detect, while the fact that migrations span national and regional borders, means implementing protection 59 strategies typically involves substantial time to coordinate (Lovejoy et al., 1987; Wilcove and Wikelski, 2008).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Lovejoy et al., 1987

      Not checked in the free run.

    • cites Wilcove and Wikelski, 2008

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given
    § 2.2. Zooniverse image counts · p. 8

    Count data from early interface versions or testing were removed prior to analysis.

    “Prior to analysis, any count data made using early versions of the counting interface or collected by either developers or citizen scientists during testing of the interface functionality was removed.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given
    § 2.3. Implementation of object detection algorithm · p. 8

    The bounding boxes for the 20,000 training images were manually checked and corrected after initial filtering.

    “After this initial filter, the bounding boxes were manually checked and corrected for each of the 20,000 training images”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1.2. Automated computer vision · p. 6

    Computer vision and machine learning are increasingly becoming essential components of the ecologist's toolbox and have been applied previously to counting aerial images of animals, including the Serengeti wildebeest population.

    “Computer vision and machine learning are increasingly becoming essential components of the ecologist's toolbox (Bruijning et al., 2018; Christin et al., 2018; Dell et al., 2014; Mac Aodha et al., 2018; Valletta et al., 2017; Weinstein, 2018) and have been applied previously to the task of counting aerial images of animals (Bajzak and Piatt, 1990; Chabot et al., 2018; Laliberte and Ripple, 2003; Mc”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Bruijning et al., 2018

      Not checked in the free run.

    • cites Torney et al., 2016

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data
    § 2.4. Expert count · p. 9

    The expert count was conducted by counting each image twice: first as a running tally during marking and second as a recount of the marks within the image.

    “The number of animals in each image were counted twice - first as running tally during marking and secondly as a recount of the marks within the image.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1. Introduction · p. 3

    Measuring population abundance enables the early detection of population declines or increases and is a precursor for adaptive management and conservation strategies.

    “Measuring population abundance enables the early detection of population declines (caused by disease, 50 over-harvesting, or changing patterns of land-use), or population increases and expansions; therefore it is a precursor 51 for adaptive management and conservation strategies (Walters, 1986).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Walters, 1986

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisereasoning
    § 1.2. Automated computer vision · p. 6

    Another potential approach to replacing dedicated professional counters is to use machine learning algorithms.

    “Another potential approach to replacing dedicated professional counters is to use machine learning algorithms.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given
    § 2.4. Expert count · p. 9

    There is potential for bias in the expert count.

    “While there remains the potential for bias in the expert count”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 4. Discussion · p. 11

    Both citizen science and deep learning methods are capable of producing highly accurate image counts.

    “From our results we see that both citizen science and deep learning methods are capable of producing highly 223 accurate image counts.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The single expert count is the true number of wildebeest in each image.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § 4. Discussion · p. 11

    Given some guidance on identifying wildebeest, false positives should be minimal.

    “This is in line with prior expectations; given some guidance on identifying wildebeest (as was available on the zooniverse project page), false positives should be minimal.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning
    § 4. Discussion · p. 11

    Eliminating false negatives requires substantial focus, and concentration will likely wane over time or volunteers will become distracted.

    “However eliminating false negatives requires substantial focus, and it's likely that concentration will wane over time, or volunteers will become distracted.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 3. Results · p. 11

    Both the mean and the median display systematic bias, therefore it is not simply due to the mean being a less robust estimator but instead reveals a tendency for all zooniverse counters, on average, to undercount.

    “It should be noted that both the mean and the median 218 display this systematic bias, therefore it is not simply due to the mean being a less robust estimator (Galton, 1907) but 219 instead reveals a tendency for all zooniverse counters, on average, to undercount.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Galton, 1907

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationown data
    § 4. Discussion · p. 12

    Given the range of counts provided by volunteers and the large errors observed in the baseline metrics, there will need to be a rigorous process of validation before a citizen science count could be used as the sole basis for a population estimate.

    “however given the range of counts provided by the volunteers and 242 the large errors we observe in the baseline metrics ( 11% and 9% undercount for the mean and median respectively) 243 there will need to be a rigorous process of validation before a citizen science count could be used as the sole basis for 244 a population estimate.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 4. Discussion · p. 12

    Image classification using pretrained DCNNs with state-of-the-art architectures can be achieved with a few lines of code using open source libraries such as Keras, while object detection algorithms are increasingly being integrated into libraries such as TensorFlow.

    “Image classification using pretrained DCNNs with state-of-the-art architectures (He et al., 2016; Simonyan and Zisserman, 2014) can be achieved with a few lines of code using open source libraries such as Keras (Chollet et al., 2015), while object detection algorithms are increasingly being integrated into libraries such as TensorFlow (Abadi et al., 2016).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites He et al., 2016

      Not checked in the free run.

    • cites Simonyan and Zisserman, 2014

      Not checked in the free run.

    • cites Chollet et al., 2015

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisenone given
    § 4. Discussion · p. 12

    YOLO has the advantage of being a single-pass object detector that is fast and accurate.

    “YOLO has the advantage of being a single-pass object detector that is fast and accurate.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4. Discussion · p. 12

    Other more sophisticated approaches to processing citizen science data have been proposed.

    “Other more sophisticated approaches to processing citizen science 241 data have been proposed (Swanson et al., 2016)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Swanson et al., 2016

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 4. Discussion · p. 12

    DCNNs are able to process wildlife images for classification tasks and also detect and localise animals.

    “As other authors have shown, DCNNs are able to process wildlife images for classification tasks (Chen et al., 2014; Norouzzadeh et al., 2018; Villa et al., 2017) and also detect and localise animals (Maire et al., 2015; Schneider et al., 2018).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Chen et al., 2014

      Not checked in the free run.

    • cites Norouzzadeh et al., 2018

      Not checked in the free run.

    • cites Villa et al., 2017

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationnone given
    § 4. Discussion · p. 12

    Suppression of these types of outliers is fundamentally part of the ethos of citizen science, but this illustrates the broad range of responses from volunteers.

    “While suppression of these types of outliers is fundamentally 246 part of the ethos of citizen science, this does illustrate the broad range of responses from volunteers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § 4. Discussion · p. 12

    In this regard citizen scientists have a clear role to play.

    “In this regard citizen scientists have a clear role to play.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Obtaining sufficiently large training datasets is the greatest challenge for implementing these algorithms for bespoke applications.

  • May combine two claims.

    Interpretationown data
    § 4. Discussion · p. 12

    Deep learning algorithms are now at a state where they can legitimately replace manual counters and remove a large burden from conservation organisations.

    “Our results show that deep learning algorithms are now at a state where they can legitimately replace manual 264 counters and remove a large burden from conservation organisations.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationcitation
    § 4. Discussion · p. 12

    Both methods should be viewed as complementary approaches with citizen science data forming the foundation for automated algorithms.

    “Hence, both methods should be viewed as complementary approaches with citizen science data forming the foundation for automated algorithms (Rey et al., 2017).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Rey et al., 2017

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given
    § 4. Discussion · p. 12

    As we move towards fully automated wildlife counts, it only remains to ensure the availability of sufficient training data that is representative of all potential survey images.

    “As we move towards fully automated wildlife 268 counts, it only remains to ensure the availability of su cient training data that is representative of all potential survey 269 images.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given
    § 3. Results · p. 10

    The optimal filtering was determined on the same images used to assess the method, hence there is no division of train and test images.

    “We stress that the optimal filtering was determined on the same images used to assess the method hence there is no division of train and test images.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given
    § 4. Discussion · p. 12

    Automated image processing will allow us to leverage emerging image collection technologies, such as unmanned aerial vehicles, satellite platforms, or fixed camera traps; coupling these advances with automated processing will greatly increase the accuracy of population estimates.

    “The further great advantage of automated 265 image processing is that it will allow us to leverage emerging image collection technologies, such as unmanned aerial 266 vehicles, satellite platforms, or fixed camera traps; coupling these advances in image collection tools with automated 267 processing will greatly increase the accuracy of population estimates.”
    Quote in the paperNot a substantial claim
37 of 37

  • The paper relies on this without support.

    Manually counting aerial images is highly labour intensive.

    “images is highly labour intensive and is often the limiting step in making a population estimate.”

  • The paper relies on this without support.

    Manually counting aerial images is often the limiting step in making a population estimate.

    “images is highly labour intensive and is often the limiting step in making a population estimate.”

  • The authors state this as an assumption.

    The expert counter's count is the true number of wildebeest per image.

    “Error rates for the di erent methods calculated against the assumed true number of wildebeest per image as determined by the expert counter.”

  • The authors state this as an assumption.

    The expert count is taken to be the gold standard for comparison, despite potential bias.

    “we take this count to be the gold standard”

  • The authors state this as an assumption.

    The single expert count is the true number of wildebeest in each image.

    “We compared the accuracy of the methods by calculating the deviation of each method from the single expert 193 count which we assume to be the true number of wildebeest in each image.”

  • The authors state this as an assumption.

    Completely empty images will be processed by citizen scientists with perfect accuracy.

    “the methods in this way we are assuming that completely empty images will be processed by citizen scientists with perfect accuracy.”

  • The paper relies on this without support.

    Obtaining sufficiently large training datasets is the greatest challenge for implementing these algorithms for bespoke applications.

    “Currently the greatest challenge for implementing these algorithms for bespoke applications is obtaining su ciently large training datasets.”
7 of 7

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Recent research shows that machine learning models trained on community data can perform as well as annotators.

    “Recent research shows that ML models trained on community data can perform as well as annotators91.”

    From Perspectives in machine learning for wildlife conservation (2022) · cites this paper as [91]

Cites

Flagged first, then by citations

Research works printed but not linked: 5

  1. Campbell K., 1995, Population trends and distribution of Serengeti herbivores: Implications for management, 117
  2. Laliberte A. S., 2003, Automated wildlife counts from remotely sensed imagery, Wildlife Society Bulletin, 31, 362
  3. Lovejoy T. E., 1987, Conservation strategy for migratory species, American Scientist, 75, 19
  4. Ren S., 2015, Advances in neural information processing systems, 91
  5. Surowiecki J., 2005, The wisdom of crowds

Cited by

Sorted by citations

Related

Authors
Torney, C. J., Lloyd‐Jones, D. J., Chevallier, M., Moyer, D. C., Maliti, H. T., Mwita, M., Kohi, E. M., & Hopcraft, G. C.
Affiliations
  • University of Glasgow (GB)
  • University of Cape Town (ZA)
  • Field Museum of Natural History (US)
  • Tanzania Wildlife Research Institute (TZ)
  • Institute of Biodiversity, Animal Health and Comparative Medicine University of Glasgow Glasgow UK
Corresponding author
Torney, Colin J.
ORCID
Venue
Methods in Ecology and Evolution
Publisher
Wiley
Published
2019-03-05
Volume
10
Issue
6
Pages
779-787
ISSN
2041-210X (ISSN-L)
DOI
10.1111/2041-210x.13165
OpenAlex
W2921243212
Semantic Scholar
0a65791bd2c57440aeb5d3d938f59b4de46420f7
Cited by
110
References
56
Citations by year
2026: 6 · 2025: 20 · 2024: 19 · 2023: 16 · 2022: 17 · 2021: 18 · 2020: 11 · 2019: 3
FWCI
7.93
Type
journal article
Language
en
Open access
Licence
Locations
  • Methods in Ecology and Evolution (publishedVersion)
  • ENLIGHTEN (Jurnal Bimbingan dan Konseling Islam) (acceptedVersion, cc-by-sa)
Funders
  • Horizon 2020 Framework Programme (641918)
  • James S. McDonnell Foundation
Topics
  • Physical Sciences › Environmental Science › Ecology › Wildlife Ecology and Conservation
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Life Sciences › Agricultural and Biological Sciences › Plant Science › Smart Agriculture and AI
Keywords
wildlife abundance estimation, animal counting, citizen science, deep learning, convolutional neural networks, wildebeest, Connochaetes taurinus, Serengeti National Park, Zooniverse, automated image analysis, training data collection

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review