Figure 3. Total number of simulations from an agent-based model that
either returned the correct number of animals (one animal; Correct),
omitted the animal (Omissions), or had multiple counts (Multiple) among
various drone flight patterns, animal movement speeds, and animal
movement patterns (Random = Random Walk, CRW = Correlated Random Walk;
BRW = Biased Random Walk).
Discussion
This research represents the first work investigating the interactions
of multiple drone flight patterns and animal movement behaviors during
simulated surveys. We demonstrate that drone flight patterns can greatly
influence animal count accuracy even over relatively small areas (herein
~22-24 ha) and provide support for use of a rarely
considered drone flight pattern (a lawnmower pattern with 0% image
overlap) for animal monitoring. While lawnmower patterns with large
overlaps allow for development of image mosaics for landscape mapping
(Frazier and Singh, 2021), we found that these flight patterns
increasingly lead to overestimated counts of mobile animals as percent
overlap increased, even when accounting for mosaicking. Our results have
important and often overlooked implications for drone surveys compared
to more commonly applied practices.
Easily programmed drone lawnmower pattern surveys typically use 60-80%
overlapping imagery (Lyons et al. , 2019; Aubert et al. ,
2021), but our results indicate this may have major implications for
multiple count concerns during drone surveys if the animal of interest
is mobile. While it is acknowledged that animals move during surveys
(Brack, Kindel and Oliveira, 2018), many drone surveys assume animals
are stationary (Sudholz et al. , 2022) and create a mosaic image
to more easily count animals and understand distributions (Ezat, Fritsch
and Downs, 2018; De Kock et al. , 2021) without quantifying the
effect of animal movement on counting accuracy. A few field drone
studies have attempted to address animal movement issues post data
collection (Linchant et al. , 2018) and with manual image searches
for clones, partial, or blurred animals after mosaicking (Barbedo and
Vieira Koenigkan, 2018; Lenzi et al. , 2023). Another approach
reviews individual overlapping images, comparing animal shapes, sizes,
and positions to reduce the number of multiple counted animals (Witczuket al. , 2018; Cleguer et al. , 2021; Sudholz et al. ,
2022). This additional post-processing of imagery can be helpful, but
uncertainty remains in their effectiveness considering most animals are
unmarked or indistinguishable from other individuals. These image
reviews are also very labor intensive, time-consuming, and do not
address animal omissions due to their movements (Brack, Kindel and
Oliveira, 2018). Automated image classification approaches are being
developed (Gonzalez et al. , 2016; Dujon et al. , 2021;
Chabot, Stapleton and Francis, 2022; Krishnan et al. , 2023), but
the development of accurate algorithms for aerial animal imagery is
still in its infancy and has many challenges to overcome (Corcoranet al. , 2021; Sudholz et al. , 2022). Addressing the issue
of multiple counting during data collection, as opposed to during post
processing could reduce the likelihood of the multiple-count problem.
Incorporating lawnmower patterns with minimal image overlap may be key,
as noted in the increase in count accuracy of our simulations as overlap
percentage decreased.
Subsampling the landscape or spreading sampling intervals has been
suggested as a means to avoid issues of multiple counts of the same
animal (Witczuk et al. , 2018). However, we found that animal
movement can still influence counts of an individual in these scenarios.
The average count and variation of the transect flight pattern both
doubled when the animal was moving, as opposed to when the animal was
stationary on the landscape. Similarly, for the systematic points flight
pattern, the average count increased 1.5 times, with a slight increase
in count variation as well. However, we also found that during the
systematic points flight (Fig. 1f), an animal moving with directional
persistence resulted in a large percentage of accurate surveys, which
was consistent with other drone flight patterns. Thus, surveys at
systematic points for animals behaving this way may be accurate and
would result in less imagery for post-processing, sequentially leading
to additional time-savings during data preparation and image evaluation.
Ultimately, subsampling the landscape, compared to a full census, will
require correction of counts (Buckland et al. , 2001), but as we
have shown, these corrections should vary depending on if the animals of
interest are anticipated to be mobile during the survey period.
Our simulations confirm that animal movement patterns and speeds
influence whether an animal is correctly counted in drone imagery. The
random and biased random walking animal movement patterns often resulted
in overestimates from multiple counts of an animal traveling back into
the path of the drone after its initial “capture.” An increase in the
animal speed lowered the precision for most flight patterns and animal
walks, with an exception for the correlated walking animal. Therefore,
researchers need to consider animal movement behaviors to avoid count
bias and consequential incorrect management prescriptions (Guerrasioet al. , 2022). Overall, our results emphasize that knowledge of
animal movement patterns can help identify the optimal survey periods
and drone flight patterns to minimize sampling error. To minimize count
error, one might survey using a systematic points flight pattern during
crepuscular periods when certain species, such as white-tailed deer, are
most active (Kammermeyer and Marchinton, 1977). Or depending on the
research question, 0% overlap lawnmower pattern surveys during other
times of day or year when individuals are more stationary, such as when
juveniles have not yet dispersed from natal areas, may also be
appropriate.
Even on our simplified landscape, we observed large amounts of bias
among animal counts during scenario simulations with one mobile animal.
While our assumptions of 100% availability and detectability are highly
unlikely in real-world applications (Gilbert et al. , 2021), for
example, due to visual obstructions above the animals or the ability of
the animal to dive underwater or move under cover (Hodgson, Peel and
Kelly, 2017; Brunton, Leon and Burnett, 2020), this assumption allowed
us to simplify our scenarios and better understand how flight patterns
and animal movements may create counting errors. Typically, surveyors
are concerned with omission rates associated with conventional animal
survey methods (i.e., occupied aircraft and ground surveys) due to
detectability issues, and there are means of addressing some of these
problems (Steinhorst and Samuel, 1989; Samuel et al. , 1992;
Hamilton et al. , 2018; Brack, Kindel, de Oliveira, et al. ,
2023). For example, the inclusion of detection probabilities in
statistical models has greatly improved our ability to estimate animal
populations (Martin et al. , 2012; Corcoran, Denman and Hamilton,
2021), and incorporating detection probabilities into drone-based
estimates would be a helpful advancement (Hodgson, Peel and Kelly, 2017;
Brack, Kindel, de Oliveira, et al. , 2023; Hodgson, Kelly and
Peel, 2023). It is also notable that false positives (i.e., multiple
counts) are less frequent during ground-based and occupied aircraft
surveys, something that researchers using drones need to carefully
consider moving forward (Brack, Kindel and Oliveira, 2018).
We acknowledge other tradeoffs must be considered for drone surveys such
as balancing battery life and line-of-sight limitations during survey
planning (Linchant et al. , 2015; Baxter and Hamilton, 2018).
Hence, trade-offs between the area sampled and survey accuracy may need
to be considered for larger sampling areas. There may also be potential
for increased accuracy with alternative flight patterns that we did not
consider. For example, sea turtle density estimates were calculated
using a modified strip-transect approach with 35-45% frontal overlap
and sequential images used in counts to reduce multiple counting
potential (Sykora-Bodie et al. , 2017; Brack, Kindel, Berto,et al. , 2023), and point count drone surveys with 360-degree
rotations were found to be a promising approach for mesocarnivore
abundance estimates (Bushaw, Ringelman and Rohwer, 2019). In either
case, the method was not thoroughly vetted for accuracy, although this
could be explored in future work. Barbedo and Vieira Koenigkan (2018)
suggest flying multiple drones in formation to collect accurate counts,
acknowledging that animals could otherwise move between survey efforts.
However, they also note that this would greatly increase survey cost,
and that formation flights have many technical challenges. In any case,
researchers should not assume mosaics composed of overlapped images can
be used for both vegetation mapping and animal surveying simultaneously.
Instead, careful thought is needed for drone flight patterns with
objectives related to animal monitoring.
Conclusions
As the use of drones in animal monitoring continues to grow,
consideration of how these survey platforms can be appropriately
incorporated into animal survey techniques is vital. Based on our
results, when using a drone to survey areas similar to our simulations
(~22-24 ha), we recommend that, researchers interested
in animal counts should consider a lawnmower flight pattern with 0%
overlap as an alternative to other more easily programmed, overlapping
patterns. We also recommend that animal life history knowledge be
incorporated in survey design, aligning with stages and times of day
when animals may exhibit more sedentary or more directional movements.
This will allow for the most accurate counts as well as maximize overall
ground coverage area when accounting for limited battery capabilities
(Linchant et al. , 2015). The simulated approach we utilized also
allows for robust inference to investigate a myriad of animal behaviors
and population processes that can be broadly applied across many taxa
and provide guidance of drone applications in a variety of wildlife
management applications. Future efforts with our agent-based modeling
approach can help assess the influence of animal density, distributions,
and detection probabilities to better simulate real-world environments.
Author Contributions
Emma A. Schultz: Conceptualization, Data curation, Formal
analysis, Methodology, Writing- Original draft and EditingNatasha Ellison: Conceptualization, Coding design,
Visualization, Validation, Methodology, Writing- Review and EditingMelanie R. Boudreau: Conceptualization, Methodology, Writing-
Review and Editing Garrett M. Street: Conceptualization,
Methodology, Project administration, Writing- Review and EditingLandon R. Jones: Conceptualization, Methodology, Writing-
Review and Editing Kristine O. Evans: Conceptualization,
Methodology, Writing- Review and Editing Raymond B. Iglay:Conceptualization, Funding acquisition, Methodology, Project
administration, Writing- Review and Editing.
Acknowledgements
We thank Dr. Sathish Samiappan for his review of the manuscript. We also
thank Dr. Pat Clark for his sponsorship and use of high-performance
computing resources. This publication is a contribution of the
Mississippi State University Forest and Wildlife Research Center
[FWRC; McIntire-Stennis Project MISZ-085160] and Mississippi
Agricultural and Forestry Experiment Station [MAFES; HATCH Project
MIS-085180].
Funding
This work was funded by FWRC and MAFES of Mississippi State University
and United States Department of Agriculture Agricultural Research
Service Non-Assistance Cooperative Agreement #58-0200-0-002 via the
Graduate Summer Research Experience in High-Performance Computing and
Agriculture fellowship program in the Geosystems Research Institute at
Mississippi State University.
Data Availability Statement
Data will be available in Supporting Information, Appendix 2, upon
acceptance.