ArticlePDF Available

Abstract and Figures

Population density estimates are important for wildlife conservation and management. Several camera trapping‐based methods for estimating densities have been developed, one of which, the random encounter model (REM), has been widely applied due to its practical advantages such as no need for species‐specific study design. Nevertheless, most of the studies in which REM has been assessed against referenced methods have sampled one population, precluding evaluation of the circumstances under which REM does or does not perform well. At this point, a review of all REM assessments could be useful to provide an overview of method reliability and highlight the main factors determining REM performance. Here we used a combination of literature review and empirical study to compare the performance of REM with independent methods. We reviewed 34 studies where REM was applied to 45 species, reporting 77 REM‐reference density comparisons; and we also sampled 13 populations (ungulates and lagomorphs) in which we assessed REM performance against independent densities. The results suggested that appropriate procedures to estimate REM parameters (namely day range, detection zone and encounter rate) are mandatory to obtain unbiased densities. Deficient estimates of day range and encounter rate lead to an overestimation of density, while deficient estimates of detection zone conducted to underestimations. Finally, the precision achieved by REM was lower than reference methods, mainly because of the high levels of spatial aggregation observed in natural populations. In this situation, simulation‐based results suggest that c. 60 camera placements should be sampled to achieve acceptable precision (i.e. coefficient of variation below 0.20). The wide range of situations and scenarios included in this study allow us to conclude that REM is a reliable method for estimating wildlife population density when using appropriate estimates of REM parameters and sampling designs. Overall, these results pave the way to wider application of REM for monitoring terrestrial mammals. Several camera trapping‐based methods for estimating wildlife population densities have been developed, one of which, the random encounter model (REM), has been widely applied because of its practical advantages such as no need for species‐specific study design. Here we used a combination of literature review and empirical study to compare the performance of REM with “gold standard” reference methods. We reviewed 77 REM‐reference density comparisons, and we also sampled 13 mammal populations. The results suggested that appropriate procedures to estimate REM parameters (namely day range, detection zone and encounter rate) are mandatory to obtain unbiased densities. Deficient estimates of day range and encounter rate led to overestimation of density, while deficient estimates of detection zone resulted in underestimations. In conclusion, the REM is a reliable method for estimating wildlife population densities when using appropriate estimates of REM parameters and sampling designs. Overall, these results pave the way for wider application of REM for monitoring terrestrial mammals.
Content may be subject to copyright.
ORIGINAL RESEARCH
Random encounter model is a reliable method for
estimating population density of multiple species using
camera traps
Pablo Palencia
1
, Patricia Barroso
1
, Joaquı´n Vicente
1
, Tim R. Hofmeester
2
, Javier Ferreres
1
& Pelayo Acevedo
1
1
Instituto de Investigaci ´
on en Recursos Cineg´
eticos (IREC) CSIC-UCLM-JCCM, C/Ronda de Toledo 12, 13071, Ciudad Real, Spain
2
Department of Wildlife, Fish, and Environmental Studies, Swedish University of Agricultural Sciences, SE-90183, Ume˚
a, Sweden
Keywords
Camera trapping, non-invasive, population
abundance, population density, random
encounter model, unmarked
Correspondence
Pablo Palencia, Instituto de Investigaci ´
on en
Recursos Cineg´
eticos (IREC) CSIC-UCLM-
JCCM, C/Ronda de Toledo 12, 13071 Ciudad
Real, Spain. Tel: 926295450 - Ext: 96015;
Fax: 34 926 295 451; E-mail:
palencia.pablo.m@gmail.com
Editor: Marcus Rowcliffe
Associate Editor: Anthony Caravaggi
Received: 16 October 2021; Revised: 25
March 2022; Accepted: 4 April 2022
doi: 10.1002/rse2.269
Remote Sensing in Ecology and
Conservation 2022;8(5):670–682
Abstract
Population density estimates are important for wildlife conservation and man-
agement. Several camera trapping-based methods for estimating densities have
been developed, one of which, the random encounter model (REM), has been
widely applied due to its practical advantages such as no need for species-
specific study design. Nevertheless, most of the studies in which REM has been
assessed against referenced methods have sampled one population, precluding
evaluation of the circumstances under which REM does or does not perform
well. At this point, a review of all REM assessments could be useful to provide
an overview of method reliability and highlight the main factors determining
REM performance. Here we used a combination of literature review and empir-
ical study to compare the performance of REM with independent methods. We
reviewed 34 studies where REM was applied to 45 species, reporting 77 REM-
reference density comparisons; and we also sampled 13 populations (ungulates
and lagomorphs) in which we assessed REM performance against independent
densities. The results suggested that appropriate procedures to estimate REM
parameters (namely day range, detection zone and encounter rate) are manda-
tory to obtain unbiased densities. Deficient estimates of day range and encoun-
ter rate lead to an overestimation of density, while deficient estimates of
detection zone conducted to underestimations. Finally, the precision achieved
by REM was lower than reference methods, mainly because of the high levels of
spatial aggregation observed in natural populations. In this situation,
simulation-based results suggest that c. 60 camera placements should be sam-
pled to achieve acceptable precision (i.e. coefficient of variation below 0.20).
The wide range of situations and scenarios included in this study allow us to
conclude that REM is a reliable method for estimating wildlife population den-
sity when using appropriate estimates of REM parameters and sampling
designs. Overall, these results pave the way to wider application of REM for
monitoring terrestrial mammals.
Introduction
Together with occupancy and species richness, population
density (i.e. the number of individuals per unit area) is a
key state variable in wildlife management and conserva-
tion (Nichols & Williams, 2006). However, obtaining
such information is particularly difficult in some wildlife
species due to low detectability, usually associated with
low population density, elusive behaviour and certain
habitat features, among others (Kindberg et al., 2009). A
plethora of methods (e.g. distance sampling or spatial
capturerecapture [SCR]) have been developed to esti-
mate wildlife population density (e.g. Borchers
et al., 2002; Seber, 1982). Moreover, comparative studies
assessing methods’ performance and reviews of their
applicability to different species have also been developed
670 ª2022 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.
(e.g. Acevedo et al., 2008; ENETWILD consortium
et al., 2019; Meriggi et al., 2008). In this connection, a
recent review concluded that, in general, camera traps are
an effective sampling tool when compared with alterna-
tive ones to collect data about medium-to-large ground-
dwelling mammals and birds (Wearn & Glover-
Kapfer, 2019). Camera traps allow highly standardized
data collection on multiple species with minimal distur-
bance to wildlife, and do not require expert knowledge
for their basic use. From camera trap data, a wide range
of methods can be applied to estimate population density
(Rovero & Zimmermann, 2016).
In this context, the use of camera traps has been firmly
established in recent decades among the non-invasive tools
available to support monitoring programmes for wildlife
population density (Delisle et al., 2021). Initially, the esti-
mation of population density from camera trap data was
earlier limited to marked populations (i.e. those where ani-
mals can be identified individually by natural or artificial
marks) when capturerecapture methods are applied
(Royle et al., 2013). However, most wildlife species do not
have natural marks that enable individual recognition
(hereafter unmarked species). To monitor unmarked popu-
lations with camera traps, physical capture has been
required for individual tagging, which greatly limits the
applicability of capturerecapture methods. Some of its
main limitations are as follows: (i) ethics committee
approval is required for the capture of animals; (ii) highly
qualified staff are needed (e.g. vets to anaesthetize the ani-
mals); (iii) the economic costs and human effort associated
with the capture and tagging of animals are high and (iv)
it defeats the non-invasive nature of camera traps and
could harm the animals. Against this background, methods
to estimate population density without the need for indi-
vidual identification emerged (see Gilbert et al., 2020 for a
review). Specifically, Rowcliffe et al. (2008) described the
random encounter model (REM).
The REM is based on modelling random encounters
between moving animals and static camera traps, taking
into account key variables that affect the encounter rate
(i.e. number of animals detected per sampling unit).
These variables are camera detection zone, defined by its
radius and angle, and the daily distance travelled by an
animal in the population (hereafter, day range). The main
advantage of REM is that individual identification is not
needed, so then REM can be used to monitor both
unmarked and marked populations without the need to
capture and tag animals. Additionally, since the survey
design is not based on target species (i.e. it is not needed
that animals have a reasonable chance of being detected
at more than one camera, so camera spacing is not deter-
mined by target species), more than one species can be
potentially monitored during the same survey (Palencia,
Rowcliffe, et al., 2021; Pfeffer et al., 2018). For all the
above reasons, REM is one of the most widely used meth-
ods to estimate population density of unmarked popula-
tions today (Gilbert et al., 2020)
,
and has been
recommended when tested against other methods. For
instance, when problems related to burst mode perfor-
mance in the camera are observed (Palencia, Rowcliffe,
et al., 2021). In these scenarios, bias is expected in other
methods such as camera trap distance sampling because
some photos are not recorded at the predetermined snap-
shot moments (Howe et al., 2017).
The application of REM was originally limited because
of the difficulties of estimating the parameters necessary
to apply the method, especially day range (Nakashima
et al., 2018; Rovero & Marshall, 2009). In recent years,
however, procedures have been described for estimating
all the parameters required considering camera trap data
only (Hofmeester et al., 2017; Palencia, Fern´andez-L´opez,
et al., 2021; Rowcliffe et al., 2011,2016). These studies
clearly improved the applicability of the method and are
increasing its use in wildlife monitoring (Palencia, Row-
cliffe, et al., 2021; Pfeffer et al., 2018). Other studies have
focused on the statistical development of the method and
software development (Caravaggi, 2017; Jourdain
et al., 2020; Lucas et al., 2015). The REM has been used
in species with different behavioural and ecological traits,
and c. 30 REM studies have been published so far. For
instance, it has been compared against reference densities
on gregarious and non-gregarious carnivores (e.g. lion
Panthera leo, Cusack et al., 2015; red fox Vulpes vulpes
Palencia, Rowcliffe, et al., 2021), ungulates (e.g. Grevy’s
zebra Equus grevyi, Zero et al., 2013; chamois Rupicapra
rupicapra, Kavˇ
ci´
c et al., 2021; moose Alces alces and roe
deer Capreolus capreolus, Pfeffer et al., 2018), lagomorphs
(e.g. European hare Lepus europaeus and Irish hare Lepus
timidus hibernicus, Caravaggi et al., 2016) and Eulipoty-
phla (e.g. European hedgehogs Erinaceus europaeus Schaus
et al., 2020), among others. Nevertheless, (i) most studies
have monitored only a single species/population, preclud-
ing evaluation of the circumstances under which REM
does or does not perform well (but see Rowcliffe
et al., 2008; Pfeffer et al., 2018; Palencia, Rowcliffe,
et al., 2021) and (ii) some of them have considered bibli-
ographic data for some of the parameters required to
derive densities which may introduce bias (Caravaggi
et al., 2016; Cusack et al., 2015; Manzo et al., 2012).
Broadly, most of these studies reported comparable
results (Pfeffer et al., 2018; Rowcliffe et al., 2008; Schaus
et al., 2020), but others reported considerable discrepan-
cies (Chauvenet et al., 2017). Thus, a global overview of
all the available comparisons between REM and reference
densities is timely to provide further insights into which
factors determine the performance of this method.
ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 671
P. Palencia et al. Assessing Random Encounter Model Reliability
This study aimed to provide a comprehensive view of
REM performance across a wide range of species/popula-
tions with different behavioural traits and a large range of
densities and evaluate which factors determine REM relia-
bility. We did this by comparing the densities obtained
with REM with those obtained with other independent
reference methods, mainly drive count, SCR and distance
sampling (see below). We reviewed published studies on
REM and sampled 13 populations. By discussing its
strengths and weaknesses when monitoring wild popula-
tions, the results reported here allowed us to draw robust
conclusions about the potential of REM for monitoring
wildlife populations.
Materials and Methods
Review of published studies
We reviewed all applications of REM reported in pub-
lished peer-reviewed studies. The results were retrieved in
March 2022 by searching the Scopus, PubMed and Web
of Science databases using “random encounter model”,
“unmarked” and “density” as keywords. Of all the studies
retrieved during the search, we were focused on those in
which REM had been compared against a reference
method. From these studies, we extracted the mean value
of the estimated densities for both methods, the target
species, the independent method considered and the
number of camera trap placements sampled (see Appen-
dix S1). Additionally, we evaluated the procedures used
to estimate REM parameters (detection zone, day range
and encounter rate). We considered two categories:
appropriate and deficient. For instance, and considering
the day range as an example, we considered as ‘deficient’
quality those cases in which day range values were
imported from other populations, because of the expected
variation in movement behaviour among populations.
Moreover, those cases in which day range was estimated
for the target population using telemetry data but without
accounting for tortuosity were also considered as ‘defi-
cient’ quality. Day range is expected to be underestimated
(e.g. Marcus Rowcliffe et al., 2012). We considered as ‘ap-
propriate’ quality those cases in which day range was esti-
mated for the target population by (i) using telemetry
data and accounting for tortuosity (Marcus Rowcliffe
et al., 2012), (ii) applying the camera trap-based method
(Palencia, Fern´andez-L ´opez, et al., 2021; Rowcliffe
et al., 2016) or (iii) when observers followed animals and
recorded the total distance covered (Cusack et al., 2015;
Rowcliffe et al., 2008). Further details about the criteria
considered for detection zone and encounter rate, as well
as the qualification reported to each study are shown in
Appendix S1. We did not consider precision in density
estimates in the published studies because most did not
adequately describe how they were estimated (e.g. whether
or not they accounted for variance in all parameters) or
reported explicitly that they had not considered precision
in some of the measured variables (e.g. Balestrieri
et al., 2016; Garrote et al., 2021; Pfeffer et al., 2018).
Field surveys
Study areas and target species
We sampled wild mammal populations at six sites in
Spain. These included a protected area (site A), three
fenced hunting estates (sites B, C and D) and two open
areas where cattle farming and hunting were the main
uses (sites E and F). Site A was located in southern Spain
(Do˜
nana National Park), sites B, C, D, E and F were
located in two mountain chains in central Spain: the
Montes de Toledo (B, C and D) and the Sistema Central
(E and F). Although sites C and D are situated next to
each other, we considered them as two independent study
areas, because they were fenced off and separated by a
road. Sites E and F were sampled over two consecutive
years. Further details of the environmental characteristics
of the sampled sites are given in Appendix S2.
For the target species, we sampled 13 wild populations
(Table 1), including five species of ungulates (red deer
Cervus elaphus; roe deer C. capreolus; fallow deer Dama
dama; mouflon Ovis musimon; and wild boar Sus scrofa)
and one lagomorph (Iberian hare Lepus granatensis). Each
population was surveyed applying REM alongside an
independent reference method (Fig. 1) for comparative
purposes in terms of precision (coefficient of variation,
CV) and consistency in average density values (see details
below). Both surveys overlapped spatially and temporally.
REM: rationale and surveys
The REM models the encounters between animals and
passive detectors (here camera traps) without the require-
ment for individual identification of animals (Rowcliffe
et al., 2008). The REM equation is:
D¼y
tπ
vr2þθðÞ
where yis the number of encounters, tis the total survey
effort, vis the day range and rand Ɵrefer to the effective
radius and angle of the camera detection zone, respec-
tively. To estimate encounter rate, we considered each
time that an individual of the target species entered the
detection zone of the camera trap as a new encounter.
Day range was estimated following Palencia, Rowcliffe,
et al. (2021) using the activity v.1.3.1 and trappingmotion
672 ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Assessing Random Encounter Model Reliability P. Palencia et al.
v.1.0.0 packages in R (Palencia, 2020; Rowcliffe, 2019).
Briefly, speed was measured on each encounter by divid-
ing the distance travelled by the duration of the encoun-
ter; we subsequently identified different movement
behaviours based on the speed measurements. Second, we
estimated activity level, following Rowcliffe et al. (2014).
For each behaviour, we estimated the average speed and
weighted the activity level, taking into account the pro-
portion of time that the population spent on each beha-
viour. Day range was estimated as the sum of the product
of the mean speed and the proportion of the activity level
associated with each behaviour. To estimate detection
zone, we recorded the position (radial distance and angle)
of an animal when it first triggered the camera trap and
then applied a distance sampling analysis to estimate
effective radius and angle (Rowcliffe et al., 2011). The
variance associated with the encounter rate was estimated
by bootstrapping, resampling camera trap locations with
replacements. The overall variance of density estimates
was computed using the delta method (Seber, 1982) and
Table 1. Summary of REM surveys.
Site Species
No. CT
placements Survey period Survey length (days)
CT deployment
height (cm)
Grid spacing
(km) CT brand
Reference
method
A Wild boar 24 SeptemberNovember 21 50 1.5 BUSAT SMR
Fallow deer 24 SeptemberNovember 57 50 1.5 BUSAT TC
B Iberian hare 30 MayJuly 93 20 0.75 BRW-SF DS
Roe deer 20 DecemberApril 138 50 2 BUSAT SCR
Red deer 19 October 15 140 2 BRW-SF DS
C Mouflon 7 MarchMay 58 50 2 LTL DS
Red deer 7 March 25 50 2 LTL DS
D Mouflon 9 MarchMay 58 50 2 LTL DS
Red deer 9 March 25 50 2 LTL DS
E
1
Wild boar 37 FebruaryMarch 30 50 1.5 LTL DC
E
2
Wild boar 17 NovemberDecember 44 50 1.5 LTL DC
F
1
Wild boar 10 FebruaryMarch 30 50 1.5 LTL DC
F
2
Wild boar 14 NovemberDecember 44 50 1.5 LTL DC
All the sites (study areas) are located on Spain. ‘A’ is a protected area, ‘B’, ‘C’ and ‘D’ fenced hunting states, and ‘E’ and ‘F’ two open areas
where cattle farming and hunting were the main uses.
1
represents year 1 surveys and
2
represents year 2 surveys. REM, random encounter
model; CT, camera trap; BRW-SF, Browning Strike Force HD Pro X; BUSAT, Bushnell Aggressor Trophy Cam; LTL, Little Acorn 5310 Series; DS, dis-
tance sampling; SCR, spatial capturerecapture; SMR, spatial markresight; TC, total count; DC, drive counts.
Figure 1. Example of the experimental design of one of the populations surveyed in this study (Iberian hare, Lepus granatensis site B). Crosses
represent camera trap placements for random encounter model application; the grey dashed line represents the line transects for distance
sampling application (the reference method in this case); the continuous black line marks the boundary of the study area. Panels to the right
represent the study area locations in Spain.
ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 673
P. Palencia et al. Assessing Random Encounter Model Reliability
the emdbook v.1.3.12 package in R (Bolker, 2019). The
latter incorporated the variance of all the parameters (en-
counter rate, day range and detection zone). Log-normal
confidence intervals were presented for density estimates
because of the limited sample size and to prevent negative
limits. In the roe deer population, only males were con-
sidered for estimation of encounter rate and speed,
because we were only able to identify males individually
for an independent estimate of density using the reference
method (see details below). Since we did not expect dif-
ferences in the effective detection zone between male and
female roe deer, we also considered females for estimating
the detection zone to increase the sample size and preci-
sion of the estimates.
With respect to field sampling, we used in all the popu-
lations a systematic design with random origin to meet the
assumption of random camera placement relative to ani-
mal movement (Rowcliffe et al., 2013). Camera traps were
deployed facing north on the nearest vertical feature (trees,
poles and so on). Three different camera trap models were
used: Browning Strike Force HD Pro X, Bushnell Aggressor
Trophy Cam and Little Acorn 5310 Series, although the
same model was used within each population (Table 1).
Camera traps were configured to record a burst of photos
at each activation, with the minimum time lapse between
consecutive activations, which allowed us to record the
entire passage of an animal. Cameras were set to be opera-
tional all day, recording nocturnal photos using an infrared
flash. For the deployment of cameras in the field, we fol-
lowed the procedure described by Palencia, Fern´andez-
L´opez, et al. (2021) and in the 10 m closest to the camera,
natural markers such as rocks or branches were placed in
the field of view of the camera at 2.5 m intervals using
ground distance (i.e. accounting for inclination). In the
case of the Iberian hare population, we also placed markers
at 3.7 m from the camera. These markers were later used
to locate the position of the animals in the eld of view of
the camera trap.
Finally, we evaluated the aggregation in encounter
rates. It is well established that most of the variance in
REM is attributable to the variation in encounter rate
between sampling points (Palencia, Rowcliffe, et al., 2021;
Rowcliffe et al., 2008), so a better understanding of the
spatial aggregation in this parameter could be useful to
provide future insights to improve precision. For that, we
fitted Poisson and negative binomial distributions to the
observed encounter rates (Appendix S3).
Independent density estimates from reference
methods
We also sampled all populations using a reference method
generally applied and recognized as reliable for wildlife
population monitoring. All the populations, except for
fallow deer, were monitored exclusively for this study (see
details below). Briefly, we considered distance sampling
for red deer, mouflon and hare populations (e.g. Acevedo
et al., 2008), total counts (TC) for fallow deer (e.g. Grig-
nolio et al., 2020), SCR for roe deer (e.g. Jim´enez
et al., 2013) and spatial markresight (SMR, an extension
of SCR for partially marked populations) and drive
counts (DC) for wild boar (e.g. ENETWILD consortium
et al., 2019; Jim´enez et al., 2017). Further details are
shown in Appendix S2.
Distance sampling (DS): We performed line transect
distance sampling to estimate the density of all red deer,
mouflon and Iberian hare populations. A set of transects
was distributed across the study areas, overlapping the
areas sampled with the REM design (Fig. 1). We carried
out the surveys in September (for red deer and mouflon
populations) and April (for Iberian hare populations),
beginning 1 h after sunset from a vehicle moving at an
average speed of 10 kmh
1
, using a handheld 100 W
spotlight to search a 180°arc in front of the vehicle. We
repeated the surveys over five consecutive nights. When
an animal/group of animals was detected, radial distance
between animal(s) and observer, and the angle between
animals and transect were measured with a telemeter
(Nikon Laser 550AS) and a compass, respectively. We
used Distance Sampling 6.2 software to analyse the data
(Thomas et al., 2010). Data were right truncated when
the probability of detection was lower than 0.15; half-
normal, uniform and hazard rate detection functions were
fitted to the data using cosine, hermite polynomial and
simple polynomial adjustment terms. The best model was
selected according to the AIC (Buckland et al., 2001).
TC: Total counts were performed as part of the
Do˜
nana National Park monitoring program (http://icts.
ebd.csic.es/monitoring-data) and were applied to estimate
fallow deer density at site A. During the rutting season,
two gamekeepers simultaneously sampled open areas and
grassland in a single afternoon (2 h before sunset). The
survey was carried out from a vehicle at an average speed
of 10 kmh
1
. When a group of animals was detected, the
size of the group and the sex and age classes of individu-
als were recorded. We estimated density by dividing the
number of animals observed by the total size of the study
area. Since only one survey was performed, it was not
possible to estimate precision of density. This method
assumes perfect detection (i.e. all the individuals in the
population are detected). To increase the reliability of this
assumption, we carried out the count in the period of
higher detectability of the species (i.e. rutting season) and
at the peak of the activity pattern during the day (i.e.
sunset). Additionally, based on telemetry data from fallow
deer tagged in the study area (Triguero-Oca ˜
na
674 ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Assessing Random Encounter Model Reliability P. Palencia et al.
et al., 2020), we designed a targeted survey, and we sur-
veyed those areas used by the species.
SCR: To estimate roe deer density at site B, we sam-
pled with the camera traps 17 artificial feeding points for
3 months, designed exclusively for roe deer and dis-
tributed throughout the study area. Since we were only
able to identify individuals based on the antlers (size,
shape, length, curvature and number of points), we dis-
carded females and calves and estimated the density of
males (Jim´enez et al., 2013). Data were analysed with
SCR using the oSCR v.0.42.0 package in R (Sutherland
et al., 2019). We tested the performance of both M
0
(which assumes constant baseline detection probability,
p0) and M
b
(where p0 is allowed to vary depending on
the previous capture). We tested M
b
because we used bai-
ted sampling points, animals might respond positively
(trap happiness) and they might be more likely to be cap-
tured subsequent to their initial capture. We also run
three models including factors affecting density. In one of
them, we included habitat (open areas and dense vegeta-
tion areas), in the second we included region (north and
south, because of the natural density gradient reported in
previous studies; Jim´enez et al., 2013) and in the third
model we included statistical interaction between habitat
and region. We also tested a model in which region was
included as factor of p0. Models were compared on the
basis of AIC values (Royle et al., 2013). As the study area
was fenced, we restricted the state space to the fenced
area.
At site A, to estimate wild boar density, we captured
and ear-tagged seven wild boar. Two of these individuals
were also tagged with GPS-GSM collars programmed to
acquire one location every 10 min. As it was not possible
to recognize all the wild boar, we applied SMR. Specifi-
cally for the SMR method, we deployed 61 cameras with
a 500 m inter-camera spacing in two regular grids (5 ×5,
6×6) representatives of the study area. We analysed the
photographic captures of both marked and unmarked
wild boar using an extension of SMR model with incom-
plete identification: the generalized SMR model (gen-
SMR-ID, Jim´
enez et al., 2019). This approach solved two
common problems in SMR studies: the difficulty of read-
ing all the marks and recognizing individuals, and equal
encounter rates in marked and unmarked animals. We
fitted the null model and included telemetry data to allow
inferences about the posterior distribution of σ, and con-
sidered a survey period of 25 days to avoid the effect of
transient animal movement.
DC: We applied DC to estimate wild boar densities at
sites E and F. An average of four drives of
229 54.90 ha (SE) in different scrubland zones in the
study areas were surveyed on separate days. Observers
were placed at fixed locations with an open field of view
(e.g. firebreaks). The DC started at 11:00 and lasted for
4 h. While the observers were in their positions, beaters
with dogs moved across the area. An experienced beater
(J. Ferreres) supervised all the DC, collected all the infor-
mation and minimized the likelihood of double counting.
Assuming that all animals were detected, we estimated
densities by dividing the number of observed animals by
the surveyed area. Multiplying these densities by the area
covered by scrubland, we estimated the total number of
wild boar. It was assumed that at the time of the DC, all
the wild boar were in the scrubland areas, and animals in
the open grassland zones were ignored. Finally, by divid-
ing the total number of individuals by the total area of
the population, we estimated the density.
Comparison of density results
To identify the factors that determine the reliability in
REM, we run a linear mixed-effects model using the bias
as response variable. Bias for each population was esti-
mated as the difference of the REM-density minus the
reference-density, and this value was divided by the
reference-density. Thus, negative values indicate an
underestimation of REM, positive values overestimation
of REM, and 0 correspondence between the densities
obtained with REM and reference method. At this point
we would like to highlight that because of the absence of
reliable precision of densities reported on literature, it
was not possible to include uncertainty in the response
variable, that can lead to an overestimation of the preci-
sion in the model parameters (Behney, 2020; Cressie
et al., 2009). As explanatory variables we considered the
number of camera trap placement log
10
-transformed as
continue, the reference method as a factor with five cate-
gories (distance sampling, DC, dung count, spatial explicit
capturerecapture and TC), the species taxonomic group
as a factor with six categories (Artiodactyl, Carnivora,
Diprotodontia, Eulipotypha, Lagomorpha and Rodentia),
and the quality of the estimation of the REM parameters
(i.e. day range, detection zone and encounter rate) used
as a factor each one with two categories (appropriate and
deficient). All these variables were included as predictors
in a full model. The study was included as a random
effect factor. Raw data for these variables are found in
Appendix S1. The assumptions of normality, homogeneity
and independence in the residuals were assessed following
Zuur et al. (2010).
Results
In the review of the existing literature, we found 34 studies
in which REM was applied to a total of 45 species.
Reported REM densities ranged from 0.07 individualskm
2
ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 675
P. Palencia et al. Assessing Random Encounter Model Reliability
(cougars Puma concolor, Loonam et al., 2021)to468indi-
vidualskm
2
(wallabies Macropus rufogriseus,Rowcliffe
et al., 2008)(AppendixS1). In 25 of these studies, REM
estimates were compared with a reference method, generat-
ing a total of 77 REM-reference method comparisons
(Fig. 2).
In the populations surveyed in this study, the densities
obtained using REM ranged from 0.44 individualskm
2
(males in roe deer population at site B) to 60.55 individ-
ualskm
2
(red deer population at site D) (Fig. 3and
Table 2).
With respect to the LLM, the results did not show an
association of the reference method, the number of place-
ments sampled and the species taxonomic group with the
bias (Fig. 4). We observed an effect of the quality of REM
parameters (appropriate/deficient) in density bias (Fig. 4).
Deficient procedures for the estimation of day range and
encounter rate led to an overestimation of density when
applying REM; while an underestimation in density was
observed when deficient procedures were applied to the
estimation of detection zone (Fig. 4). We also reported a
tendency to overestimate density when appropriate proce-
dures were applied to estimate detection zone (Fig. 4).
The model has an R
2
of 0.51, while the R
2
associated with
the random effect (here the ‘study’) was 0.26.
With respect to precision, we observed overdispersed
encounter rates in all the populations surveyed in this
study, where parameter kranged from 0.04 (fallow deer
population at site A) to 1.39 (wild boar population at site
F
2
). The mean CV was 0.47 in the REM estimates, rang-
ing from 0.34 to 0.75. In contrast, the mean CV of the
reference methods was 0.25, ranging from 0.13 to 0.47.
The REM achieved lower precision than the reference
method in all populations, except for the mouflon popu-
lation at site D (REM-CV =0.42, reference method-
CV =0.47) and the wild boar at site E
2
(REM CV =0.40,
reference method CV =0.39), in which reference meth-
ods were distance sampling and DC, respectively. Based
on this result, we included in Appendix S3 a brief simula-
tion to evaluate the survey design needed for a given level
of precision considering the high levels of aggregation
observed in the populations sampled in this study.
Discussion
The development of new methods to estimate population
densities without the need for individual recognition has
improved the applicability of camera trapping for wildlife
monitoring. However, comparative studies surveying
more than one population and assessing methods reliabil-
ity are scarce. This study, based on a combination of
reviewed and empirical data, shows the potential of REM
for estimating population density, as well as what factors
determines its reliability.
Broadly, we found a strong equivalence between REM
and reference densities (Figs. 2and 3). These results are
Figure 2. Densities plotted in a pairwise comparison between
random encounter model (REM) and reference methods reported in
published studies. Crosses represent mean density values; translucent
ellipses represent 95% log-normal confidence intervals. Those popula-
tions without ellipses are those which did not report variance values
in the original study. A detailed list of references and density values
can be found in Appendix S1. The diagonal line is the line of equality.
Figure 3. Densities plotted in a pairwise comparison between
random encounter model (REM) and reference methods for
populations surveyed in this study. Symbols represent mean density
values, and translucent ellipses, 95% log-normal confidence intervals.
Note that species are grouped according to colour and symbol type.
The diagonal line is the line of equality. Capital letters above the sym-
bols represent the populations at the different sites.
676 ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Assessing Random Encounter Model Reliability P. Palencia et al.
in agreement with most of the studies that have assessed
REM, most of which reported comparable results (Pfeffer
et al., 2018; Rowcliffe et al., 2008; Schaus et al., 2020),
although others reported discrepancies (Chauvenet
et al., 2017). In this respect, our results suggested that
biased REM densities are obtained when REM parameters
(namely day range, detection zone and encounter rate)
are estimated applying deficient procedures (Fig. 4). First,
should be mentioned that, considering the number of
REM and reference method comparisons (N =90), we
are going to discuss the observed effect (point estimate)
in the statistical model, since the absence of significant
differences in some factors could be due to the wide
intervals obtained likely related with limited sample size
(Amrhein et al., 2019). Focusing on encounter rate, we
observed that some studies applied not random (i.e. tar-
geted) designs by setting cameras in placements in which
the presence of animals such as dung piles, footprints or
wildlife trails were observed (e.g. Rahman et al., 2017;
Rovero & Marshall, 2009; Soofi et al., 2017). Additionally,
other studies have considered regular grids, but the place-
ment selected around the predefined point was based on
the presence of wildlife signs (Pfeffer et al., 2018; Zero
et al., 2013), so the camera trap placement is not random.
Targeted designs when applying REM could lead to an
increase in encounter rate, and this could explain the ten-
dency to overestimate density observed (Fig. 4). Regard-
ing the day range, we also reported a tendency to an
overestimation in REM densities when applying deficient
procedures to estimate day range. Looking into the bibli-
ography (Appendix S1), we observed that most of the
deficient procedures to estimate day range are those in
which tagged animals with GPS collars were used to esti-
mate day range without accounting for tortuosity (e.g.
Caravaggi et al., 2016; Garrote et al., 2021; Massei
et al., 2018; Rovero & Marshall, 2009; Zero et al., 2013).
It is well described that estimate day range assuming
straight-line distances between consecutive fixes notably
underestimate day range, and some studies concluded
that more than 5 fixesmin
1
would be required to get tol-
erably accurate estimates (Marcus Rowcliffe et al., 2012;
Sennhenn-Reulen et al., 2017). If day range is underesti-
mated, densities are overestimated when applying REM.
Finally, an underestimation of density when applying
REM using deficient procedures to estimate detection
zone was observed. Looking into literature (Appendix S1),
we observed that habitual practice is to determine the
dimensions of the detection zone by a series of trials in
which the camera was approached by a person from vary-
ing directions (e.g. Cusack et al., 2015; Garrote
et al., 2021; Loonam et al., 2021; Massei et al., 2018;
Rowcliffe et al., 2008). In this respect, some studies have
evidenced that detection zone is determined by different
factors such as environmental conditions and camera trap
settings (Palencia, Vicente, et al., 2021; Rowcliffe
et al., 2011). More relevant, other studies have described
a positive relationship between species body mass and
detection zone dimensions (Hofmeester et al., 2017).
Thus, if detection zone is estimated using human trails,
an overestimation is expected because target species are
usually shorter and smaller than humans. If detection
zone is overestimated, REM densities are underestimated.
Additionally, other habitual practices for the calculation
of detection zone are to use the values reported on
Table 2. Estimated random encounter model (REM) parameter values for each population, where y/t is the encounter rate; v, the average dis-
tance travelled by an individual during a day (day range); r, the radius of detection; and θ, the angle of detection.
Populations Parameters
Site Species y/t(ind(camday)
1
)v(kmday
1
)r(km) θ(rad)
A Fallow deer 0.203 (0.194) 5.776 (1.596) 0.0088 (0.0004) 0.733 (0.083)
Wild boar 0.600 (0.270) 15.770 (1.931) 0.0080 (0.124) 0.733 (0)
B Iberian hare 0.144 (0.040) 4.069 (0.752) 0.0059 (0.0005) 0.911 (0.121)
Roe deer 0.012 (0.005) 6.644 (2.436) 0.0049 (0.0003) 0.733 (0.238)
Red deer 0.670 (0.254) 4.020 (0.420) 0.0046 (0.0001) 0.959 (0.0001)
C Mouflon 0.181 (0.056) 6.112 (1.213) 0.0049 (0.0003) 0.959 (0.083)
Red deer 0.317 (0.039) 1.840 (0.446) 0.0045 (0.0002) 0.641 (0.049)
D Mouflon 0.063 (0.026) 6.112 (1.213) 0.0049 (0.0003) 0.959 (0.083)
Red deer 0.995 (0.714) 4.383 (0.472) 0.0045 (0.0002) 0.641 (0.049)
E Wild boar
1
0.063 (0.013) 7.097 (2.086) 0.0034 (0.0004) 0.380 (0.052)
Wild boar
2
0.149 (0.063) 6.482 (1.465) 0.0034 (0.0004) 0.380 (0.052)
F Wild boar
1
0.280 (0.091) 6.326 (1.345) 0.0030 (0.0001) 0.582 (0.050)
Wild boar
2
0.135 (0.059) 7.753 (1.960) 0.0029 (0.0006) 0.582 (0.050)
Data represent means (standard error). All the sites (study areas) are located on Spain. ‘A’ is a protected area, ‘B’, ‘C’ and ‘D’ fenced hunting
states, and ‘E’ and ‘F’ two open areas where cattle farming and hunting were the main uses. Subscripts:
1
: year 1,
2
: year 2.
ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 677
P. Palencia et al. Assessing Random Encounter Model Reliability
manuals (e.g. Pettigrew et al., 2021; Popova et al., 2019)
or to take reference values from literature (e.g. Soofi
et al., 2017; Zero et al., 2013); both approaches could lead
to biased detection zones, and consequently, biased densi-
ties. On the other hand, underestimation of detection
zone could lead to overestimation of density. Theoreti-
cally, detection zone size is estimated accurately using
detection distances and applying distance sampling if
detection probability is certain for at least some distance
from the camera (Rowcliffe et al., 2011), which can be
achieved by setting the cameras at shoulder height of the
target species (Palencia, Vicente, et al., 2021). Thus,
future studies in which appropriate procedures are
applied to estimate all the parameters are necessary to
confirm that, under these scenarios, accurate REM densi-
ties are estimated. Broadly, deficient procedures for the
estimation of REM parameters will be lead to biased den-
sities. Thus, a best practice guide for the application of
REM should include the estimation of all the parameters
for the target population applying reliable procedures (see
Appendix S1) together with random camera placements
relative to animal movement (Rowcliffe et al., 2013).
Additionally, variance in all the parameters should be
considered when estimating density precision. Random
designs can lead to low sample size but increasing the
sampling effort by increasing the number of cameras or
survey length should be always considered, rather than
using attractants or targeted designs.
Regarding the number of camera trap placements,
the reference method and the species taxonomic group,
we did not observe relevant relationships (i.e. values for
the estimate close to 0, Fig. 4) with bias in density esti-
mations from REM. With respect to the reference
methods, we acknowledge that there are relevant differ-
ences among them, for instance, in the estimation of
probability of detection. However, other practical rea-
sons usually determine which method is applied in
monitoring programmes, so we decided to include all
of them in the comparisons, and not only the robust
ones (Borchers et al., 2002). Regarding the species taxo-
nomic group, the most relevant patterns were associated
with Eulipotypha and Lagomorpha. Both groups showed
a slight tendency to underestimate REM densities. Con-
sidering the low number of studies that sampled these
groups, further comparisons in these taxonomic groups
are still necessary.
Figure 4. Left panel: coefficients for the predictors included in the LLM model to valuate REM reliability. Bias was estimated as the difference of
the REM-density minus the reference-density, and this value divided by the reference density. Colours represent different levels of the same factor.
The reference categories for day range, encounter rate and detection zone quality were “appropriate”, for reference method was “distance sam-
pling” and for taxonomic group was “Artiodactyl”. Right panel: predicted values for day range, encounter rate and detection zone considering
two categories (app.: appropriate and def.: deficient) according to the quality of the procedures applied to estimate these parameters. Error bars
represent 95% confidence intervals. REM, random encounter model.
678 ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Assessing Random Encounter Model Reliability P. Palencia et al.
After estimating all REM parameters for the target pop-
ulation, the Spanish surveys demonstrated their reliability
in ungulates and Iberian hare in different environmental
scenarios and a wide range of densities (from 0.44 to
60.55 individualskm
2
). This range of densities covers the
vast majority of the population densities reported in wild-
life monitoring programmes (Fig. 2). We also surveyed
gregarious (e.g. fallow deer, mouflon) and non-gregarious
(e.g. Iberian hare) mammals and the results showed a high
degree of equivalence between REM and reference densities
(Fig. 3). For REM application, it is not needed that ani-
mals have a reasonable chance of being detected at more
than one camera, which means that multispecies studies
can be considered. Here, we surveyed red deer and mou-
flon populations at sites C and D in parallel, while previ-
ous studies have surveyed the ungulate and carnivore
community (Palencia, Rowcliffe, et al., 2021; Pfeffer
et al., 2018). Thus, in addition to the advantages of REM
discussed above, we would like to highlight the potential of
REM for monitoring more than one species using the same
survey design. This is less common in other reference
monitoring methods. Distance sampling surveys, for exam-
ple, should be conducted during the season of highest
detectability when target species activity is at its peak
(Buckland et al., 2001). On the other hand, when using
SCR methods, the distance between traps depends on the
home range of the species (Royle et al., 2013). Home range
and activity periods are usually species specific.
A potential point of concern is that our REM estimates
showed relatively low precision (Figs. 2and 3). The low
precision of REM estimates has been reported before
(ENETWILD consortium et al., 2019; Palencia, Rowcliffe,
et al., 2021). Although we considered variance in all esti-
mated model parameters, most of the final density vari-
ance was attributable to the variation in encounter rate
between sampling points (Table 2), and it has been
described in other studies (Howe et al., 2017). The distri-
bution of animals is not uniform but aggregated, and
usually overdispersed (i.e. the variance is greater than the
mean). In this study, for example, we monitored highly
aggregated populations (maximum kof negative binomial
distribution for encounter rate was 1.4) throughout the
study area. In highly aggregated scenarios (e.g. k=0.05),
when applying REM, a minimum of 60 camera trap
placements (sampling locations) should be sampled to
obtain a CV of <0.2 (which is a rule of thumb for moni-
toring programmes; Pollock et al., 1990) (Appendix S3).
The human effort and cost associated with sampling more
than 60 placements would not be feasible in some man-
agement programmes, which may limit the applicability
of REM for wildlife monitoring (but CV of <0.2 may not
always be necessary’). In this respect, some studies have
shown seasonal variation in encounter rates (Kays
et al., 2021; Kolowski et al., 2021), so a general recom-
mendation when applying REM could be to survey popu-
lations when low aggregation is expected. This could help
to optimise human effort. Considering all the above,
future advances in REM should be focused on optimizing
surveys design to improve density precision.
In addition to the advantages highlighted above, we
would also like to highlight that previous studies evaluat-
ing the costs associated with REM and reference methods
have concluded that REM is cost-effective in the long
term despite the high start-up costs (Cusack et al., 2015;
Pettigrew et al., 2021; Rovero & Marshall, 2009; Schaus
et al., 2020; Zero et al., 2013). The REM is recommended
particularly when the assumption of population closure is
violated (i.e. density is expected to change during the sur-
vey), since it provides an average density across the sam-
pling period, but not biased results. It should be
mentioned that violation of closure is common, for
instance, when monitoring game species during the hunt-
ing season (ENETWILD consortium et al., 2019). The
REM could also be recommended for well-defined areas
(such as forests surrounded by agricultural lands or
fenced hunting areas). On the other hand, a significant
limitation is that REM estimates average density over the
entire study area and survey period, which limits its
potential to identify spatial variation in densities (Car-
avaggi et al., 2016; Rowcliffe et al., 2008).
In conclusion, our results showed that REM could be a
reliable alternative for monitoring wildlife populations and
is highly recommended when parameters (day range,
encounter rate and detection zone) are adequately esti-
mated, and survey effort, in terms of camera trap place-
ments, is appropriate to obtain precise estimates. Since it
was first described, the REM has been well received and
widely applied by the scientific community (c. 30 applica-
tions) and even included in citizen science projects (Schaus
et al., 2020). It has also been proposed as a reference
method for monitoring certain species at European level
(ENETWILD consortium et al., 2019). Here we provide
strong evidence of the reliability of REM, highlighting the
priority aspect of estimating REM parameters properly and
using appropriate survey design. These results, along with
the practical recommendations for improving precision,
and its methodological advantages relative to other meth-
ods described above, allow us to conclude that REM could
be recommended for monitoring wildlife population den-
sity, especially managers and partitioners responsible for
monitoring wildlife populations.
Acknowledgements
The authors would like to thank the Quintos de Mora
and Do˜
nana Biological Reserve gamekeepers and
ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 679
P. Palencia et al. Assessing Random Encounter Model Reliability
managers for their kind support during the fieldwork, and
especially
´
Angel Moreno, Amanda Garcı´a, Alejandro Ara-
sanz, Rafael Corredor, Enrique Corredor and Jes´us. We
also would like to thank Jos´e Jimenez for his advice and
assistance in the gen-SMR-ID analysis, and Pablo Iglesias
for his assistance in image processing. We are grateful to
the two anonymous reviewers for their detailed and helpful
comments, which lead to a substantially improved manu-
script. PP received support from the MINECO-UCLM
through an FPU grant (FPU16/00039) and mobility grant
(EST19/00481). This work was partly funded by the
HAWIPO project MICINN (PID2019-111699RB-I00).
Data Availability Statement
Data available via Zenodo https://zenodo.org/record/
6784645#.Ytaw47bP07E. The R code, working data and
vignette to run a REM analysis are available on https://
github.com/PabloPalencia/CameraTrappingAnalysis/tree/
main/REM.
References
Acevedo, P., Ruiz-Fons, F., Vicente, J., Reyes-Garcı´a, A.R.,
Alzaga, V. & Gort´azar, C. (2008) Estimating red deer
abundance in a wide range of management situations in
Mediterranean habitats. Journal of Zoology,276(1), 3747.
https://doi.org/10.1111/j.1469-7998.2008.00464.x
Amrhein, V., Greenland, S. & McShane, B. (2019) Scientists
rise up against statistical significance. Nature,567(7748),
305307. https://doi.org/10.1038/d41586-019-00857-9
Balestrieri, A., Ruiz-Gonzalez, A., Vergara, M., Capelli, E.,
Tirozzi, P., Alfino, S. et al. (2016) Pine marten density in
lowland riparian woods: a test of the random encounter
model based on genetic data. Mammalian Biology,81(5),
439446. https://doi.org/10.1016/j.mambio.2016.05.005
Behney, A.C. (2020) Ignoring uncertainty in predictor
variables leads to false confidence in results: a case study of
duck habitat use. Ecosphere,11(10), e03273. https://doi.org/
10.1002/ecs2.3273
Bolker, B. (2019). emdbook: Ecological Models and Data in R;
R package version 1.3.12. Available from: https://cran.
microsoft.com/snapshot/2022-02-14/web/packages/emdbook/
index.html
Borchers, D.L., Buckland, S.T. & Zucchini, W. (2002)
Estimating animal abundance: closed populations. London:
Springer. ISBN 1852335602. 2002.
Buckland, S.T., Anderson, D.R., Burnham, K.P., Laake, J.L.,
Borchers, D.L. & Thomas, L. (2001) Introduction to distance
sampling: estimating abundance of biological populations.
Oxford, UK: Oxford University Press.
Caravaggi, A. (2017) remBoot: an R package for random
encounter modelling. The Journal of Open Source Software,2
(10), 176. https://doi.org/10.21105/joss.00176
Caravaggi, A., Zaccaroni, M., Riga, F., Schai-Braun, S.C., Dick,
J.T.A.,Montgomery,W.I.etal.(2016) An invasive-native
mammalian species replacement process captured by camera trap
survey random encounter models. Remote Sensing in Ecology and
Conservation,2(1), 4558. https://doi.org/10.1002/rse2.11
Chauvenet, A.L.M., Gill, R.M.A., Smith, G.C., Ward, A.I. &
Massei, G. (2017) Quantifying the bias in density estimated
from distance sampling and camera trapping of unmarked
individuals. Ecological Modelling,350,7986. https://doi.org/
10.1016/j.ecolmodel.2017.02.007
Cressie, N., Calder, C.A., Clark, J.S., Hoef, J.M.V. & Wikle,
C.K. (2009) Accounting for uncertainty in ecological
analysis: the strengths and limitations of hierarchical
statistical modeling. Ecological Applications,19(3), 553570.
https://doi.org/10.1890/07-0744.1
Cusack, J.J., Swanson, A., Coulson, T., Packer, C., Carbone, C.,
Dickman, A.J. et al. (2015) Applying a random encounter
model to estimate lion density from camera traps in
Serengeti National Park, Tanzania. The Journal of Wildlife
Management,79(6), 10141021. https://doi.org/10.1002/
jwmg.902
Delisle, Z.J., Flaherty, E.A., Nobbe, M.R., Wzientek, C.M. &
Swihart, R.K. (2021) Next-generation camera trapping:
systematic review of historic trends suggests keys to
expanded research applications in ecology and conservation.
Frontiers in Ecology and Evolution,9, 97.
ENETWILD consortium, Grignolio, S., Apollonio, M., Brivio,
F., Vicente, J., Acevedo, P. et al. (2020) Guidance on
estimation of abundance and density data of wild ruminant
population: methods, challenges, possibilities. EFSA
Supporting Publications,17(6), 1876E. https://doi.org/10.
2903/sp.efsa.2020.EN-1876
ENETWILD-consortium, Vicente, J., Palencia, P., Plhal, R.,
Blanco-Aguiar, J.A., Laguna, E. et al. (2019) Harmonization of
the use of hunting statistics for wild boar density estimation in
different study areas. EFSA Supporting Publications,16(9).
https://doi.org/10.2903/sp.efsa.2019.EN-1706
Garrote, G., P´
erez de Ayala, R.,
´
Alvarez, A., Martı´n, J.M., Ruiz,
M., de Lillo, S. et al. (2021) Improving the random
encounter model method to estimate carnivore densities
using data generated by conventional camera-trap design.
Oryx,55(1), 99104. https://doi.org/10.1017/
S0030605318001618
Gilbert, N.A., Clare, J.D.J., Stenglein, J.L. & Zuckerberg, B.
(2020) Abundance estimation of unmarked animals based
on camera-trap data. Conservation Biology, cobi.13517.
https://doi.org/10.1111/cobi.13517
Hofmeester, T.R., Rowcliffe, J.M. & Jansen, P.A. (2017) A
simple method for estimating the effective detection distance
of camera traps. Remote Sensing in Ecology and Conservation,
3(2), 8189. https://doi.org/10.1002/rse2.25
Howe, E.J., Buckland, S.T., Despr´es-Einspenner, M.L. & K ¨
uhl,
H.S. (2017) Distance sampling with camera traps. Methods
in Ecology and Evolution,8(11), 15581565.
680 ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Assessing Random Encounter Model Reliability P. Palencia et al.
Jim´enez, J., Chandler, R.B., Tobajas, J., Descalzo, E., Mateo, R.
& Ferreras, P. (2019) Generalized spatial markresight
models with incomplete identification: an application to red
fox density estimates. Ecology and Evolution,9(8), ece3.5077.
https://doi.org/10.1002/ece3.5077
Jim´enez, J., Higuero, R., Charre-Medellin, J.F. & Acevedo, P.
(2017) Spatial mark-resight models to estimate feral pig
population density. Hystrix,28(2), 208213. https://doi.org/
10.4404/hystrix
Jim´enez, J., Rodrı´guez, C. & Moreno,
´
A. (2013) Estima de una
poblaci´on de corzo mediante modelos de captura-recaptura
cl´asicos y espacialmente explı´citos. Galemys,25,112.
https://doi.org/10.7325/Galemys.2013.A01
Jourdain, N.O.A.S., Cole, D.J., Ridout, M.S. & Rowcliffe, J.M.
(2020) Statistical development of animal density estimation
using random encounter modelling. Journal of Agricultural,
Biological and Environmental Statistics. https://doi.org/10.
1007/s13253-020-00385-4
Kavˇci´c, K., Palencia, P., Apollonio, M., Vicente, J. &
ˇ
Sprem, N.
(2021) Random encounter model to estimate density of
mountain-dwelling ungulate. European Journal of Wildlife
Research,67(5), 18.
Kays, R., Hody, A., Jachowski, D.S. & Parsons, A.W. (2021)
Empirical evaluation of the spatial scale and detection
process of camera trap surveys. Movement Ecology,9(1), 1
13. https://doi.org/10.1186/s40462-021-00277-3
Kindberg,J.,Ericsson,G.&Swenson,J.E.(2009)
Monitoring rare or elusive large mammals using
effort-corrected voluntary observers. Biological
Conservation,142(1), 159165. https://doi.org/10.1016/j.
biocon.2008.10.009
Kolowski, J.M., Oley, J. & McShea, W.J. (2021) High-density
camera trap grid reveals lack of consistency in detection and
capture rates across space and time. Ecosphere,12(2),
e03350. https://doi.org/10.1002/ecs2.3350
Loonam, K.E., Ausband, D.E., Lukacs, P.M., Mitchell, M.S. &
Robinson, H.S. (2021) Estimating abundance of an
unmarked, low-density species using cameras. The Journal of
Wildlife Management,85(1), 8796. https://doi.org/10.1002/
jwmg.21950
Lucas, T.C.D., Moorcroft, E.A., Freeman, R., Rowcliffe, J.M. &
Jones, K.E. (2015) A generalised random encounter model
for estimating animal density with remote sensor data.
Methods in Ecology and Evolution,6(5), 500509. https://doi.
org/10.1111/2041-210x.12346
Manzo, E., Bartolommei, P., Rowcliffe, J.M. & Cozzolino, R.
(2012) Estimation of population density of European pine
marten in central Italy using camera trapping. Acta
Theriologica,57(2), 165172. https://doi.org/10.1007/s13364-
011-0055-8
Marcus Rowcliffe, J., Carbone, C., Kays, R., Kranstauber, B. &
Jansen, P.A. (2012) Bias in estimating animal travel distance:
the effect of sampling frequency. Methods in Ecology and
Evolution,3(4), 653662.
Massei, G., Coats, J., Lambert, M.S., Pietravalle, S., Gill, R. & Cowan,
D. (2018) Camera traps and activity signs to estimate wild boar
density and derive abundance indices. Pest Management Science,
74(4), 853860. https://doi.org/10.1002/ps.4763
Meriggi, A., Sotti, F., Lamberti, P. & Gilio, N. (2008) A review
of the methods for monitoring roe deer European
populations with particular reference to Italy. Hystrix,19(2),
103120. https://doi.org/10.4404/hystrix-19.2-4421
Nakashima, Y., Fukasawa, K. & Samejima, H. (2018)
Estimating animal density without individual recognition
using information derivable exclusively from camera traps.
Journal of Applied Ecology,55(2), 735744. https://doi.org/
10.1111/1365-2664.13059
Nichols, J. & Williams, B. (2006) Monitoring for conservation.
Trends in Ecology & Evolution,21(12), 668673. https://doi.
org/10.1016/j.tree.2006.08.007
Palencia, P. (2020) Trappingmotion: integrate camera-
trapping in movement and behavioural studies. R package
version 1.0.0. Available at: https://github.com/PabloPalencia/
trappingmotion [Accessed 29th April 2021].
Palencia, P., Fern´andez-L ´opez, J., Vicente, J. & Acevedo, P.
(2021) Innovations in movement and behavioural ecology
from camera traps: day range as model parameter. Methods
in Ecology and Evolution, 2041210X.13609. https://doi.org/
10.1111/2041-210X.13609
Palencia, P., Rowcliffe, J.M., Vicente, J. & Acevedo, P. (2021)
Assessing the camera trap methodologies used to estimate
density of unmarked populations. Journal of Applied
Ecology, 1365-2664.13913. https://doi.org/10.1111/1365-2664.
13913
Palencia, P., Vicente, J., Soriguer, R.C. & Acevedo, P. (2021)
Towards a best-practices guide for camera trapping:
assessing differences among camera trap models and settings
under field conditions. Journal of Zoology.https://doi.org/10.
1111/jzo.12945
Pettigrew, P., Sigouin, D. & St-Laurent, M. (2021) Testing the
precision and sensitivity of density estimates obtained with a
camera-trap method revealed limitations and opportunities.
Ecology and Evolution, ece3.7619. https://doi.org/10.1002/
ece3.7619
Pfeffer, S.E., Spitzer, R., Allen, A.M., Hofmeester, T.R.,
Ericsson, G., Widemo, F. et al. (2018) Pictures or pellets?
Comparing camera trapping and dung counts as methods
for estimating population densities of ungulates. Remote
Sensing in Ecology and Conservation,4(2), 173183. https://
doi.org/10.1002/rse2.67
Pollock, K.H., Nichols, J.D., Brownie, C. & Hines, J.E. (1990)
Statistical inference for capture-recapture experiments.
Wildlife Monographs,397.
Popova, E.D., Ahmed, A.A., Genov, P.V., Hodzha, M.M. &
Zlatanova, D.P. (2019) Estimating population density of roe
deer Capreolus capreolus (L.) (Artiodactyla: Cervidae) for
better management: comparison of three methods. Acta
Zoologica Bulgarica,71(4), 531536.
ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London. 681
P. Palencia et al. Assessing Random Encounter Model Reliability
Rahman, D.A., Gonzalez, G. & Aulagnier, S. (2017) Population
size, distribution and status of the remote and critically
endangered Bawean deer Axis kuhlii.Oryx,51(4), 665672.
https://doi.org/10.1017/S0030605316000429
Rovero, F. & Zimmermann, F. (2016) Camera Trapping for
Wildlife Research (Pelagic Pu).
Rovero, F. & Marshall, A.R. (2009) Camera trapping
photographic rate as an index of density in forest ungulates.
Journal of Applied Ecology,46(5), 10111017. https://doi.org/
10.1111/j.1365-2664.2009.01705.x
Rowcliffe, J.M. (2019). activity: Animal activity statistics. R
package version 1.3. Available from: https://CRAN.R-project.
org/package=activity
Rowcliffe, J.M., Carbone, C., Jansen, P.A., Kays, R. &
Kranstauber, B. (2011) Quantifying the sensitivity of camera
traps: an adapted distance sampling approach. Methods in
Ecology and Evolution,2(5), 464476. https://doi.org/10.
1111/j.2041-210X.2011.00094.x
Rowcliffe, J.M., Field, J., Turvey, S.T. & Carbone, C. (2008)
Estimating animal density using camera traps without the
need for individual recognition. Journal of Applied Ecology,
45(4), 12281236. https://doi.org/10.1111/j.1365-2664.2008.
01473.x
Rowcliffe, J.M., Jansen, P.A., Kays, R., Kranstauber, B. &
Carbone, C. (2016) Wildlife speed cameras: measuring
animal travel speed and day range using camera traps.
Remote Sensing in Ecology and Conservation,2(2), 8494.
https://doi.org/10.1002/rse2.17
Rowcliffe, J.M., Kays, R., Carbone, C. & Jansen, P.A. (2013)
Clarifying assumptions behind the estimation of animal
density from camera trap rates. Journal of Wildlife
Management.https://doi.org/10.1002/jwmg.533
Rowcliffe, J.M., Kays, R., Kranstauber, B., Carbone, C. &
Jansen, P.A. (2014) Quantifying levels of animal activity
using camera trap data. Methods in ecology and evolution,5
(11), 11701179.
Royle, J.A., Chandler, R.B., Sollmann, R. & Gardner, B. (2013)
Spatial capture-recapture. Oxford, UK: Academic Press.
Schaus, J., Uzal, A., Gentle, L.K., Baker, P.J., Bearman-Brown,
L., Bullion, S. et al. (2020) Application of the random
encounter model in citizen science projects to monitor
animal densities. Remote Sensing in Ecology and
Conservation.https://doi.org/10.1002/rse2.153
Seber, G.A.F. (1982) The estimation of animal abundance and
related parameters, Vol. 16. New York: Macmillan. https://
doi.org/10.1002/bimj.19740160115
Sennhenn-Reulen, H., Diedhiou, L., Klapproth, M. & Zinner,
D. (2017) Estimation of baboon daily travel distances by
means of point samplingthe magnitude of underestimation.
Primate Biology,4(2), 143151. https://doi.org/10.5194/pb-4-
143-2017
Soofi, M., Ghoddousi, A., Hamidi, A.K., Ghasemi, B., Egli, L.,
Voinopol-Sassu, A.J. et al. (2017) Precision and reliability of
indirect population assessments for the Caspian red deer
Cervus elaphus maral.Wildlife Biology,2017(17). https://doi.
org/10.2981/wlb.00230
Sutherland, C., Royle, J.A. & Linden, D.W. (2019) oSCR: a
spatial capturerecapture R package for inference about
spatial ecological processes. Ecography,42(9), 14591469.
https://doi.org/10.1111/ecog.04551
Thomas, L., Buckland, S.T., Rexstad, E.A., Laake, J.L.,
Strindberg, S., Hedley, S.L. et al. (2010) Distance software:
design and analysis of distance sampling surveys for
estimating population size. Journal of Applied Ecology,47(1),
514. https://doi.org/10.1111/j.1365-2664.2009.01737.x
Triguero-Oca˜
na, R., Vicente, J., Palencia, P., Laguna, E. &
Acevedo, P. (2020) Quantifying wildlife-livestock
interactions and their spatio-temporal patterns: is regular
grid camera trapping a suitable approach? Ecological
Indicators,117, 106565. https://doi.org/10.1016/j.ecolind.
2020.106565
Wearn, O.R. & Glover-Kapfer, P. (2019) Snap happy: camera
traps are an effective sampling tool when compared with
alternative methods. Royal Society Open Science,6(3),
181748. https://doi.org/10.1098/rsos.181748
Zero, V.H., Sundaresan, S.R., O’Brien, T.G. & Kinnaird, M.F.
(2013) Monitoring an endangered savannah ungulate,
Grevy’s zebra Equus grevyi: choosing a method for
estimating population densities. Oryx,47(3), 410419.
https://doi.org/10.1017/S0030605312000324
Zuur, A.F., Ieno, E.N. & Elphick, C.S. (2010) A protocol for
data exploration to avoid common statistical problems.
Methods in Ecology and Evolution,1(1), 314. https://doi.
org/10.1111/j.2041-210X.2009.00001.x
Supporting Information
Additional supporting information may be found online
in the Supporting Information section at the end of the
article.
Appendix S1. Summary of random encounter model
published studies.
Appendix S2. Study areas and reference methods details.
Appendix S3. Assessing improvements in precision.
682 ª2022 The Authors. Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.
Assessing Random Encounter Model Reliability P. Palencia et al.
... Both the FMP and viewshed density estimators rely on the assumption that the space sampled by transects/cameras are representative of the study area. In their review of 34 studies comparing REM estimates with reference methods, Palencia et al. (2022) found that studies with targeted camera placement (e.g., on game trails or latrine sites) had positive bias in density estimates due to increased encounter rates, as many species preferentially use roads and trails to travel (e.g., Tanwar, Sadhu, and Jhala 2021;Montalvo et al. 2023). Lyet et al. (2023) similarly found that targeted camera placement can confound the violation of other assumptions: the negative bias in STE density estimates from inflated viewshed measurements was FIGURE 2 | Relationship between estimates of population density from random FMP track surveys (x-axis) and other methods (y-axis). ...
... Simulation work by Chauvenet et al. (2017) and Hayashi and Iijima (2022) suggests bias can be introduced by group size; however, the camera detection radii used in these two studies were large (18 and 13 m, respectively), and simulated animals grouped relatively close together; whereas our estimated detection distances were small (ranging 4-7 m), and even for large groups of sika deer or wild boar, it was rare for more than one individual to be within that 4-7 m range. Other studies from the field suggest mixed results: REM estimates of wild boar (a social, grouping species) from Palencia et al. (2022) were generally larger than the reference estimates, but earlier work by Cusack et al. (2015) found that REM estimates of lion density in Serengeti National Park matched well with reference estimates, so long as encounters were considered of individuals, not groups. Lyet et al. (2023) found that STE density estimates became more negatively biased for species with larger group size. ...
... Two other assumptions related to animal movement rate should be especially considered by managers. First, for the FMP, REM, and TTE, practitioners should assess their ability to obtain independent estimates of movement rate because of this parameter's influence on density estimates Keeping and Pelletier 2014;Palencia et al. 2022). We believe our use of movement rate estimates from the literature was justified as they came from the same study area; however, these movement rates can vary by season ) and year (Waller et al. 2024), whereas we used the same estimates across years. ...
Article
Full-text available
Population density is a valuable metric used to manage wildlife populations. In the Russian Far East, managers use the Formozov‐ Malyushev‐Pereleshin (FMP) snow tracking method to estimate densities of ungulates for hunting management. The FMP also informs Amur tiger (Panthera tigris altaica) conservation since estimates of prey density and biomass help inform conservation interventions. Yet, climate change and challenges with survey design call into question the reliability of the FMP. Camera traps offer a promising alternative, but they remain unexplored for monitoring tiger prey density. Over three years (2020‐2022), we used the FMP and camera‐based methods to estimate densities of four prey species of the Amur tiger in the Sikhote‐ Alin Biosphere Reserve, Russian Far East: wild boar (Sus scrofa), red deer (Cervus canadensis), roe deer (Capreolus pygargus), and sika deer (Cervus nippon). We compared FMP results from snow track survey routes either along trails, or along routes representative of the study area, and estimates derived from camera data using the random encounter model (REM), space‐to‐event model (STE), and time‐to‐event model (TTE). We found that density estimates from representative routes were typically lower than routes along trails and indicated different relative densities of prey. Density estimates from camera traps and representative track surveys were generally similar with no significant relative bias, but precision was poor for all methods. Differences between estimates were amplified when converted to prey biomass, particularly with larger, more abundant prey, which poses a challenge for their utility for tiger managers. We conclude camera traps can offer an alternative to snow track surveys when monitoring unmarked prey, but we caution that they require considerably more resources to implement. Tiger managers should be especially cautious when extrapolating density to estimates of prey biomass, and we encourage future research to develop more robust methods for doing so.
... One common goal of camera trapping is to estimate population densities of mammal species as to inform conservation and management. Among these, the random encounter model (REM) (Rowcliffe et al., 2008) is one of the most reliable and most often used approaches (Palencia et al., 2022;Schaus et al., 2020). REM estimates animal densities from camera-trap data by correcting capture rates for a set of biological variables of the animals (average animal group size, speed and activity level) and characteristics of camera sensors. ...
... The minimum sample size for parameter estimation was set at 150 (Bollen et al., 2023;Palencia et al., 2022; Density estimation with vertical camera traps S. He et al. ...
... One unique feature of our field application is that it yielded density estimates for 12 different species, ranging from small to megafauna, that together formed the bulk of the mammal community in a single protected area. Most REM applications to date concern one or a few species rather than entire communities (Palencia et al., 2022;Schaus et al., 2020;Wearn et al., 2022), and smaller, more cryptic species in mammal communities are often lacking in density estimates. The application in BNP explores the effective use of REM in a wider range of body sizes in a mammal community with a novel camera orientation. ...
Article
Full-text available
The random encounter model (REM) estimates animal densities from camera‐trap data by correcting capture rates for a set of biological variables of the animals (average group size, speed and activity level) and characteristics of camera sensors. The REM has been widely used for setups in which cameras are mounted on trees or other structures aimed parallel to the ground. Here, we modify the REM formula to accommodate an alternative field of view acquired with vertically oriented camera traps, a type of deployment used to avoid camera theft and damage. We show how the calculations can be adapted to account for a different detection zone with minor modifications. We find that the effective detection area can be close to a rectangle with dimensions influenced by the properties of the Fresnel lens of the camera's motion sensor, the body mass of different species and the height of the camera. The other REM parameters remain the same. We tested the modified REM (vREM) by applying it to wildlife data collected with vertically oriented camera traps in Bardia National Park, Nepal. We further validated that the effective detection area for the camera model used was best approximated as a rectangle shape using maximum likelihood estimation. Density estimates obtained broadly matched independent density estimates for nine species from the previous studies in Bardia with varying body sizes by four orders of magnitude. We conclude that these modifications allow the REM to be effectively used for mammal density estimation for species with a wide range of body sizes, with vertically oriented camera traps.
... The REM was among the first methods developed for estimating density of unmarked wildlife populations, and rapid application of this approach has occurred since the seminal publication in 2008 (Appendix A, Figure A1A,B). However, conflicting results from both empirical and simulation-based studies have caused controversy over the reliability of REM density estimates and the usefulness of those estimates for informing conservation and management [33,[93][94][95]. Results from analysis of our multi-year empirical predator-prey dataset demonstrated that using detection data from strategically placed cameras can cause substantial inflation of REM density estimates, and that borrowing movement velocity (day range) values from other studies, time periods, or species can introduce volatility in REM density estimates. ...
... In contrast, we deployed 83 camera-traps continuously for multiple consecutive years to produce density estimates during~180-day seasonal periods within each year. A recent review of 34 studies that compared density estimates from the REM with density estimates from other methods (e.g., spatial mark-resight) found an average of~50% positive bias in REM densities when detection data from strategically placed cameras were used [95]. Our results indicated that the positive bias in REM density estimates from using data obtained at strategically placed cameras could be much more severe, likely depending on the behavior and ecology of the target species and characteristics of the survey design (e.g., number of cameras, survey duration, etc.). ...
... Our empirical study results also confirm the findings of Palencia et al. [52] regarding the consequences of borrowing movement velocity values. In a review of studies that compared density estimates from the REM with densities from other methods, a tendency was demonstrated for researchers to use underestimates of movement velocity that, in turn, caused overestimation of density by the REM [95]. In contrast, the recent study by Palencia et al. [52] that investigated the potential effects when non-survey-specific velocity values were used found both positive and negative bias in REM density estimates. ...
Article
Full-text available
Identifying population-level relationships between predators and their prey is often predicated on having reliable population estimates. Camera-trapping is effective for surveying terrestrial wildlife, but many species lack individually unique natural markings that are required for most abundance and density estimation methods. Analytical approaches have been developed for producing population estimates from camera-trap surveys of unmarked wildlife; however, most unmarked approaches have strict assumptions that can be cryptically violated by survey design characteristics, practitioner choice of input values, or species behavior and ecology. Using multi-year datasets from populations of an unmarked predator and its co-occurring unmarked prey, we evaluated the consequences of violating two requirements of the random encounter model (REM), one of the first developed unmarked methods. We also performed a systematic review of published REM studies, with an emphasis on predator–prey ecology studies. Empirical data analysis confirmed findings of recent research that using detections from non-randomly placed cameras (e.g., on trails) and/or borrowing movement velocity (day range) values caused volatility in density estimates. Notably, placing cameras strategically to detect the predator, as is often required to obtain sufficient sample sizes, resulted in substantial density estimate inflation for both the predator and prey species. Systematic review revealed that 91% of REM density estimates in published predator–prey ecology studies were obtained using camera-trap data or velocity values that did not meet REM requirements. We suggest considerable caution making conservation or management decisions using REM density estimates from predator–prey ecology studies.
... Advances in remote sensing technologies, automated classification methods and statistical models for observation processes have revolutionised the study of the natural world (Besson et al., 2022). Ecological sensors such as camera traps and autonomous recording units (ARUs) are now routinely deployed to monitor species' distribution (Chambert, Waddle, et al., 2018), abundance (Augustine et al., 2018;Palencia et al., 2022) and behaviour (Buxton et al., 2018;Caravaggi et al., 2017), and method-specific uncertainties are well-documented (Burton et al., 2015). A growing and widely applied suite of hierarchical models, particularly occupancy models, have been developed to draw ecological inferences from sensor data while accounting for false negative errors: the ubiquitous failure to detect a species or individual when present, and false positives: the less common, but still problematic, erroneous detection of a species or individual when absent (Kéry & Royle, 2020;Royle & Link, 2006). ...
Article
Full-text available
Artificial intelligence (AI) has revolutionised the process of identifying species and individuals in audio recordings and camera trap images. However, despite developments in sensor technology, machine learning and statistical methods, a general AI‐assisted data‐to‐inference pipeline has yet to emerge. We argue that this is, in part, due to a lack of clarity around several decisions in existing workflows, including: the choice of classifier used (e.g. semi‐ vs. fully automated); how classifier confidence scores are used and interpreted; and the availability and selection of appropriate statistical methods for drawing ecological inferences. Here, we attempt to conceptualise a general workflow associated with automated tools in ecology. We motivate this perspective using our experiences with occupancy modelling using monitoring data collected through passive acoustic monitoring and camera trapping, identifying priority areas for future developments. We offer an accessible guide to support the ecological community in navigating and capitalising on rapid technological and methodological advances. We describe how different error types arise from both sensor‐based monitoring and from classifiers themselves; how different error types are handled at each stage of the workflow; and finally, implications and opportunities associated with deciding on methods used at each step of the pipeline. We recommend that ‘black box’ tools like neural network classification algorithms should be embraced in ecology, but widespread uptake requires more formal integration of AI into the existing ecological inference workflows. Like ecological AI more broadly, however, successful development of new data‐to‐inference pipelines is a multidisciplinary endeavour that requires input from everyone invested in collecting, processing, analysing and using ecological monitoring data.
... This approach has been widely applied due to its practical advantages such as no need for species-specific study design. Cameras would be set in areas with low visibility or dense vegetation, obtaining precise estimates for a wide range of ungulate species, as shown by Palencia et al. (2022). ...
Article
Full-text available
Mountain environments pose challenges for accurately assessing the population size of ungulates, which may hamper confidence in using abundance estimates for decision-making. Thus, evaluations of the accuracy of methods using known-size populations are valuable. Here, we tested the accuracy of direct counts and distance sampling to assess the size of a fenced population of Iberian ibex (Capra pyrenaica) of known size. We evaluated the influence of observer expertise (beginners and experts) on the detection error of female and male ibexes and whether training improved accuracy. The ibex population comprised 18 individuals (ten males and eight females) living in a 17-ha enclosure in northeastern Spain. After 27 surveys, experts detected 16% more ibexes than beginners. Male ibexes were ~ 13% easier to detect than females, and experts were more accurate than beginners in sexing individuals. Additionally, the detection error in absolute counts was similar among beginners but different among experts (> 10%). Despite reducing detection error over subsequent surveys, under-detection was > 50% in all events (> 85% and > 67% for beginners and experts, respectively). Our data indicate a systematic underestimation of direct counts of our mountain ungulate populations and the contribution of expertise to improving accuracy. Our results show that managers assessing mountain ungulate populations for management purposes should consider alternative methods.
Preprint
Full-text available
1. Artificial Intelligence (AI) has revolutionised the process of identifying species and individuals in audio recordings and camera trap images. However, despite developments in sensor technology, machine learning, and statistical methods, a general AI-assisted data-to-inference pipeline has yet to emerge. 2. We argue that this is, in part, due to a lack of clarity around several decisions in existing workflows, including: the choice of classifier used (e.g., semi- vs. fully automated); how classifier confidence scores are used and interpreted; and the availability and selection of appropriate statistical methods for drawing ecological inferences. 3. Here, we attempt to conceptualise a general workflow associated with automated tools in ecology. We motivate this perspective using our experiences with occupancy modelling using monitoring data collected through passive acoustic monitoring and camera trapping, identifying priority areas for future developments. 4. We offer an accessible guide to support the ecological community in navigating and capitalising on rapid technological and methodological advances. We describe how different error types arise from both sensor-based monitoring and from classifiers themselves; how different error types are handled at each stage of the workflow; and finally, implications and opportunities associated with deciding on methods used at each step of the pipeline. 5. We recommend that black box tools like neural network classification algorithms should be embraced in ecology, but widespread uptake requires more formal integration of AI into the existing ecological inference workflows. Like ecological AI more broadly, however, successful development of new data-to-inference pipelines is a multidisciplinary endeavour that requires input from everyone invested in collecting, processing, analysing, and using ecological monitoring data.
Article
Full-text available
The European Observatory of Wildlife EOW, as part of the ENETWILD project, represents a collaborative network that has been operating since 2021 to develop and implement standardized protocols to obtain harmonized data on distribution and density of target mammal species. In so doing, the EOW aims at contributing to improving the quality of data that are available for wildlife management and risk assessment on a European scale. This report describes the activities carried out during the 2023 EOW campaign, which was joined by a total of 30 organizations who committed to collect data in 44 sites across 22 different countries. We present data on the distribution and density of three species – wild boar (Sus scrofa), European roe deer (Capreolus capreolus), and red fox (Vulpes vulpes) – obtained by implementing a camera trapping protocol and by fitting the random encounter model (REM) for density estimation. Camera‐trap images were processed using the Agouti platform and some of its tools specifically designed for the management of camera trapping projects. This includes the use of photogrammetry to obtain parameters for the REM directly from the sequences of images. A total of 24 EOW sites were monitored in past years as well, providing multiannual density estimates and population trends and highlighting an improvement in the precision of the estimates, related to the improved study design and protocol implementation. We also describe the activities of the 2024 campaign, carried out as part of ENETWILD 2.0, where big efforts were made to expand the network, focusing on sites at risk of African Swine Fever, with wild boar/pig interactions and containing wetlands, as potential hubs for Avian Influenza. This effort resulted in the engagement of 40 participants monitoring 64 study sites (27 countries), including 28 study sites located either in infected areas or < 100km from the ASF frontline, and 25 sites with wetland habitats. Furthermore, in at least 20 sites pig farming is practised either intensively, extensively or as backyard farming. Finally, synergies were established with other international initiatives related to wildlife monitoring and disease prevention, with the aim of sharing experiences and sustaining a transnational data collection and harmonization.
Preprint
Full-text available
Ecological corridors are essential for maintaining ecosystem functionality, allowing species movement between protected areas. In the Central Italian Apennines, five corridors have been identified to enhance habitat connectivity for the critically endangered Marsican brown bear (Ursus arctos marsicanus). This study focuses on two of these corridors to investigate their support of other mammal species populations. We collected data from camera traps over four months, and applied the Random Encounter Model to estimate the population densities of eight meso- and macro-mammal species. We compared the densities we estimated with those reported in the literature for different locations across Europe. The results indicated higher-than-average densities for several species compared to published data, especially for ungulates. These findings underscore the importance of Marsican bear corridors for a number of other mammals, as they provide important habitats for several of them. Effective management of these corridors, with a focus on reducing human disturbance and improving habitat connectivity, will be critical for the long-term survival of both the Marsican bear and its coexisting species.
Article
The diversity and abundance of vectors are essential parameters in the transmission dynamics of West Nile virus (WNV) between its avian reservoirs and clinically susceptible mammalian species. Knowing the determinants of vector abundance could be thus useful in preventing West Nile fever (WNF) cases and associated socio-economic impact. We designed a survey at the wildlife-livestock interface to test the hypothesis that variations in environmental favorability between anthropized and wild scenarios modulate WNV vector abundance and transmission risk. In a continental Mediterranean region where WNF has recently emerged, we selected nine sampling sites and allocated three areas to every site with a decreasing gradient of wildlife-livestock interaction: A1) a horse farm where interaction is maximal; A2) a zone of intermediate interaction 500-1000m from the farm; and A3) an entirely wild zone of low interaction 3-5 km from the farm. At a fortnightly frequency, we estimated mosquito abundance at each of the 27 study sites in May-December 2018 and April-July 2019. We estimated bird and mammal abundance, collected meteorological information, and characterized mosquito habitat at the site scale. Thereafter, we studied the determinants of Culex spp., Cx. pipiens sensu lato (s.l.) and Cx. theileri abundance by constructing negative binomial generalized linear mixed models. We identified 20 mosquito species, with a notable predominance of Culex spp. and, particularly, of Cx. pipiens s.l.. We found differences in the spatiotemporal distribution of Culex spp. abundance and confirmed our hypothesis by finding important effects of local environmental variations in abundance. The accumulated rainfall in fortnights 4 to 14 and the mean temperature of the two fortnights before sampling were positively and statistically significantly associated with the abundance of Cx. pipiens s.l. (Z=13.09, p<0.001, and Z=9.91, p<0. 001, respectively) and Culex spp. (Z=13.35, p<0.001, and Z=6.99, p<0.001, respectively), while the mean temperature of the two previous fortnights was a positive statistically significant predictor (Z=14.69, p<0.001) of the abundance of Cx. theileri. The farm environment was the most conducive predictor to hosting Culex spp. compared to wild settings. Our results indicate that continental Mediterranean environments are favorable for WNV circulation and maintenance, especially the environment of anthropized rural settings such as farms. These results will have an impact on the spatiotemporal risk prediction of WNF emergence in continental Mediterranean environments.
Article
Full-text available
Camera trapping is a widely used tool in wildlife research and conservation, and a plethora of makes and models of camera traps have emerged. However, insufficient attention has been paid to testing their performance, particularly under field conditions. In this study, we have comparatively tested five of the most frequently used makes of camera trap (Bushnell, KeepGuard, Ltl Acorn, Reconyx and Scoutguard) to identify the key factors behind their probability of detection (i.e. the probability that the camera successfully capturing a usable photograph of an animal passing through the field of view) and trigger speed (i.e. the time delay between the instant at which a motion is detected, and the time at which the picture is taken). We used 45 cameras (nine devices of each make) with infrared flash in a field experiment in which a continuous remote video was used in parallel (as a gold‐standard) to discover the animals that entered the camera trap detection zone. The period (day/night), distance between animals and cameras, model, species, deployment height and activation sensitivity were significantly related to the probability of detection. This probability was lower during the night than during the day. There was a greater probability of detecting a given species when the cameras were set at its shoulder height. The interaction between species and the distance between the animals and the cameras significantly affected the trigger speed, meaning that the closer the animals that entered the detection zone, the higher the trigger speed, with substantial differences among species. This was probably related by movement speed. In conclusion, this study shows differences in the performance of camera trap models and settings, signifying that caution is required when making direct comparisons among results obtained in different experiments, or when designing new ones. These results provide empirical guidelines for best practices in camera trapping and highlight the relevance of field experiments for testing the performance of the camera traps.
Article
Full-text available
Methods for estimating population densities of unmarked species using camera traps are still under development. One such method is called "Random Encounter Model (REM)" and, to our knowledge, has never been used to estimate densities of mountaindwelling ungulates. In this study, we tested the REM method to estimate the density of Balkan chamois ( Rupicapra r. balcanica ) in a Mediterranean habitat, Mountain Biokovo. To meet the assumptions of REM, we systematically placed 25 camera traps throughout the known range of the population (approximately 65 km 2 ) at the intersections of 2-km grid cells. Prior to data collection, population density was estimated by visual counts on sample plots in August 2020. Cameras were operational between July 2020 and September 2020 and active throughout the 24-hour period. All parameters for REM (i.e., average movement speed, angle and radius) were estimated using exclusively camera trap data. We obtained 279 independent events of chamois from 2503 camera trap days. The density estimate obtained by REM resulted to be 20.65 ± 5.27 ind. km -2 , slightly higher than the reference value obtained by visual counts: 17.33 ± 0.98 ind. km -2 . Other parameters required to calculate density were speed (1.62 km·day -1 ± 0.21), detection radius (5.56 m ± 0.21) and detection angle (1.16 + 0.05 radians). Therefore, REM has shown comparable results to visual counts and may have potential for estimating density of mountain ungulates, especially in rugged and inaccessible mountainous areas with low detectability where other approaches are inadequate or impossible.
Article
Full-text available
Background Camera traps present a valuable tool for monitoring animals but detect species imperfectly. Occupancy models are frequently used to address this, but it is unclear what spatial scale the data represent. Although individual cameras monitor animal activity within a small target window in front of the device, many practitioners use these data to infer animal presence over larger, vaguely-defined areas. Animal movement is generally presumed to link these scales, but fine-scale heterogeneity in animal space use could disrupt this relationship. Methods We deployed cameras at 10 m intervals across a 0.6 ha forest plot to create an unprecedentedly dense sensor array that allows us to compare animal detections at these two scales. Using time-stamped camera detections we reconstructed fine-scale movement paths of four mammal species and characterized (a) how well animal use of a single camera represented use of the surrounding plot, (b) how well cameras detected animals, and (c) how these processes affected overall detection probability, p. We used these observations to parameterize simulations that test the performance of occupancy models in realistic scenarios. Results We document two important aspects of animal movement and how it affects sampling with passive detectors. First, animal space use is heterogeneous at the camera-trap scale, and data from a single camera may poorly represent activity in its surroundings. Second, cameras frequently (14–71%) fail to record passing animals. Our simulations show how this heterogeneity can introduce unmodeled variation into detection probability, biasing occupancy estimates for species with low p. Conclusions Occupancy or population estimates with camera traps could be improved by increasing camera reliability to reduce missed detections, adding covariates to model heterogeneity in p, or increasing the area sampled by each camera through different sampling designs or technologies.
Article
Full-text available
Population density estimations are essential for wildlife management and conservation. Camera traps have become a promising cost‐effective tool, for which several methods have been described to estimate population density when individuals are unrecognizable (i.e. unmarked populations). However, comparative tests of their applicability and performance are scarce. Here, we have compared three methods based on camera traps to estimate population density without individual recognition: Random Encounter Model (REM), Random Encounter and Staying Time (REST) and Distance Sampling with camera traps (CT‐DS). Comparisons were carried out in terms of consistency with one another, precision and cost‐effectiveness. We considered six natural populations with a wide range of densities, and three species with different behavioural traits (red deer Cervus elaphus, wild boar Sus scrofa and red fox Vulpes vulpes). In three of these populations, we obtained independent density estimates as a reference. The densities estimated ranged from 0.23 individuals/km² (fox) to 34.87 individuals/km² (red deer). We did not find significant differences in terms of density values estimated by the three methods in five out of six populations, but REM has a tendency to generate higher average density values than REST and CT‐DS. Regarding the independents’ densities, REM results were not significantly different in any population, and REST and CT‐DS were significantly different in one population. The precision obtained was not significantly different between methods, with average coefficients of variation of 0.28 (REST), 0.36 (REM) and 0.42 (CT‐DS). The REST method required the lowest human effort. Synthesis and applications. Our results show that all of the methods examined can work well, with each having particular strengths and weaknesses. Broadly, Random Encounter and Staying Time (REST) could be recommended in scenarios of high abundance, Distance Sampling with camera traps (CT‐DS) in those of low abundance while Random Encounter Model (REM) can be recommended when camera trap performance is not optimal, as it can be applied with less risk of bias. This broadens the applicability of camera trapping for estimating densities of unmarked populations using information exclusively obtained from camera traps. This strengthens the case for scientifically based camera trapping as a cost‐effective method to provide reference estimates for wildlife managers, including within multi‐species monitoring programmes.
Article
Full-text available
The use of camera traps in ecology helps affordably address questions about the distribution and density of cryptic and mobile species. The random encounter model (REM) is a camera‐trap method that has been developed to estimate population densities using unmarked individuals. However, few studies have evaluated its reliability in the field, especially considering that this method relies on parameters obtained from collared animals (i.e., average speed, in km/h), which can be difficult to acquire at low cost and effort. Our objectives were to (1) assess the reliability of this camera‐trap method and (2) evaluate the influence of parameters coming from different populations on density estimates. We estimated a reference density of black bears (Ursus americanus) in Forillon National Park (Québec, Canada) using a spatial capture–recapture estimator based on hair‐snag stations. We calculated average speed using telemetry data acquired from four different bear populations located outside our study area and estimated densities using the REM. The reference density, determined with a Bayesian spatial capture–recapture model, was 2.87 individuals/10km² [95% CI: 2.41–3.45], which was slightly lower (although not significatively different) than the different densities estimated using REM (ranging from 4.06–5.38 bears/10km² depending on the average speed value used). Average speed values obtained from different populations had minor impacts on REM estimates when the difference in average speed between populations was low. Bias in speed values for slow‐moving species had more influence on REM density estimates than for fast‐moving species. We pointed out that a potential overestimation of density occurs when average speed is underestimated, that is, using GPS telemetry locations with large fix‐rate intervals. Our study suggests that REM could be an affordable alternative to conventional spatial capture–recapture, but highlights the need for further research to control for potential bias associated with speed values determined using GPS telemetry data.
Article
Full-text available
Camera‐trapping methods have been used to monitor movement and behavioural ecology parameters of wildlife. However, when considering movement behaviours to estimate DR is mandatory to include in the formulation the speed ratio, otherwise DR results will be biased. For instance, some wildlife populations present movement patterns characteristic of each behaviour (e.g. foraging or displacement between habitat patches), and further research is needed to integrate the behaviours in the estimation of movement parameters. In this respect, the day range (average daily distance travelled by an individual, DR) is a model parameter that relies on movement and behaviour. This study aims to provide a step forward concerning the use of camera‐trapping in movement and behavioural ecology. We describe a machine learning procedure to differentiate movement behaviours from camera‐trap data, and revisit the approach to consider different behaviours in the estimation of DR. Second, working within a simulated framework we tested the performance of three approaches to estimate DR: DROB (i.e. estimating DR without behavioural identification), DRTB (i.e. estimating DR by identifying behaviours manually and weighting each behaviour on the basis of the encounter rate obtained) and DRRB (i.e. estimating DR based on the classification of movement behaviours by a machine learning procedure and the ratio between speeds). Finally, we evaluated these approaches for 24 wild mammal species with different behavioural and ecological traits. The machine learning procedure to differentiate behaviours showed high accuracy (mean = 0.97). The DROB approach generated accurate results in scenarios with a speed‐ratio (fast relative to slow behaviours) lower than 10, and for scenarios in which the animals spend most of the activity period on the slow behaviour. However, when considering movement behaviours to estimate DR is mandatory to include in the formulation the speed ratio, otherwise the DR results will be biased. The new approach, DRRB, generated accurate results in all the scenarios. The results obtained from real populations were consistent with the simulations. In conclusion, the integration of behaviours and speed‐ratio in camera‐trap studies makes it possible to obtain unbiased DR. Speed‐ratio should be considered so that fast behaviour is not overrepresented. The procedures described in this work extend the applicability of camera‐trap‐based approaches in both movement and behavioural ecology.