Wildlife Picture Index: Implementation Manual

Wildlife Picture Index: Implementation Manual
wo r k i n g pa p e r n o . 3 9 J u n e 2 010
Wildlife Picture Index:
Implementation Manual Version 1.0
Tim O’Brien
WCS Working Paper No. 39
June 2010
IMPlementation Manual Version 1.0
By Tim O'Brien
Tim O'Brien
Senior Conservation Scientist, Conservation Support
Wildlife Conservation Society
2300 Southern Boulevard
Bronx, NY 10460
254 0 62 32751
[email protected]
WCS Working Papers: ISSN 1530-4426
Online posting: ISSN 1534-7389
Copies of WCS Working Papers are available from:
Wildlife Conservation Society
Global Conservation
2300 Southern Boulevard
Bronx, NY 10460-1099 USA
Suggested citation:
O’Brien, Tim. Wildlife Picture Index: Implementation Manual Version
1.0. WCS Working Papers No. 39, June 2010.
Cover image: Aardvark (Orycteropus afer) camera trap photograph © WCS
The contents of this paper are the sole property of the author, and cannot
be reproduced without permission of the author.
The Wildlife Conservation Society saves wildlife and wild places worldwide. We do so through science, global conservation, education, and the
management of the world’s largest system of urban wildlife parks, led by
the flagship Bronx Zoo. Together these activities change attitudes towards
nature and help people imagine wildlife and humans living in harmony.
WCS is committed to this mission because it is essential to the integrity of
life on Earth.
Established 10 years ago as the Living Landscapes Program, Conservation
Support works with WCS staff from around the world to develop and
deploy wildlife-focused tools and strategies that help to save wildlife and
wild places. Conservation Support provides technical assistance, analysis,
training and capacity building to help strengthen the practice of conservation, both within WCS and more broadly in the conservation community.
The Conservation Support Program works closely with WCS’s Regional,
Policy, Species and Global Health Programs, with Global Challenges (climate adaptation, extractive industries, livelihoods and health) and with
our Living Institutions to help target and prioritize technical assistance and
training across the organization.
The Zoological Society of London (ZSL) is a WCS partner organization
that hosted the development of the WPI.
The WCS Working Paper Series, produced through the WCS Institute, is
designed to share with the conservation and development communities
in a timely fashion information from the various settings where WCS
works. These Papers address issues that are of immediate importance to
helping conserve wildlife and wild lands either through offering new data
or analyses relevant to specific conservation settings, or through offering
new methods, approaches, or perspectives on rapidly evolving conservation issues. The findings, interpretations, and conclusions expressed in the
Papers are those of the author(s) and do not necessarily reflect the views
of the Wildlife Conservation Society. For a complete list of WCS Working
Papers, please see the end of this publication.
Wildlife Conservation Society | WORKING PAPER NO. 39
This protocol has been developed as a collaborative effort between the Wildlife
Conservation Society (WCS) and the Zoological Society of London (ZSL). It
relies heavily on development work for the Terrestrial Vertebrate (Camera Trap)
Monitoring Protocol produced by WCS and the Conservation International
Tropical Ecology Assessment and Monitoring (TEAM) Initiative. I have also relied heavily on camera trap documents by Scott Silver, Justina Ray and Phillipp
Henschel. I thank Jonathan Baillie for hosting me at ZSL, and for stimulating
discussions and encouragement, Jim Nichols for his comments on spatial sampling and distribution of camera points, and Jorge Ahumada for many discussions about how to develop implementation manuals and for writing the WPI
software program. I am especially grateful to Margaret Kinnaird for her insightful comments, and to Jonathan Baillie, Marcus Rowcliffe and James Reardon
for taking the time to read and comment on an earlier draft of the manual. Erika
Reuter provided superb support in layout and final edits. Finally, I thank Steve
Buckland and Rachel Fewster for permission to use their description of general
additive models in Appendix I.
Introduction and Justification…………………………………………………………… 1
State Variables, Indices and Estimators .………………………………………….. 2
Wildlife Picture Index………..................………………………………………………. 5
Sampling Design: Equipment, Effort, Time, Spacing…………………………… 14
Practicalities of Setting Camera Traps……………………………………….......... 18
Date & Time Settings ………………………….................................................... 20
Time Delays …………………………………………………………………............ 20
Setting the Camera Traps ………....................……………………………………. 20
Monitoring the Cameras ……………………………………………………………. 22
Bibliography …………………………………..................................................…….. 23
Appendix 1: Model Specification using Generalized Additive Models. 28
Appendix 2: R-Program to Calculate WPI, Confidence Intervals and
Change Points ……………………………………………................. 31
Wildlife Conservation Society | WORKING PAPER NO. 39
WPI: Implementation
Manual 1.0
Introduction and Justification
Worldwide, biodiversity is being lost at a rate comparable in magnitude only
to a handful of cataclysmic mass extinction events in the Earth’s geological
history (Pimm et al. 1995, Raffaelli 2004). Loss of biodiversity has major
implications for ecosystem health and function (Zavaleta and Hulvey 2004,
Solan et al. 2004), provision of goods and services (Hooper et al. 2005, OdlingSmee 2005), and the impoverishment of quality of life (Millenium Ecosystem
Assessment 2005). However, we possess few indicators capable of assessing
the extent or location of biodiversity loss on a global scale, and thus lack the
knowledge with which to respond to the underlying drivers of loss (Parrish et
al. 2003, Balmford et al. 2005, Mace and Baillie 2007).
One of the first lines of defense in the conservation of biodiversity is the
global network of legally mandated protected areas and wilderness areas
(Balmford et al. 2002, Rodrigues et al. 2004, deVries et al. 2005). They are critical for preserving natural habitats and wildlife communities and, in many cases,
offer the last remaining refuge for rare and/or threatened species. Although they
are essential for conserving biodiversity, there is currently little information on
where conservation areas are, and are not, stemming the tide of biodiversity
loss (Andelman and Willig 2003, DeVries et al. 2005, Joppa et al. 2008) and
this information is often contested (Bruner et al. 2001a,b, Vanclay 2001, Ervin
2003). Information on the effectiveness of parks is especially scarce in the tropical regions, where much of the world’s biodiversity resides. Such information is
needed to assess the overall status of biodiversity and to identify regions where
more resources are urgently required (Meir et al. 2004. Andam et al. 2008).
Such information also will help to assess management effectiveness.
In 2002, 188 signatory countries to the Convention on Biological Diversity
(CBD) committed themselves to “achieve by 2010 a significant reduction of the
current rate of biodiversity loss (emphasis added) at the global, regional and
national level” (Decision VI/26; CBD Strategic Plan). This ambitious target has
highlighted the lack of knowledge with which to assess biodiversity trends and
the need for effective biodiversity indicators to report national and global trends
for 2010 and beyond (Dobson 2005). At the Ninth Meeting of the Subsidiary
Body on Scientific Technical and Technological Advice (SBSTTA), seven focal
The WPI is
suitable for
monitoring the
component of
by mediumto large-sized
terrestrial forest
and savannah/
mammals and
areas were recommended for indicator development: (1) Status and trends of
the components of biological diversity; (2) Sustainable use; (3) Threats to biodiversity; (4) Ecosystem integrity and ecosystem goods and services; (5) Status
of traditional knowledge, innovations and practices; (6) Status of access and
benefit sharing; and (7) Status of resource transfers. The broad range of indicator focal areas highlighted the conceptual complexity of biodiversity and lack
of knowledge regarding biodiversity trends.
Monitoring change in biodiversity requires gathering data on many species, often of different taxonomic groups positioned at different trophic levels.
Because different groups of species require different sampling techniques, single
monitoring programs can only target components of biodiversity and often
these results are amalgamated into composite or headline indices (Gregory et
al. 2008). A large number of composite indices that combine information across
species have been proposed as indicators of components of biodiversity including the Living Planet Index (LPI; Loh et al. 2005), Biodiversity Intactness Index
(BII; Schole and Biggs 2005), Red List Index (RLI; Butchart et al. 2004, 2007),
and the Sampled Red List Indicator (SRLI; Baillie et al. 2008). Fundamental
problems beset these indicators to varying degrees, including assumptions
about initial conditions, the subjective nature of underlying species data, reliance on expert opinion, and use of secondary data from a variety of published
and unpublished sources, collected under a variety of methods and subject to
very different degrees of precision. While statistically robust indicators can be
designed (i.e. UK Wild Bird Index; Gregory et al. 2003, 2005), few have been
implemented on a regional or global level that rely on a sound underpinning of
coordinated and consistent data collection.
This manual presents a new biodiversity indicator, the Wildlife Picture Index
(WPI). The WPI combines camera trapping, a field technique that is rapidly gaining
acceptance and use throughout the world (O’Brien 2008, Rowcliffe and Carbone
2008), with occupancy analysis (MacKenzie et al. 2006) and generalized additive
models or GAMs (Hastie and Tibshirani 1990, Fewster et al. 2000). The WPI is
suitable for monitoring the component of biodiversity represented by medium- to
large-sized terrestrial forest and savannah/grassland mammals and birds.
Camera trapping offers a non-intrusive, low cost, and verifiable means of
sampling rare and elusive birds and mammals that might react to sampling
methods that require human presence. Already, camera trapping is a standard
tool in the study of large forest cats (tigers: Karanth and Nichols 1998; jaguars:
Silver et al. 2004; pumas: Kelly et al. 2008), applications to the study of birds
are increasing (O’Brien and Kinaird 2008), and applications to biodiversity are
just beginning (Tobler et al. 2008a, 2008b, O’Brien et al. in press). Just as the
availability of clear methodology and guidelines for study designs aided the
development of capture-recapture studies based on camera trapping (Karanth
and Nichols 1998, 2002), I hope that this manual will serve as a practical guide
to developing sampling designs and analytical approaches for species richness
surveys and biodiversity monitoring. Because of the charismatic appeal of camera trap photographs and the potential to monitor entire communities of medium- to large-sized terrestrial vertebrates, the WPI will be well-suited for reporting geared to audiences that include policymakers and the general public.
Wildlife Conservation Society | WORKING PAPER NO. 39
State Variables, Indices and Estimators
Yoccoz et al. (2001) emphasize the need to pay attention to three basic questions
when developing monitoring programs: (1) Why Monitor? (2) What should be
monitored? and (3) How should monitoring be carried out? With respect to ‘why
monitor’, programs to monitor biodiversity components arise for a number of
reasons and at a number of spatial scales. The Tropical Ecology and Assessment
Monitoring (TEAM) program aims to be a surveillance system that provides an
early warning for the impacts of climate change and deforestation on tropical
rainforest biodiversity at a global level (www.teamnetwork.org). Other programs (i.e. U.K. Breeding Bird Survey) look for trends in the avian component of
biodiversity at the regional level. The important part of planning a biodiversity
monitoring program is to have a clear idea a priori of the objectives of the monitoring program. Objectives may include better scientific understanding of factors
that affect the rate of change in biodiversity. When competing hypotheses can be
formulated and tested through manipulative experiments, we can gain powerful
insights into the dynamics of biodiversity change. For large mammals and birds,
community-level manipulations often are not possible. It is still possible, however, to gain insights from monitoring data when a priori hypotheses are used to
make comparisons among alternatives. Combining biodiversity monitoring with
management interventions, such as the renewed commitment to maintenance
of biodiversity within the world’s protected area system, may yield information
about the current state of biodiversity and the impact of management activities on biodiversity. This is particularly relevant to the primary objective of the
Convention on Biological Diversity. Achieving a significant reduction in the rate
of loss of biodiversity is unlikely to occur without major management interventions at sites around the world.
‘What to monitor’ follows from the monitoring program objectives.
Objectives should focus on state variables (density, occupancy), rate parameters
that characterize the system dynamics, and other variables that are believed to
influence the system dynamics. In biodiversity monitoring, the state variable
can be a measure of species richness, or some combination of ‘abundance and
diversity’ (Magurran 2004). The rate parameters may be extinction and colonization rates, or measures of change in overall species abundance (turnover).
Abundance can be measured directly (an estimate of numbers of animals or the
biomass of the species), or indirectly (a measure of occupancy for a species), as
long as detectability is incorporated (Pollock et al. 2002, Buckland et al. 2005).
Diversity indices then combine abundance and species richness in a number
of variations of weighted sums of relative abundance (Yoccoz et al. 2002,
Margurran 2004). The sampling design for a monitoring program obviously
will depend on the choice of biodiversity measures. Some monitoring programs
may rely on estimates of species richness and associated rate parameters (colonization, extinction, and turnover) and there are a number of unbiased maximum
likelihood estimators of species richness and relative species richness (proportion of potential species present) available (Bunge and Fitzpatrick 1993, Cam et
al. 2000, Boulinear et al. 1998, MacKenzie et al. 2006). More often, it is desirable to include some measure of abundance/biomass/occupancy in the diversity
measure, increasing the complexity of the monitoring program but providing
better information on the tradeoffs between species richness, species abundance
and species evenness, and a better understanding of system function.
‘How to monitor’ should follow best practices for sampling. There is a large
literature on biodiversity monitoring and species richness inventories. Much
of this literature is devoted to the ‘How’ question and the merits of indices
versus estimators of species abundance or richness. The ideal monitoring program would account for variation in detectability among species, over time,
and across space. It would also account for spatial variation and survey error.
Accounting for variation in detection is normally done by estimating the detection probability for a species at a time and at a site and correcting the count
statistic (number of observed individuals [Ci], number of observed occupied
sites [sd], number of observed species [Sobs]) by the estimate of detection probability, ^
p, where the ^ (hat) denotes an estimated value of p. Suppose we wish
to estimate species richness when species differ in detectability due to rareness,
nocturnal versus diurnal habit, and shyness. In this case the count statistic is the
number of observed species Sobs. The relationship between the total number of
species in the community and the Sobs can be written as:
E(Sobs) = Si pi (1)
where E(Sobs) is the expected value of a random variable, the observed sample
of species, and pi is the probability that one of Si species is detected and included
in the Sobs, or the proportion of species detected at i. Species richness can be
estimated as:
= obs
pˆ i
From Eq. 2 it is clear that the precision of the estimate of Si is a function of the
precision of the estimate of pi, as long as we can count observed species without
The ease with which count statistics can be collected and pi estimated varies widely for state variables of abundance, biomass, occupancy, and species
richness. Usually, it will be easier to collect data on occupancy and species richness than on abundance and biomass when working with mammals and birds.
Often, there is a temptation to use the count statistics directly as indices of the
variable of interest under the assumption that detection probabilities are either
equal or are constant over space and time (Conroy 1996). This is usually not a
good idea. Let λij measure the rate of change in species richness between time i
and time j. λij is calculated as the ratio of species richness, Sj/Si. The counts of
species, Sobs at times i and j, are used as indices and λij is estimated as:
ˆ S obs, j
λij =
S obs, i
The expected value of λ is estimated as:
E�S obs, j � S j p j
E�S obs,i � S i p i
Wildlife Conservation Society | WORKING PAPER NO. 39
The use of species counts as an index of rate of change in species richness is
only warranted when pi = pj. The violation of this assumption can have many
unintended consequences and makes interpretation of λij difficult or impossible.
Although an index usually has a smaller variance than a corresponding unbiased estimate based on maximum likelihood methods, the gain in precision is
offset by the unpredictable loss of accuracy. In short, when we monitor, do we
want precise metrics with unknown bias, or less precise but unbiased metrics?
Wildlife Picture Index
For the Wildlife Picture Index, I have followed the recommendation of Buckland
et al. (2005) and substituted occupancy for abundance as the state variable for a
community of terrestrial mammals and birds weighing more than one kilogram
(Box 1). I restrict the community to terrestrial species weighing at least one kilogram because smaller species of rodents and birds are not reliably detected in
camera traps. This is due, in part, to their small heat signature (camera traps are
triggered by heat and motion sensors) and, in part, to the fact that many small
mammals and birds are semi-terrestrial, and may be present but not detected
owing to vertical habitat gradients. More importantly, the larger mammals and
birds are well-described and represent the highest trophic levels in most communities (Dobson et al. 2006). This high-level community is composed of strong
interactors (Power et al. 1996) including top carnivores, ecosystem engineers,
large grazers and browsers, seed dispersers and seed predators. These are important components of terrestrial biodiversity because they are vulnerable to legal
and illegal consumption and exploitation (Pimm et al. 1988), and often are the
targets of wildlife management and eco-tourism (Ray 2005, Norton-Griffiths
2007). Because they tend to have large area requirements, they are susceptible
to extinction due to habitat loss (Purves et al. 2000). Species that occupy higher
trophic levels typically are lost more rapidly than species from lower trophic levels
as habitat quality and quantity decline (Dobson et al. 2006), and their loss is often
The larger
mammals and
birds are well
described and
represent the
highest trophic
levels in most
(Dobson et al.
2006); this highlevel community
is composed of
strong interactors
(Power et al.
BOX 1. Criteria for a Biodiversity Indicator
Buckland et al. (2005) suggest a set of criteria for a biodiversity measure when it is
used to assess changes over time. They assume that three aspects of biodiversity are
of primary interest: number of species, overall abundance, and species evenness. For
a group of similar species, abundance may be used. Biomass or occupancy may be
substituted when the species vary in size.
1. For a system that has a constant number of species, overall abundance and
species evenness, but with varying abundance of individual species, the index
should show no trend.
2. If overall abundance is decreasing, but number of species and species evenness
are constant, the index should decrease.
3. If species evenness is decreasing, but number of species and overall abundance
are constant, the index should decrease.
4. If number of species is decreasing, but overall abundance and species evenness
are constant, the index should decrease.
5. The index should have an estimator whose expected value is not a function of
sample size.
6. The estimator of the index should have good and measurable precision.
linked to trophic cascade and collapse (Terborgh et al. 2001, Pringle et al. 2007).
Dobson et al. (2006) argue that many ecosystem services result from activities of
species at specific trophic levels, and ecosystem services that rely on high trophic
level species are especially sensitive to small changes in biodiversity. Such services
include seed dispersal, browsing, predation on lower trophic levels, and ecotourism. Loss of upper trophic level species can have large indirect impacts several
levels lower, affecting the structure of plant communities, bird communities, and
water quality (Ripple and Beschta 2004, Hollenbeck and Ripple 2007). Dobson
et al. conclude that the status of species at higher trophic levels may serve as an
important indicator for maintenance of species and ecosystem services at lower
trophic levels, where services are more closely linked to human health and economic benefits. Using this logic, changes in WPI may provide an early warning
system for loss of lower trophic levels and associated ecosystem services.
Buckland et al. (2005; Box 1) evaluated five potential biodiversity measures
that might use abundance or occupancy data. They found that the geometric
mean of relative abundance (defined as abundance at time t divided by abundance at time 1) and a Shannon Index modified to fit Buckland et al.'s performance criteria were most satisfactory. I chose the geometric mean because the
modified Shannon Index had no theoretical justification other than fitting the
criteria. Note that the form of the geometric mean index anchors the index to
the value of the first abundance or occupancy estimate. This is considered to be
more efficient compared to the more usual index of xt+1 divided by xt because we
do not lose information when a year of surveys is missed (Fewster et al. 2000).
The geometric mean performs best if a small value is added to all observations
to remove zeroes from the dataset, and if there are not too many rare species in
the community. Rare species tend to inflate the variance estimates, but they are
a typical feature of most mammal and bird communities. The geometric mean
has several advantageous features (Limpert et al. 2001). First, it is useful for
averaging ratios when it is desirable to give each ratio equal weight (Zar 1999).
Second, because we are interested in rates of change of a group of species, the
geometric mean, unlike an arithmetic mean, tends to dampen the effect of very
high or low values, which might otherwise introduce bias. The geometric mean
thus can be used to develop the trend for a population of species (Gregory et
al. 2003). Geometric means can also be combined and scaled upward, making
it desirable for comparisons at regional and global scales. Composite indices
based on geometric means at a number of sites can be combined to generate a
regional index that, in turn, can be combined to generate a global geometric
mean (Collen et al. 2008, 2009).
To develop a WPI, we begin with occupancy estimates using data that are
typically collected during a camera trap study, photographic identifications of
species that can be assigned to specific days of a survey. Occupancy surveys
are relatively easy to carry out and to interpret. We start with the objective of
estimating the proportion of an area (actually a collection of sampling units)
that is inhabited by a target mammal or bird. The sampling units are camera
trap points at a site of interest. We assume that the points are selected to be
representative of the larger area for which we wish to make an inference (e.g. a
random or systematic sampling array). We assume that a species is not detected
at a site when it is absent (no false positives). The K surveys are conducted over
Wildlife Conservation Society | WORKING PAPER NO. 39
a period of time during which the population is assumed to be closed to changes
in state of occupancy. The period of population closure is considered a “season”
and, for most species of medium- to large-sized mammals and birds, population closure may be between one and five months. We then conduct K repeated
surveys within a season to establish the status of a species at each point using
camera traps. Each camera records a history of daily occurrence of each species
in the community at the sampling point within a season. Species status can take
3 states: present, absent, and present but not detected.
The definition of season as a period of population closure requires that we
be familiar with the behavior of all species within the community. Some species are territorial, some residential, some nomadic and some migratory. It is
therefore likely that not all species using an area of interest are present at any
given point or period in time. Careful consideration is required to ensure that
the ‘season’ of closure coincides with the time that the maximum number of
species occupies the area of interest, and avoids transition periods when species may be moving in and out of the area in an unpredictable, and possibly
nonrandom, manner.
Cameras operate for K days during which they record the presence or detection (designated as 1) and nondetection (designated as 0) for each day of the
survey. Each point i has a detection history for each species in the community
represented by a vector of K 1’s and 0’s that describe the detection history for the
species. A K=5-day survey at camera point i=1 might photograph species x on day
1 and day 5 but not on days 2, 3, and 4. This can be expressed as a detection history of [10001] for the 5-day period. Similar detection histories are accumulated
for each species at each camera point. For each species i, in year j at site k, we use
the species’ detection histories to estimate occupancy for that species.
It is unlikely that a target species will always be detected when present at a
point. This is especially true for camera traps because the sampled area is actually the field of view of a camera. In developing a model to estimate occupancy,
we can first consider the simplest case, a single species, single season occupancy
model with survey-specific detection probabilities (MacKenzie et al. 2006). In the
first of three situations, we can assume perfect detection (p=1) of the target species when it is present at a point, and that all points have the same probability of
occupancy, ψ1. The proportion of points occupied is number of points where the
target species is detected (sD) divided by the total of s random points:
ψ1 ==
sps *
Next, assume the target species is detected imperfectly and the probability of
detecting the species during a single survey of a point where the species occurs
is p, which is known exactly. The probability of detecting a species at least
once after K surveys is 1 minus the probability of never being detected during
K surveys, p* = 1 – (1 – p)K. The number of points where a species is detected
is again sD out of s random sites. The proportion of points occupied when p*
is known is:
ψ =
sp *
Eqs. 5 and 6 assume knowledge about p which is unlikely to exist. The models
we use in occupancy analysis therefore do not assume knowledge of p. Rather,
these models consider the likelihood of an observed outcome in a framework that
allows simultaneous estimation of occupancy and the associated detection parameters using maximum likelihood estimation (MacKenzie et al. 2006). The model
assumes that two processes affect the detection process at a sample point. First, a
point may be occupied by a target species with probability ψ, or unoccupied with
probability 1 – ψ. If the point is occupied, then there is some chance of detecting
the target species during a survey, pj, and a probability 1 - pj of not detecting the
species during a survey. Under this model, we can describe all possible outcomes
of K surveys as a set of detection histories in which each detection history has
an associated probability. For the detection history [10001], the likelihood of
this particular history (hi, where h symbolizes a vector of outcomes of surveys)
is described as Pr(hi = 10001) = ψp1(1-p2)(1-p3)(1-p4)p5. This translates to the
likelihood that the site was occupied by the target species and was detected the
first and last surveys during K=5 surveys. For the special case of the site being
occupied but the target species not detected we would have a detection history
reflecting no detections, hi = [00000]. The interpretation here is that either the
species was not present (1-ψ) or that it was occupied and the species was not
detected [ψ(1-p1)(1-p2)(1-p3)(1-p4)(1-p5)]. Because we cannot distinguish the
correct state, the likelihood incorporates both states as Pr(hi = 00000) = ψ(1-p1)
(1-p2)(1-p3)(1-p4)(1-p5) + (1- ψ). We use this approach to describe the detection
history (hi) for s points and K survey days in a model that describes the likelihood
that ψ and p occur given a series of s detection histories of length K:
L(ψ, p | h1 , h 2
,..., h s ) = ∏ Pr(h i )
i =1
Which describes the product of all possible outcomes of surveys, present and
detected, present but not detected, and absent:
s −s D
s −s ⎤ ⎡
(ψ,pp|| hh11 ,, h
Pr(h ii)� = ⎢ψsD ∏ p j j �1 − p j � D j ⎥ ⎢ψ∏ �1 ‐ p j � + �1 − ψ�⎥ (8)
h22 ,...,
,..., hhss�) = ∏ ���h
⎦ ⎣ j=1
where sD is the number of points where the target species was detected at least
once, and sj is the number of points where the species was detected during the jth
survey. The main assumptions for this model are: (1) the occupancy state of each
point is constant during the season (season closure); (2) the probability of occupancy is equal across all points; (3) detection of a species in each survey of a point
is independent of detection during other surveys at the point; and (4) detection
histories at each point are independent of other points. Often, a particular model
is used that assumes that detection is equal across all sites (all pi’s are the same).
We develop an occupancy estimate for each species in a community that
is detected during a season. A species that is present but not detected has an
occupancy estimate of zero for the season. The geometric mean is restricted to
values greater than 0, however, so the occupancy estimates must be adjusted to
eliminate 0-values. Adjustments terms are arbitrary, and I recommend that all
zero estimates of ψ be adjusted by:
ψ* = ψ +
Wildlife Conservation Society | WORKING PAPER NO. 39
for an occupancy estimate based on x camera trap points. This ensures a distribution of ψ values that is strictly non-zero, non-negative distribution and has minimal effect on the variance of the distribution. The next step is to develop an index
of relative occupancy for each species-specific occupancy estimate for species i at
site j in year k. We do this by dividing occupancy in year k by the estimated occupancy at the initial season, oijk = ψ ijk/ψ ij1. This creates a species-specific index
that measures the change in occupancy from initial conditions. The estimate for
k = 1 is always 1. The WPI for year k and site j and n species is geometric mean
of scaled occupancy statistics for n species:
WPI = n o
Or equivalently,
WPI jk = exp⎜ ∑ log(o ijk )⎟
⎝ n i =1
This formulation has several advantages. First, it possesses most of the favorable characteristics of a biodiversity index outlined by Buckland et al. (2005; see
Box 1). Second, it is intuitively understandable (it behaves like a stock exchange
index). Third, it allows for easy dissection and development of associated indices
that track subsets of the community. For instance, it would be relatively straightforward to develop a bushmeat index by restricting the analysis to those species
at a site that are harvested for food. Fourth, the index is insensitive to species-specific variation in abundance and occupancy, because each species is scaled before
entering the site index. Finally, by scaling to the initial year, the ratio is robust to
missing years of data. Most ratio estimators require evenly spaced observations
because ratios are calculated sequentially, a process called chaining. The proposed
index does not depend on chaining as all estimates are calculated based on the
temporal distance from the initial condition (Fewster et al. 2000).
A problem that will often arise is that of a species being missed initially and
then detected after the first y seasons, due to sampling error or colonization. The
problem of missed species occurring in later surveys has two solutions. The first
is to re-calculate the WPI as new species are acquired, as is done with the LPI
(Collen et al. 2009); the second is to develop an index based on a regional species
list of expected species with all species occupancies adjusted by a constant. Species
not detected in the first survey are given the minimum value for the expected community. I recommend that the index be re-calculated as new species are added to
the community as this avoids biasing the index with species that are undetected
and, in fact, extinct in the community. For species that ‘colonize’ the community,
their pre-detection occupancy values are set to ψ*.
A second situation concerns rare species. Rare species are characterized by
restricted occurrence and/or detection probabilities close to zero. For these species,
unbiased occupancy estimates may be difficult to achieve using maximum likelihood methods (MacKenzie and Royle 2005, MacKenzie et al. 2006). In general,
increasing the number of sampling occasions and number of sampling points will
increase the accuracy and precision of occupancy estimates used in the WPI. For
species with detection probabilities < 0.02, accuracy and precision may decline
substantially, even with 100 sample points and 30 days of sampling (Figure 1).
Estimated Occupancy
P = 0.02
Estimated ψ = True ψ = Tru
P = 0.03
P = 0.04
True Occupancy
Figure 1. Change in bias of estimated occupancy for species with low detectability
as true occupancy declines. Sampling based on 100 camera points surveyed over 30
days and 500 simulations per run.
Figure 1 illustrates how bias increases as detection and true occupancy (expressed
as a percentage) decline for a species.
For species with p = 0.04, the estimated occupancy is 1% – 2% greater than
true occupancy and estimated occupancy accurately reflects the declining trend
in true occupancy. For species with p = 0.03, estimated occupancy bias increases
from 2% to 9% as true occupancy declines but the estimated occupancy still
tracks the trend reasonably well. At p = 0.02, we see large discrepancies in the
estimates and poor tracking of the trend in true occupancy.
When conducting a biodiversity survey that includes rare species, I recommend
that the investigator evaluate the impact of rarity and low detectability on occupancy estimates using the simulation functions in PRESENCE. Once the level of
sampling (number of points and number of days) are determined, the simulation
is simple. For a given true level of occupancy, detection probability, number of
Figure 2. Location of simulation function in PRESENCE Software. Enter PRESENCE
and select drop down menu for Tools, then select Simulation.
Wildlife Conservation Society | WORKING PAPER NO. 39
sites and number of replications, PRESENCE simulations (Figure 2) can calculate
the expected observed occupancy, estimated occupancy and standard error. One
simply varies the detection probability and true occupancy to evaluate the point
at which bias becomes unacceptable. Program failure is easily recognized; either
the program fails to give an estimate of occupancy or it generates an estimate
approaching 100% occupancy, because as detection probability approaches 0, an
occurrence at a one or a few points is vastly inflated. When this situation arises,
there are four possible alternatives for generating occupancy estimates. First,
we can assume that detection probability does not change over time, estimate a
single detection probability using a multi-year data set and apply this detection
probability to the individual datasets. Second, we can assume that closely related
species share detectability and develop detection probabilities for species complexes that can be applied to rare members of the complex. Third, we can apply
constant detection over time to a species complex of rare species and post-stratify
to estimate occupancy for individual species. Finally, we can use the observed
occupancy as the best estimate of true occupancy for rare species.
The choice of methods to deal with rare and cryptic species depends on the nature
of the species community. The first 3 strategies are all reasonable approaches to
avoiding misleading inferences at low detection and low occupancy. Substituting
observed occupancy for estimated occupancy is a bit more complicated (Figure
3). In Figure 3 we see that, even at low occupancy and low detectability, biased
Estimated Occupancy
True Occupancy
Figure 3. Estimated occupancy versus true occupancy for a range of detection probabilities (p) that describe cryptic species, based on 100 points sampled for 30 days
and 500 simulations per run. For species with p < 0.03, bias in observed occupancy
is less than bias in estimated occupancy between true occupancy values of 0.015 and
0.010. For species with 0.03 < p <0.05, bias in observed occupancy is less than bias
in estimated occupancy between true occupancy values of 0.010 and 0.05. For species with p > 0.05, bias in expected occupancy is always less than bias in observed
occupancy. Arrows indicate the range of occupancy where the trend reverses for most
cryptic (dashed lines) and less cryptic (solid line) species.
estimates will accurately track the species trajectory up to a point. For very cryptic
species, at values of occupancy between 0.15 and 0.10, the trend reverses and
estimated occupancy tends upward. This will lead to an incorrect inference. For
cryptic species with detection probabilities between 0.03 and 0.05, the downward
trend reverses between occupancy values of 0.10 and 0.05. For species with detection probabilities of 0.05 and greater, the estimated trend tracks the real trend
throughout. If one considers substituting the observed occupancy for estimated
occupancy, one should be aware that this also will create a bias in the trend below
that of the true trend. Based on the simulations above, little is gained by substituting observed occupancies for estimated occupancies when detection probabilities
are 0.04 or greater. At this point, we trade a positive bias in trend for a negative
bias in trend. In monitoring programs, guarding against type II error (failing to
detect a real trend) usually is more important than guarding against type I error
(detecting a trend when none exists). Under this precautionary principle, one
should consider the option of substituting observed for estimated occupancy values for rare and cryptic species only after careful evaluation of the situation.
To determine trends in the WPI, we follow Fewster et al. (2000) and Buckland
et al. (2005). They recommend using generalized additive models (GAMs) to
model trends as a smooth nonlinear function of time (Hastie and Tibshirani 1990).
GAMs are similar to regressions but they do not require that the data be normally
distributed and they assume that the relationship between the index and time is
smooth but not linear. GAMs are useful because they incorporate smoothing procedures into the model fitting process, allow a range of curves to be considered,
and allow for direct incorporation of co-variates to test hypotheses of factors
influencing trends. GAMs also allow for a statistical test of changes in direction of
the index trajectory, thus satisfying the criteria of a CBD 2010 indicator.
A simple regression model has the structure y = α + βx + ε with the assumption that the error ε is normally distributed. Ter Braak et al. (1994) used a loglinear Poisson regression model to fit count data of birds. They assumed that an
observation yit at site i and time t comes from a Poisson distribution with mean
μit. Their model resembles a linear regression:
it������i ���t
where αi is called the site effect for site i and βt refers to the year effect for year t.
Both the normal linear regression and the log-linear Poisson regression model are
considered types of general additive models. In a generalized additive model:
i ����������
The error ε is not assumed to be normally distributed, and the f(t) is some nonlinear smoothing function of time. The form of the predictor function f(t) is the
principle difference between a GAM and a generalized linear model. The GAM
is fitted by estimating the parameters αi and the smooth function f in the same
way that a linear regression is fitted by estimating the parameters α and β. For
a linear trend over time (substitute t for x), f(t) = βt has a single parameter β to
be estimated. For an annual model, f(t) = βt. In this case the function is jagged
and represented by joining β’s with straight lines. Between these two limits are
Wildlife Conservation Society | WORKING PAPER NO. 39
functions f that are nonlinear, smoother than the annual model, and of greater
utility for detecting long-term, nonlinear trends. Fewster et al. (2000) provide
an excellent summary of the relationship between generalized linear models and
GAMs; Appendix I summarizes relevant sections of Fewster et al. (2000).
Before the function f can be estimated, the level of smoothing must be
specified. The degree of smoothing is flexible and controlled by the degrees of
freedom in the time series dataset, ranging from a linear trend (df = 1) to an
unsmoothed trend representing the annual change during t years (df = t – 1).
Between these two extremes, the function f is determined nonparametrically
from the data. GAMs thus allow us to explore linear trends in short time series
and more complicated nonlinear trends as t increases. The choice of df-value
is an important part of the modeling process and depends on the objectives
of a particular analysis and length of the time series (Appendix I). GAMs are
used to separate underlying trends from short-term fluctuations (noise in the
data), but the point at which this occurs is subjective and may vary depending
on the objectives of the analysis. For long-term trends, a smooth index curve
is desirable and df should be set low. If information about annual fluctuation
is required, the index should be set at t – 1 to produce a curve of maximum
fluctuations. The length of the time series is also important; it will be harder
to detect nonlinear trends in short time series. Fewster et al. (2000) suggest
that a df of 0.3t be used for long time series, but caution against setting rules
for model selection and advise plotting indices from GAMs with a range of df
values before settling on a final value.
The 95% confidence limits for the GAM trend are determined by a nonparametric bootstrap process. To develop a bootstrap confidence interval, we
first select a random sample with replacement from the species that make up
the sample for a specific time point. We repeat this process 999 times. We then
analyse each sample as if it had been our real data. The variation in estimates
of the index among bootstrap samples should give a good guide to the variation
we would expect if we could take new samples of the community. The standard
deviation of samples estimates the standard error of our index. If we take the
999 bootstrap estimates for each year in the time series, and order each bootstrap sample from smallest to largest, the 25th smallest and 25th largest estimates
represent the lower and upper 2.5% quantiles and are approximate 95% confidence limits for the index at each point in the time series.
The rate of change in diversity is measured by the slope of the smoothed trend.
Nonlinear trends allow for changes in the rate of change over time. Changes in
the rate of change (a benchmark of Convention on Biological Diversity 2010
indicators) are measured by the deriving the curve of the second derivative of
the trend and the bootstrapped 95% confidence interval around the second
derivative. If, in a given year t, the confidence interval does not include 0,
then we have evidence that the rate of change is changing. The sign (+/-) of the
confidence interval indicates the direction of the change. In principle, a crude
approximation of the second derivative of the slope at time t can be obtained
using three points and the equation:
Dt ����t‐1 ‐ ���t�����t�1
where Dt is the second derivative evaluated at time t and I is the smoothed index
value at t – 1, t and t + 1. If the time series is lengthy, a more precise second derivative can be estimated using the index value at t – 2, t – 1 , t, t + 1 and t + 2 (S.
Buckland pers. comm.):
D ��2�� ��� 1�� ��� 2�� ��� 1�� ���2�� �
A negative Dt indicates the rate of decline is accelerating and a positive Dt indicates the rate of decline is slowing. To test the significance of the Dt value, we
use the bootstrap resamples above to set the confidence interval for the measure
of change. If in a given year, the confidence interval does not include zero, then
we have evidence that the rate of change is changing. If the interval includes
only negative values, the change is for the worse; if the interval includes only
positive values, the change is for the better.
All procedures for implementing an Occupancy analysis are available in the
free software package PRESENCE (www.mbr-pwrc.usgs.gov/software/presence),
GAM modeling software are available in the mcgv software package (Wood
2006) in R (Check the R-website, www.r-project.org/, for the latest version by
the R Development Core Team). Rachel Fewster provides GAM modeling software for monitoring of wildlife populations on her website (www.stat.auckland.
ac.nz/~fewster/gams/R/). Jorge Ahumada (Technical Director of CI/TEAM) has
written a program in R to calculate the WPI, the bootstrap confidence intervals,
and the significance of changes in slopes (Appendix II).
Sampling Design: Equipment, Effort, Time, Spacing
Years of work in community ecology have taught us that the number of species
detected is related to the area sampled and the sampling effort. Larger areas
tend to have more species and as sampling effort increases, the number of rare
species detected increases (Table 1). The size of the study site is an important
consideration for estimating alpha diversity, since the area sampled should
adequately represent the area used by the community of interest, including
rare species (Buckland et al. 2005). As a general rule, sample points should be
randomly or systematically assigned, and sampling should be sufficient at each
point to provide a reasonable chance of detecting a species if it is present. It is
important to keep the sampling quadrats of equal size for comparability.
For camera trapping studies, the sampling quadrats are analogous to the camera trap points and the quadrat size is measured as the area in which all individuals
have a chance of being detected. The larger the sampling area, the more likely a
species will be detected in the field of view. The sampling sensitivity or the ability of a camera to capture a species that is in the field of view is determined by a
combination of the field of view of the sensor (or detection zone), the distance that
the sensor and camera can trigger, the trigger speed of the shutter, and, at night,
the strength of the flash. Camera trap detection zones range from 345 ft2 to 4,185
ft2. Perhaps a more useful standard of comparison among camera traps is the field
of view measured at 30 ft from the camera. Here we find that most camera traps
have a field of view either in the range of 3 – 6 ft or they jump to a width of 26
ft. Clearly a camera with a wide field of view and a strong sensor will have a large
Wildlife Conservation Society | WORKING PAPER NO. 39
detection zone. Trigger speeds range from 0.15 sec to 4.52 sec. Since most camera
trap sensors reach 50 feet or more, one can compensate for slower trigger speeds
by setting the camera unit further away from the central target. Be aware of the
range of your flash, however, since the camera sensor will trigger beyond the range
of the flash but at night you will only record eyeshines. Also, be sure that if you
use a camera with a slow shutter speed, the unit compensates with a strong sensor
and wide detection zone. Otherwise you will miss the animal as it passes through
the field of view before the camera fires. It is necessary to experiment with your
camera traps to understand exactly how they function. The Web site www.trailcompro.com provides reviews of commercially available camera traps that include
cost, speed, sensor width and distance, flash range, and other useful information.
Table 1. Sampling effort, number of terrestrial forest mammals detected and proportion
of community considered rare, based on a detection rate less than 1 photo/1000 trapdays. Data are unpublished camera trap results from WCS Asia Program.
Rare Mammals
10 (35%)
9 (30%)
15 (43%)
20 (49%)
18 (45%)
22 (59%)
For camera trap sampling, there are few examples of statistically defendable sampling designs that follow accepted rules of randomized, systematic and stratified
spatial sampling. For capture-recapture sampling designs that focus on single species, guidelines include camera trap placement that ensures no individual territory
can fit between camera traps. Good sampling also considers adequate spatial coverage for the species in question and trap placement that maximizes detectability
of the target species (Karanth et al. 2008). This design is easy to generalize to
single species occupancy surveys, but difficult to generalize to multi-species surveys when species differ in ranging patterns, unless you assume that the trapping
design be appropriate for the species with the smallest home ranges. The use of
trapping webs (Anderson et al. 1983, Buckland et al. 1993) and Spatially Explicit
Capture-Recapture designs (SECR; Efford 2004, Borcher and Efford 2007) relaxes
requirements of spacing because the analysis uses distances moved between captures to calculate detection and effective sampling area. There are no examples in
the literature, however, of multi-species SECR surveys for mammals although at
least one such survey is underway (O’Brien and Kinnaird unpubl. data). For multispecies sampling, especially at the regional level or higher, monitoring programs
should be designed to take account of spatial variation (Yoccoz et al. 2001, Pollock
et al. 2002).
Spatial variation arises at two levels in regional biodiversity monitoring programs. First, there is between-site variation in biodiversity due to habitat, topography, climate and anthropogenic disturbance. In order to capture that variation,
the site selection should be representative of the region. Nonrandom selection
of sampling sites is a common feature of long-term wildlife studies (Why do we
The Web site www.
provides reviews
of commercially
available camera
traps that include
cost, speed, sensor
width and distance,
flash range, and
other useful
selection of
sampling sites is
a common feature
of long-term
wildlife studies
and can lead to
faulty inferences.
work where we work? Usually because of abundant wildlife!) and can lead to
faulty inferences. For instance, a program based on monitoring wildlife in wellmanaged national parks may not tell us much about the region where those parks
are located. However, many considerations lead us to make decisions about where
to locate a sampling site, especially in the tropics. Given the practical limitations,
non-random site location will continue to plague us and research into estimating
regional trends when sample sites are non-random will continue to be an area of
interest (Buckland et al. 2005). Within a sample site, there is also variation due to
local habitat, micro-climatic, micro-topographic and anthropogenic disturbance.
We usually have more control over sampling allocation within a site as opposed
to between sites. A defendable sampling design should employ randomized or
systematic assignment of sampling points to ensure representative coverage.
Stratification by habitat or elevation may also be appropriate within a site.
Area of coverage at a sample site should ideally be determined by the distribution of species in the community to be monitored; to adequately represent
wide-ranging species, the area of coverage must be sufficiently large. Species with
ranges that cover 10 – 50 km2 will require sampling at the level of hundreds
of square kilometers. I propose sampling units of 200 km2 as a general rule of
thumb. This size allows for the spatial coverage necessary to sample most large,
wide-ranging species such as large cats and elephants. It is also a logistically feasible area to cover within the constraints of a single season, limited resources and
under difficult field conditions.
Sampling intensity is usually measured in trapdays, a combination of number of traps deployed and number of 24-hour periods of sampling. Camera
trap sampling of rare species may require several thousand trapdays to develop
an adequate number of encounters for analysis. To help plan a camera trap
design for WPI based on occupancy analyses, I considered the tradeoffs between
detection probability of a species, the expected area of occupancy, the number
of days in a trapping season, and the number of points required to achieve
relatively unbiased and precise occupancy estimates at the level of species. I
assumed that the species in a community were a mix of rare species (detected
after a minimum of 1,000 trapdays), common species, widely occurring species
and spatially restricted species (only occurring at a few points in the sample).
Since a composite index based on occupancy estimates inherits the bias and
uncertainty of the species estimates that comprise the index, I looked for a
sampling strategy that produces the most robust species occupancy estimates in
terms of accuracy, precision and cost.
I considered a species with a range of detection probabilities (0.02, 0.03,
0.04, 0.06, 0.08, and 0.10), representative of the range of uncommonly to rarely
encountered species. I considered a true occupancy of 10% to 60% of the sample
points in 10% intervals. I then evaluated the limits of reliable detection for a
range of sampling intensities (60 camera points to 100 camera points and 30 days
of sampling/point). I considered an estimate as reliable when its bias was less than
10% of true value and its coefficient of variation (CV) was 20% or lower.
Table 2 shows that we can achieve acceptable accuracy using 60 camera
points at a detection probability of 0.03 only for those species with 60% occupancy or greater. As we attempt to monitor species with lower detection probabilities and more restricted distribution, more camera points are required to
Wildlife Conservation Society | WORKING PAPER NO. 39
accurately estimate occupancy within 30 days. Even 120 trap points are insufficient to provide unbiased estimators when a species has a detection probability
of 0.02 and occupancy of 20%.
Table 2. Number of trap points operated for 30 days required to minimize bias (bias
< 10%) for a range of detection probabilities, P.
Table 3 shows the effort required to gain acceptable precision for a range of
detection probabilities and occupancies. It is much harder to increase precision
(reduce CV) using 60 trap points. Even 100 trap points produce imprecise estimates for those species with the lowest detection probabilities.
Table 3. Number of trap points operated for 30 days required to minimize CV (CV <
20%) for a range of detection probabilities, P.
Ideally, we would like to have precise and accurate occupancy estimates to enter
into the WPI for all classes of detection and occupancy. Unfortunately, achieving
the last 5% of gain can be prohibitively expensive. As a compromise between
the time required to trap a large number of trap points, the cost of camera traps,
cost of deployment in the field, and the large area to be covered, I recommend
that 100 camera points across 200 km2, or 1 camera per 2 km2, be considered
adequate coverage for a WPI survey. This is an arbitrary decision, but experience in many WCS sites and in implementing WPI monitoring for the TEAM
Program suggests that this is a feasible target that can be completed within 3-4
months, a reasonable length of time to consider a closed season for medium and
large mammals and birds. Refining the understanding of the sampling requirements for precise and unbiased WPI estimates is the topic of current analysis.
WPI surveys should be completed at each site on an annual basis. The time
of year in which surveys are conducted is a site-level decision. The choice of
season should be left up to the site managers for the camera trap protocol but,
once a season is chosen, it should not change in future years. The deployment
of cameras should be kept consistent (same season, same locations) over time
at each site in order to control for seasonally-regulated influences on animal
behavior, occupancy or abundance.
Strategies for camera trap deployment are difficult to prescribe so I will only
give suggestions that have worked for researchers in the past. Most monitoring
programs can afford to purchase 30 to 50 cameras at a time. Often, climatic conditions, theft, and wildlife damage can all take a toll on camera traps. Hot wet
climates require camera traps that can withstand the weather, whereas in tropical
savannah climates, less durable units may be used. Locks and theft-proof boxes
can add expense to a monitoring program, but are a wise investment in humandominated landscapes where theft of camera traps can be a problem.
I have envisioned that a researcher wishing to conduct a WPI survey would
be able to deploy 33 – 35 cameras for a month in three sampling blocks
totaling ~ 200 km2. The camera trap deployment rests on the assumption
that each site will deploy 30-35 operating cameras during a sampling period
(30 days) and that the cameras will not be visited until they are ready to be
moved. The precise shape of each sample is dependent, to some extent, on
landscape features and access, and the initial deployment of traps should be
determined with GIS prior to going to the field (see Figure 4). I find it useful
to determine the area to be trapped, and overlay a grid of the desired area to
be sampled by a single point, generate centroid points for the grid and use
these as the starting points for the sampling design. This can easily be done
using ArcView or ArcGIS. Samples should be oriented along a gradient from
disturbance to pristine conditions. For some sites, this will mean that the edge
of disturbance is directly adjacent to the sample blocks. Other sites will have
a buffer of undisturbed habitat before encountering an edge of disturbance.
The spacing between cameras is sufficient to ensure that the sampling occurs
at the level of habitat use by most or all of the largest mammals and birds in
the community.
Practicalities of Setting Camera Traps
Once the initial trapping design is established, the cameras should be deployed.
The deployment team should use GPS to navigate to specific trap point coordinates and, once the deployment team reaches the sample point, they will need to
find the best possible location as close as possible to the predetermined coordinates, preferably within 50 m but possibly within 100 m. The exact site is chosen
to give the highest probability of obtaining useful photographs of a range of species, usually a game trail. The goal is to photograph as many species as possible.
Although different species have different travel habits, and trail characteristics
may affect the species that use those trails, local knowledge of the situation on the
ground should assist in making the decision. Once the final location is chosen, the
leader of the camera trap deployment team should record the new longitude and
latitude coordinates of the final placement of the camera trap using a GPS unit.
This serves as the permanent location of the camera trap point in all subsequent
Wildlife Conservation Society | WORKING PAPER NO. 39
It is difficult to give unambiguous recommendations for choosing the ideal sample
point for a global, community-level monitoring program. Choice will depend on
the habitat and animal community under consideration. A few tips that have been
suggested in the past include:
• Pick a site where the travel path is restricted to the area that can be photographed by the camera. For example, a good location to place a camera trap
could be a place where there is a good deal of wildlife sign or an intersection
of several trails. A single trail with evidence of wildlife use and limited travel
alternatives is optimal for placing cameras. The maximum trail width should
be less than the flash distance; we recommend no more than 15-20 ft.
• The ground and slope under the sensor beam needs to be reasonably level.
Trails with ruts or holes in front of the camera may inhibit use, especially if
they fill with water after rains. Slopes can result in the ground obscuring animals from the sensor beams. A pronounced slope on one side of the path may
result in a sensor beam that is at shoulder height of large mammals but over
the heads of smaller animals on the down slope. Be aware of all the possibilities of travel in front of the cameras. The best way to do this is to test cameras
for the ability to detect animals with a shoulder height of 20-50 cm.
The goal is to
photograph as
many species as
Figure 4. An example of a basic design for deploying camera trap across a 200 km2
landscape in a systematic grid that ensures a spacing of 1 camera per 2 km2. Points
represent camera trap locations. Dark line is boundary of landscape, and light lines
represent a road system.
Increasingly, camera trap projects are converting to digital camera traps. It is
critically important that projects organize data management in a careful manner.
I recommend a catalogue of directories that begin at the site level, blocks within
sites, points within blocks, cameras within points, discs within cameras and photographs within discs. The TEAM network has produced a terrestrial vertebrate
monitoring protocol that provides detailed advice for camera deployment and
data management (available for download at www.teamnetwork.org)
Date & Time Settings
Photographs without an accurate date and time stamp are practically useless. The
date on the photograph is essential for determining the individual capture event
for occupancy analysis. Each 24-hour period is considered one of 30 sampling
periods so that all pictures of an individual photographed on the same date occur
within a single capture period. If you wish to use a filter to determine which
photographs constitute independent events (O’Brien et al. 2003), then the time
stamp can be used to distinguish adjacent film frames taken 1 minute apart versus
1 hour apart. While camera models may differ slightly in setting the time/date
stamp the important consideration is that it is consistent among all cameras in
the monitoring program. Digital cameras allow the option of multiple frames per
trigger event. This may be useful to help in identification of species.
Time Delays
All camera traps can be programmed with a delay between successive pictures.
This is important as group-living species or animals that linger in front of the
camera can result in many wasted pictures, and more importantly, fill the memory
card or deplete the battery before the sampling period is finished. A non-functioning camera creates a data gap in the survey design that may result in the loss of
data. The delay setting should be based upon the likelihood of encountering large
groups of non-target animals: experimentation during the pilot study period will
assist in selecting the length of the delay setting for your study site. Because a longer delay increases the probability of missing a capture, the rule of thumb should
be to use the minimum length of delay you feel comfortable with. For instance,
a 2 GB flashcard can store 3,000 images or more so the trigger interval is not a
large concern so long as you do not have false triggers due to moving vegetation.
For the WPI, a setting of 1 minute between triggers is appropriate to detect species. This will reduce the number of photos of group-living species passing by the
camera, yet allows classification of independent photo events.
Setting the Camera Traps
Once all these factors have been considered and the optimal sample point location is determined, the camera traps must be set. Find a location where there is
a suitable tree or insert a post at an optimal site. Suitable trees have trunks that
are reasonably straight, thin enough to tie a chain or wire around, but not so
thin that wind, people or other animals can shake it excessively. Try to minimize
direct sunlight on the cameras as excessive heat can reduce the sensitivity of the
sensors to endothermic animals. It is important to avoid setting a camera facing east or west as the sunset and sunrise may cause glare on the photograph.
Cameras should be set back at least two meters from the nearest point where an
Wildlife Conservation Society | WORKING PAPER NO. 39
animal might travel across the sensor. This allows for clear, focused pictures and
a large field of detection from the sensor. The longer an animal is in the detection
zone, the less chance of missing a photograph. Because the sensor beam should
be approximately shoulder-high on the average target species, the camera should
be set approximately 30-50 centimeters off the ground and parallel to it. Once a
camera is positioned, the details of positioning should be recorded and referred to
in subsequent sampling. The camera should be mounted to face perpendicular to
the trail. Use pliable, light gauge wire, rubber or elastic cords, or suitable strapping material to secure the cameras to the selected tree trunk. The camera should
be tightly mounted so that it does not move unless considerable force is applied.
If locks and chains (or bicycle lock) are necessary to secure the cameras against
theft, wait until cameras have been tightened with the wire before securing them.
A twig or wedge placed between the camera housing and the tree trunk can help
adjust the angle in which the sensor is pointed.
Once the camera is positioned, clear the area between the camera and the
path of travel of all vegetation. Anything that obstructs the beam reduces the
detection ability of the camera, and could result in obscured pictures. Large
leaves and blades of grass can result in false triggers when the sun heats up a
frond blowing in the wind. Also try to avoid pointing the cameras at objects in
direct sunlight that may absorb heat and trigger sensors such as large rocks or
sunlit streams. Be aware of the field of view for the camera and the sensor, and
be sure that both fields of view are clear.
Test the aim of the sensor by crossing in front of it. Do this on both the near
and far edge of the trail as well as the middle of the trail. Most camera trap
brands come equipped with an indicator light that will light up when the camera’s sensor makes detection. Approximate a typical target species by walking
in a crouch or crawling past the sensor. Make sure that every angle at which an
animal can pass in front of the camera is tested, and that in each instance the
sensor is triggered.
Occasionally, limitations in terrain or suitable trees hamper complete coverage of a trail. In such cases, lay brush or other obstructions down one side
of the trail to restrict the travel path and guide animals past the camera. This
technique is also useful if you are unable to set the camera well back from the
trail, and wish to deter an animal from passing so closely to a camera that it
cannot take a well-focused picture.
Once the camera is positioned and the field of view is clear, activate the camera.
If you rely on a camera with normal flash, be sure the flash is set to activate every
time the camera triggers, and that the red-eye and other features that slow the flash
or trigger are turned off. Be sure that the date and time stamp are activated and
properly set. To be sure that film are not later mixed, it is useful to use the first
frame of the film to identify the film number. This can be done by photographing a
white board with the film number clearly written in large numerals. If you do this,
activate the flash after the white board picture. Some teams photograph themselves
using their fingers to indicate film number (4 people; one's fingers to represent
thousands, one's hundreds, one's tens and one's the final digit [1-9]). If you use
digital cameras, some models allow additional information to be programmed
into the EXIF file that stores metadata for each image. Camera point can be programmed and later the metadata can be exported using a variety of software.
Monitoring the Cameras
The amount of animal traffic, human disturbance, and sensitivity of the camera
trap sensor will dictate how fast the camera memory fills. Film cameras should
probably be checked once a week if possible. In Sumatran forests, we left film
cameras for 30 days and most retained unexposed films (O’Brien et al. 2003).
Ideally, a digital camera should not run out of memory during the sampling
period. Given that 2GB flash cards may hold more than 2,000 images, it is
unlikely that a digital camera should stop working unless it malfunctions, fires
repeatedly, or unless animals linger in front of the camera for excessive amounts
of time. Realistically, one might expect 10% of cameras to fail for a range of
reasons or to use up the memory before the end of a 30-day sampling period.
Careful positioning of cameras should minimize sensor misfires and careful
maintenance should minimize mechanical or battery failure. It should not be
necessary to service the camera during a 30-day sampling period.
Wildlife Conservation Society | WORKING PAPER NO. 39
Andam, K.S., P.J. Ferraro, A. Pfaff, G.A. Sanchez-Azofelfa and J.A. Robalino.
2008. Measuring the effectiveness of protected area networks in reducing deforestation. PNAS 105:16089-16094.
Anderson, D.R., K.P. Burnham, G.C. White and D.L. Otis. 1983. Density
estimation of small-mammal populations using a trapping web and distance
sampling methods. Ecology 64:674-80.
Andelman, S.J. and M.R. Willig. 2003. Present patterns and future prospects for
biodiversity in the Western Hemisphere. Ecological Letters 6:818-824.
Baillie, J.E.M., B. Cullen, R. Amin, H.R. Akcakaya, S.H.M. Buchart, N
Brummitt, T.R. Meagher, M. Ram, C. Hilton-Taylor and G.M. Mace. 2008.
Toward monitoring global biodiversity. Conservation Letters 1:18-26.
Balmford, A., A. Bruner, P. Cooper, R. Constanza, S. Farber, R.E. Green, et al.
2002. Economic reasons for conserving wild nature. Science 297:950-953.
Balmford, A., P. Crane, A. Dobson, R.E. Green and G.M. Mace. 2005. The 2010
challenge: data availability, information needs and extraterrestrial insights. Phil.
Trans. R. Soc. B. 360:221-228.
Borcher, D.L. and M.G. Efford. 2007. Spatially explicit maximum likelihood
methods for capture-recapture studies. Biometrics 63:1-9.
Boulinear, T., J.D. Nichols, J.R. Sauer, J.E. Hines and K.P. Pollock. 1998.
Estimating species richness: The importance of heterogeneity in species detectability. Ecology 79:1018-10288.
Bruner, A.G., R.E. Gullison, R.E. Rice and G.A.B. da Fonsesca. 2001a.
Effectiveness of parks in protecting tropical biodiversity. Science 291:125-128.
Bruner, A.G., R.E. Gullison, R.E. Rice and G.A.B. da Fonsesca. 2001b. The
effectiveness of parks: response. Science 293:1007a.
Buckland, S.T., D.R. Anderson, K.P. Burnham and J.L. Laake. 1993. Distance
Sampling: Estimation of Biological Populations. Chapman and Hall, New York.
Buckland, S.T., A.E. Margurran, R.E. Green and R.M. Fewster. 2005. Monitoring
change in biodiversity through composite indices. Phil. Trans. R. Soc. Lond. B.
360(1454): 243 – 254.
Bunge, J. and M. Fitzpatrick. 1993. Estimmating the number of species: a
review. J. Am. Stat. Assoc. 88:364-373.
Butchart, S.H.M., A.J. Stattersfield, L.A. Bennun, S.M. Shutes, H.R. Akçakaya,
J.E.M. Baillie, S.N. Stuart, C. Hilton-Taylor and G.M. Mace. 2004. Measuring
global trends in the status of biodiversity: Red List Indices for birds. PLoS
Biology 2:2294–2304.
Butchart, S., H.R. Akçakaya, J. Chanson, J.E.M. Baille, B. Collen, S. Quader,
W.R. Turner, R. Amin, S.N. Stuart and C. Hilton-Taylor. 2007. Improvements
to the Red List Index. PLoS1 1:e140.
Cam, E., J.D. Nichols, J.R. Sauer, J.E. Hines and C.H. Flather. 2000. Relative
species richness and community completeness: Avian communities and urbanization in the mid-Atlantic states. Ecol. Appl. 10:1196-1210.
Conroy, M.J. 1996. Abundance indices. In: Measuring and Monitoring
Biological Diversity: Standard Methods for Mammals (Wilson, D.E., Cole, F.R.,
Nichols, J.D., Rudran, R., and Foster, M.S., eds.), pp 179-192, Smithsonian
Institution Press, Washington, D.C.
Collen, B., M. Ram, T. Zamin and L. McRae. 2008. The tropical biodiversity
data gap: addressing disparity in global monitoring. Tropical Conservation
Science 1:75-88.
Collen, B., J. Loh, S. Whitmee, L. McRae, R. Amin and J.E.M. Baillie. 2009.
Monitoring change in vertebrate abundance: the living planet index. Conserv.
Biol. 23:317-327
DeVries, R., A. Hansen, A.C. Newton and M.C. Hansen. 2005. Increasing isolation of protected areas in tropical forests over the past 20 years. Ecological
Applications 15:19-26.
Dobson. A. 2005. Monitoring global rates of biodiversity change: challenges
that arise in meeting the Convention on Biological Diversity (CBD) 2010 goals.
Phil. Trans. R. Soc. B. 360:229-241.
Dobson, A., D. Lodge, J. Alder, G.S. Cumming, J. Keymer, J. McGlade, H.
Mooney, J.A. Rusak, O. Sala, V. Wolters, D. Wall, R. Winfree and M.A.
Xenopoulos. 2006. Habitat loss, trophic collapse, and the decline of ecosystem
services. Ecology 87:1915-1924.
Efford, M.G. 2004. Density estimation in live-trapping studies. Oikos 106:598–
Ervin, J. 2003. Protected area assessments in perspective. BioScience 53:819-822.
Fewster, R.M., S. T. Buckland, G.M. Siriwardena, S.R. Baillie and J.D. Wilson.
2000. Analysis of population trends for farmland birds using generalized additive models. Ecology 81:1970-1984.
Gregory, R.D., D. Noble, R. Field, J.H. Marchant, M. Raven and D.W. Gibbons.
2003. Using birds as indicators of biodiversity. Ornis Hungaria 12-13:11-14.
Gregory, R.D., A. van Strien, P. Vořišek, A.W. Gmelig Meyling, D.G. Noble,
R.P.B. Foppen and D.W. Gibbons. 2005. Developing indicators for European
birds. Phil. Trans. R. Soc. B. 360:269-288.
Gregory, R.D., P. Vořišek, D.G. Noble, A. van Strien, A Klcaňová, M. Eaton,
A.W. Gmelig Meyling, A. Joys, R.P.B. Foppen and I.J. Burfield. 2008. The
generation and use of bird population inidicators in Europe. Bird Conservation
International 18:S223-S244.
Hastie, T.J. and R.J. Tibshirani. 1990. Generalized Additive Models. Chapman
and Hall, London, UK.
Hollenbeck, J.P and W.J. Ripple. 2007. Aspen snag dynamics, cavity-nesting
birds, and trophic cascades in Yellowstone’s northern range. Forest Ecology and
Management 255:1095-1103.
Hooper, D.U., F.S. Chapin III, J.J Ewel, A Hector, P. Inchausti, S. Lavorel, J.H
Lawton, D.M. Lodge, M. Loreau, S. Naeem, B. Schmid, H. Setälä, A.J. Symstad,
J. Vandemeer and D.A. Wardle. 2005. Effects of Biodiversity on ecosystem functioning: a consensus of current knowledge. Ecological Monographs 75:3-35.
Joppa, L.N., S.R. Loarie and S.L. Pimm. 2008. On the protection of protected
areas. PNAS 105:6673-6678.
Karanth, K.U. and J.D. Nichols. 1998. Estimation of tiger densities in India
using photographic captures and recaptures. Ecology 79: 2952-2862.
Karanth, K.U. and J.D. Nichols, (eds.). 2002. Monitoring Tigers and Their
Prey: A Manual for Researchers, Managers, and Conservationists in Tropical
Asia. Centre for Wildlife Studies, Bangalore, India.
Karanth, K.U. N.S. Kumar, V. Srinivas and A. Gopalaswamy. 2008. Revised
Monitoring Framework for Tigers Forever-Panthera sites. Unpublished report
to the Wildlife Conservation Society. 17 pp.
Kelly, M.J., A.J. Noss, L.R. Arispe, M. Di Bitetti, C.D. De Angelo, A. Paviolo, Y.E.
Di Blanco and L. Maffei. 2008. Estimating puma densities from remote cameras
across three study sites: Bolivia, Argentina, and Belize. J. Mammal. 89: 408-418.
Limpert, E., W.A. Stahel and M. Abbt. 2001. Log-normal distributions across
the sciences: keys and clues. Bioscience 51: 341-352.
Wildlife Conservation Society | WORKING PAPER NO. 39
Loh, J., R.E. Green, T. Ricketts, J. Lamoreux, M. Jenkins, V. Kapos and J.
Randers. 2005. The Living Planet Index: using species population time series to
track trends in biodiversity. Phil. Trans. R. Soc. 360:289-295.
Mace, G. M. and J. E. M. Baillie. 2007. The 2010 Biodiversity Indicators: challenges for science and policy. Conservation Biology 21:1406-1413.
MacKenzie, D.I. and J.A. Royle. 2005. Designing occupancy studies: general
advice and allocating survey effort. Ecol. Appl. 42:1105-1114.
MacKenzie, D.I., Nichols, J.D., Royle, J.A., Pollock, K.P., Bailey, L.L. and
Hines, J.E. 2006. Occupancy Estimation and Modeling: Inferring Patterns and
Dynamics of Species Occurrence. Academic Press, New York.
Magurran, A.E. 2004. Measuring Biological Diversity. Oxford: Blackwell
Meir, E., S. Andelman and H.P. Possingham. 2004. Does conservation planning
matter in a dynamic and uncertain world. Ecology Letters 7:615-622.
Millenium Ecosystem Assessment. 2005. Ecosystems and Human Well-being: A
Framework for Assessment. Island Press, Washington, D.C.
Norton-Griffiths, M. 2007. How many wildebeest do you need? World
Economics 8:41-64.
O’Brien, T.G. 2008. On the use of automated cameras to estimate species richness for large- and medium-sized rainforest mammals. Animal Conservation
O’Brien, T.G. and M.F. Kinnaird. 2008. A picture is worth a thousand words:
the application of camera trapping to the study of birds. Bird Conservation
International 18: S144-S162.
O’Brien, T.G., M.F. Kinnaird and H.T. Wibisono. 2003. Crouching tigers, hidden prey: Sumatran tiger and prey populations in a tropical forest landscape.
Anim. Cons. 6:131-139.
O’Brien, T.G., H.T. Wibisono and M.F. Kinnaird. In press. Species richness
estimation using camera traps: An Indonesian example. In: Camera Traps in
Animal Ecoclogy: Methods and Analyses (O’Connell, A.F., Nichols, J.D., and
Karanth, U.K. eds.), Springer Tokyo Inc.
Odling-Smee, L. 2005. Dollars and sense. Nature 436:614-616.
Parrish, J.D., D.P. Braun and R.S. Unnasch. 2003. Are we conserving what we
say we are? Measuring ecological intergrity within protected areas. BioScience
Pimm, S.L., H.L. Jones and J. Diamond. 1988. On the risk of extinction.
American Naturalist 132:757–785.
Pimm, S.L., G.J. Russell, J.L Gittleman and T.H. Brooks. 1995. The future of
biodiversity. Science 269:347-350.
Pollock, K.H., J.D. Nichols, T.R. Simons, G.L. Farnsworth, L.L. Bailey and
J.R. Sauer. 2002. Large scale wildlife monitoring studies: statistical methods for
design and analysis. Environmetrics 13:105-119.
Power, M.E., D. Tilman, J. Estes, B.A. Menge, W.J. Bond, L.S. Mills, G. Daily,
J.C. Castilla J. Lubchenco and R.T. Paine. 1996. Challenges in the quest for
keystones. BioScience 46:609-620.
Pringle, R.M., T.P. Young, D.I. Rubenstein and D.J. McCauley. 2007.
Herbivore-initiated interaction cascades and their modulation by productivity
in an African savanna. PNAS 104:193-197.
Purves, A., K.E. Jones and G.M. Mace. 2000. Extinction. Bioessays 22:11221133.
Raffaelli, D. 2004. How extinction patterns affect ecosystems. Science
Ray, J.C. 2005. Large carnivorous animals as tools for conserving biodiver-
sity: assumptions and uncertainties. Pp. 34-56 In: Large Carnivores and the
Conservation of Biodiversity. J.C. Ray, K.H. Redford, R.S. Steneck and J.
Berger (eds.) Island Press, Washington, D.C.
Ripple, W.J. and R.L Beschta. 2004. Wolves, elk, willows and trophic cascades
in the upper Gallitin range of southwestern Montana, USA. Forest Ecology and
Management 200:161-181.
Rodrigues, A.S.L., H.R. Akçakaya, S.J. Andelman, M.I. Bakaar, L. Boitani,
T.M. Brooks, J.S. Chanson, L.D.C. Fishpool, G.A.B. Da Fonseca, K.J. Gaston,
M.Hoffman, P.A. Marquet, J.D. Pilgrim, R.L. Pressey, J. Schipper, W. Secrest,
S.N. Stuart, L.G. Underhill, R.W. Waller, M.E.J. Watts and X. Yan. 2004.
Global gap analysis: priority regions for expanding the global preotected-area
network. BioScience 54:1092-1100.
Rowcliffe, J.M. and C. Carbone. 2008. Surveys using camera traps: are we
looking to a brighter future? Animal Conservation 11:185-186.
Silver, S.C., L.E.T. Ostro, L.K. Marsh, L. Maffei, A.J. Noss, M.J. Kelly, R.B.
Wallace, H. Gómez and G. Ayala. 2004. The use of camera traps for estimating
jaguar Panthera onca abundance and density using capture/recapture analysis.
Oryx 38: 148-154.
Solan, M., B.J. Cardinale, A.L. Downing, K.A.M. Engelhardt, J.L. Ruesink and
D.S. Srivastava. 2004. Extinction and ecosystem function in the marine benthos.
Science 306:1177-1180.
Terborgh, J., L. Lopez, P. Nuñez, M. Rao, G. Shahabuddin, G. Orihuela, M.
Riveros, R. Ascanio, G.H. Adler, T.D. Lambert and L. Balbas. 2001. Ecological
meltdown in Predator-free forest fragments. Science 294:1923-1926.
Tobler, M.W., S.E. Carrilo-Percastegui, R.L. Pitman and G. Powell. 2008a. An
evaluation of camera traps for inventorying large and medium sized terrestrial
rainforest mammals. Animal Conservation 11:169-178.
Tobler, M.W., S.E. Carrilo-Percastegui, R.L. Pitman, R. Mares and G. Powell.
2008b. Further notes on the analysis of mammal inventory data collected with
camera traps. Animal Conservation 11:187-189.
Vanclay, J.K. 2001. The Effectiveness of parks. Science 293:1007a.
Wood, S. 2006. Generalized Additive Models. An Introduction with R. Chapman
and Hall/CRC, Boca Raton, Fl.
Yoccoz, N.G., J.D. Nichols and T. Boulinier. 2001. Monitoring of biological
diversity in space and time. TREE 16:446-453.
Zar, J.H. 1999. Biostatistical Analysis: Fourth Edition. Prentice Hall, Upper
Saddle River, New Jersey, USA.
Zavaleta, E.S. and K.H. Hulvey. 2004. Realistic species losses disproportionately
reduce grassland resistance to biological invaders. Science 306:1175-1177.
Wildlife Conservation Society | WORKING PAPER NO. 39
Camera traps have been used by WCS and ZSL projects throughout the world to
photograph, among others, snow leopards in Afghanistan and sarus cranes and giant
ibises in Cambodia.
Model specification using Generalized Additive Models.
From Fewster et al. (2000), with permission of R. Fewster and S.
Generalized additive models (GAMs) are flexible extensions of general linear models (GLMs). We assume that the counts yit follow independent Poisson distributions, with mean μit for the count in site i in year t. However, the linear predictor
associated with the GLM, is replaced by a more general additive predictor. This
allows the change in mean abundance over time to follow any smooth curve, not
just a linear form as in linear regression or a sequence of unrelated estimates as in
log-linear Poisson regression. The form of the predictor function is the principal
difference between the GLM and the GAM. We write the additive predictor as
log(μit) = αi + f(t)
The expected count μit in site i in year t therefore depends upon the site
effect αi, and upon any number of other smoothly varying quantities, which
are summarized by the value f(t) in year t. The notation f(t) simply represents
any smooth function of time. The GAM is fitted by estimating the parameters
αi and the smooth function f. There are two special cases of the GAM formulation that fall into the category of GLMs. The first is the simple linear trend
model, in which f(t) = γt for a single parameter γ to be estimated. In this case,
the expected abundance within each site varies linearly on a logarithmic scale
with time. The second case is the log-linear Poisson regression model for which
f(t) = βt for parameters β1, . . . , βT to be estimated. In this instance, the function
f is no longer smooth, and is obtained by joining the estimates βt with straight
lines. These two cases lie at opposite extremes of the GAM framework. The first
has maximum smoothness in the function f, a single straight line; the second
has minimum smoothness, a sequence of unconstrained estimates joined by linear segments. Between the two extremes lie functions s with greater flexibility
than the linear trend f(t)=γt, but with smooth output in contrast to the discrete
annual estimates f(t) = βt. These intermediate GAM curves provide opportunities for eliciting long-term nonlinear trends that are not available at the GLM
The output from the GAM is visualized as follows. The fitted year effect curve
f(t) is common to all sites, so that for any two sites i1 and i2, the curves log(μi1t)
Wildlife Conservation Society | WORKING PAPER NO. 39
and log(μi2t) are parallel. The intercepts of these curves are determined by the
site effects, respectively αi1 and αi2. Consequently, every site is subject to the same
trend in the logarithm of expected count over time, although the absolute values
differ between sites. The assumption that log abundance follows the same trend in
every site, although simplistic, is the same as that implicit in the traditional Poisson
regression models. Indeed, the assumption is perhaps better justified in the GAM
context, because it is more reasonable to suggest that there is a smooth pattern of
change common to all sites than to assert that all sites are subject to the same series
of annual fluctuations. Nonetheless, the GAM system also accommodates models
that allow trends to vary between different regions, or even within every site.
Once an estimate f has been obtained for the smooth function f, the annual
abundance index curve I(t) is calculated as:
I (t ) =
total predicted count for year t exp( fˆ (t ))
total predicted count for year 1 exp( fˆ (1))
Before the function f can be estimated, the required level of smoothing must
be established. At the two extremes of maximum and minimum smoothing,
where the GAM is equivalent to a GLM, it is possible to specify f in parametric
form: namely f(t) = γt at the upper extreme and f(t) = βt at the lower. Between
these extremes, however, the function f is not defined in terms of parameters,
but is estimated nonparametrically from the data by means of scatterplot
smoothers. The shape of the function is therefore determined by the data rather
than being restricted to a parametric form. Note the distinction between the
smoothing procedures in a GAM, which are part of the model-fitting process,
and the smoothing of Mountford or TRIM indices in a second step after the
fit is complete.
Several options are available for the scatterplot smoother used in estimating
f, although experimence suggests that the choice has little effect on the final
index curve. All of the analyses presented in Fewster et al. (2000) were conducted using smoothing splines; other options include locally weighted regression smoothers, kernel smoothers, and running-median smoothers.
Smoothing splines fit the data using piecewise cubic polynomials. They are
chosen to satisfy a penalized least squares criterion, which optimizes the fit while
penalizing roughness to some pre-specified extent. The extent to which roughness
is penalized, or equivalently the level of smoothing that is applied, is calibrated
by a quantity known as the degrees of freedom (df). As the degrees of freedom
are increased, the function f gains in flexibility: more turning points and gradient
changes are accommodated. A straight line f(t) = γt corresponds to a fit with 1
df; the curve follows a single direction only, with no gradient changes or turning
points. Conversely, a piecewise linear fit f(t) = βt (t = 1, . . . , T) employs T - 1 df; a
separate gradient is allowed between every successive pair of points. The degrees
of freedom associated with the curve f may take any value between 1 and T - 1,
and are loosely interpreted as the number of parameters used in fitting f.
The choice of the value for df is an important part of the modeling process.
For clarity, we shall write fd for the curve f to be fitted using a smoothing spline
on the variable t with d degrees of freedom, and we shall refer to the associated
model as a ‘‘GAM with d degrees of freedom.’’ The choice of d depends largely
on the objectives of the analysis. For inference about long-term trends, a smooth
index curve is required, corresponding to low df; whereas information about
annual fluctuations requires unconstrained annual estimates and the maximum
value of d. Consideration of the length of the time series is also important,
because longer time series will demand higher values of d if it is required to
maintain a given level of flexibility in the trend curve.
Certain conceptual difficulties enter at this point, because it is not possible
to assert that a GAM with df = a is more realistic than a GAM with df = b.
The GAMs separate underlying trend from short-term fluctuation, but the precise point at which a signal can be considered fluctuation rather than trend is
ill-defined. The specification of df effectively represents models with different
criteria for achieving this separation of trend from noise, and there is no right
answer against which to judge them. To overcome this difficulty, it is necessary
to formulate our own ideas about a reasonable definition of noise and trend.
As d is raised from a low starting value, the index curve begins to reveal new
features as it more closely tracks the data, but still remains smooth until a
period of stability is reached. Further increases in d serve only to roughen the
output. It is reasonable to take those points at which the output has stabilized,
but is still smooth, as our definition of true trend. Experiments with truncation of Common Birds Census data suggested that a value for d of roughly 0.3
times the length of the time series tended to produce trend curves with suitable
complexity and smoothness, although we stress that advice will vary according
to precise objectives and data. It is always necessary to plot indices from GAMs
with a range of df before settling on a final value.
Automatic selection of the value of df in a generalized additive model is
sometimes recommended, using, for example, Akaike’s Information Criterion
(AIC) or generalized cross-validation. Each automatic procedure produces a single value of d that is optimal with respect to some criterion. Because the quantity optimized might be very different from our own requirements of the model,
however, we do not recommend the use of automatic selection procedures in
the context of trend analysis. For example, we might wish to examine the same
data both for longterm trends and for information about annual fluctuations.
Although different values of d are clearly required for the two purposes, an
automatic selection routine would be unable to distinguish between the distinct
objectives, and would provide the same value for both.
Wildlife Conservation Society | WORKING PAPER NO. 39
R-Program to calculate WPI, confidence intervals and
change points.
Written by Jorge Ahumada, Technical Director, CI/TEAM
#Boostrapping script to calculate WPI and create 95% confidence limits for
time series and change points in a generalized additive model (Fewster et al.
#The script returns the means and confidence limits of the WPI geometric mean
as well as the confidence limits of the second derivative.
#The function also returns a pdf with the graph of WPI and 95%CI with significant change points identified.
#This script can have small changes depending on the length of the time
#This version has been modified for Tim O’Brien’s analysis. Species are re-sampled, not sites. Assumes a neighborhood of one point to each side to calculate the
second derivative. Longer time series should use a neighborhood of 2 points on
each side.
#This bootstrap requires the mgcv library for to calculate the GAM routine.
#Written by Jorge Ahumada, Technical Director, CI/TEAM.
#------------------------#Data is a three column data set with the occupancies or densities. Column 1:
species code, Column2: Year (or time period), Column3: index of measurement
(density, occupancy, etc.)
# nsim: Number of resamplings in the bootstrap, 999 by default
# nyr: number of sampling periods (or years) of data
# title: any title (in double quotes) that will appear as the title in the graph
# species: number of species in the sample
#----------------------------------------f.boot.geom.mean <-function(data,nsim=999,nyr,title,species){
#several parameters for the gam
#store the results for the index
#Create table with the data
#Main loop
for(i in 1:nsim){
#create bootstrap sample
#table with bootstrapped sites
#loop to go through j years in this case 5.
for(j in 1:nyr)
#in case there are zeros
#calculate geometric mean
# fit GAM
#store the results for the geometric means
#calculate the second derivative and store the results in res.sdr
for(j in 2:(nyr-1))
res.sder[i,j] <- logsm[j-1]-2*logsm[j]+logsm[j+1]
#get rid of unwanted NaN
#calculate means and confidence limits
Wildlife Conservation Society | WORKING PAPER NO. 39
#calculate the 95% ci of the second derivative
#figure out which symbols to use in the plot
#plot the whole thing and put in a pdf
plot(time,mean.smooth,xlab=”time”,ylab=”geometric mean
of relative abundance”,
#return result matrices
#Function to calculate the geometric mean
f.geom.mean.occ <- function(data){
apply(data,2,function(x) exp(1/length(data[,1])*sum(log(x/data[,1]))))
WCS Working Paper No. 1
Bleisch, William V. (1993) Management Recommendations for Fanjing
Mountain Nature Reserve and Conservation at Guizhou Golden Monkey &
WCS Working Paper No. 2
Hart, John A. and Claude Sikubwabo. (1994) Exploration of the Maiko
National Park of Zaire, 1989-1994, History, Environment and the Distribution
and Status of Large Mammals.
WCS Working Paper No. 3
Rumiz, Damian and Andrew Taber. (1994) Un Relevamiento de Mamíferos y
Algunas Aves Grandes de la Reserva de Vida Silvestre Ríos Blanco y Negro,
Bolívia: Situación Actual y Recomendaciones.
WCS Working Paper No. 4
Komar, Oliver and Nestor Herrera. (1995) Avian Density at El Imposible
National Park and San Marcelino Wildlife Refuge, El Salvador.
WCS Working Paper No. 5
Jenkins, Jerry. (1995) Notes on the Adirondack Blowdown of July 15th, 1995:
Scientific Background, Observations, and Policy Issues.
WCS Working Paper No. 6
Ferraro, Paul, Richard Tshombe, Robert Mwinyihali, and John Hart. (1996)
Projets Integres de Conservation et de Developpement; un Cadre pour
Promouvoir la Conservation et la Gestion des Ressources Naturalles.
WCS Working Paper No. 7
Harrison, Daniel J. and Theodore G. Chapin. (1997) An Assessment of Potential
Habitat for Eastern Timber Wolves in the Northeastern United States and
Connectivity with Occupied Habitat on Southeastern Canada.
WCS Working Paper No. 8
Hodgson, Angie. (1997) Wolf Restoration in the Adirondacks? The Question
of Local Residents.
WCS Working Paper No. 9
Jenkins, Jerry. (1997) Hardwood Regeneration Failure in the Adirondacks:
Preliminary Studies of Incidence and Severity.
WCS Working Paper No. 10
García Víques, Randall. (1996) Propuesta Técnica de Ordenamiento Territorial
con Fines de Conservación de Biodiversidad en Costa Rica: Proyecto GRUAS.
WCS Working Paper No. 11
Thorbjarnarson, John and Alvaro Velasco. (1998) Venezuela’s Caiman Harvest
Program: A Historical Perspective and Analysis of its Conservation Benefits.
WCS Working Paper No. 12
Bolze, Dorene, Cheryl Chetkiewicz, Qui Mingjiang, and Douglas Krakower.
(1998) The Availability of Tiger-Based Traditional Chinese Medicine Products
and Public Awareness about the Threats to the Tiger in New York City’s Chinese
Communities: A Pilot Study.
WCS Working Paper No. 13
O’Brien, Timothy, Margaret F. Kinnaird, Sunarto, Asri A. Dwiyahreni, William
M. Rombang, and Kiki Anggraini. (1998) Effects of the 1997 Fires on the Forest
and Wildlife of the Bukit Barisan Selatan National Park, Sumatra.
Wildlife Conservation Society | WORKING PAPER NO. 39
WCS Working Paper No. 14
McNeilage, Alistair, Andrew J. Plumptre, Andy Brock-Doyle, and Amy Vedder.
(1998) Bwindi Impenetrable National Park, Uganda. Gorilla and Large Mammal
Census, 1997.
WCS Working Paper No. 15
Ray, Justina C. (2000) Mesocarnivores of Northeastern North America: Status
and Conservation Issues.
WCS Working Paper No. 16
Kretser, Heidi. (2001) Adirondack Communities and Conservation Program:
Linking Communities and Conservation Inside the Blue Line.
WCS Working Paper No. 17
Gompper, Matthew E. (2002) The Ecology of Coyotes in Northeastern North
America: Current Knowledge and Priorities for Future Research.
WCS Working Paper No. 18
Weaver, John L. (2001) The Transboundary Flathead: A Critical Landscape
for Carnivores in the Rocky Mountains.
WCS Working Paper No. 19
Plumptre, Andrew J., Michel Masozera, Peter J. Fashing, Alastair McNeilage,
Corneille Ewango, Beth A. Kaplin, and Innocent Liengola. (2002) Biodiversity
Surveys of the Nyungwe Forest Reserve in S.W. Rwanda.
WCS Working Paper No. 20
Schoch, N. (2003) The Common Loon in the Adirondack Park: An Overview
of Loon Natural History and Current Research.
WCS Working Paper No. 21
Karasin, L. (2003) All-Terrain Vehicles in the Adirondacks: Issues and Options.
WCS Working Paper No. 22
Clarke, Shelly. (2002) Trade in Asian Dry Seafood, Characterization, Estimation
and Implications for Conservation.
WCS Working Paper No. 23
Mockin, Miranda H., E.L. Bennett, and D.T. LaBruna. (2005) Wildlife Farming:
A Viable Alternative to Hunting in Tropical Forests?
WCS Working Paper No. 24
Ray, Justina C., Luke Hunter, and Joanna Zigouris. (2005) Setting Conservation
and Research Priorities for Larger African Carnivores.
WCS Working Paper No. 25
Redford, Kent H. and Michael Painter. (2006) Natural Alliances Between
Conservationists and Indigenous Peoples.
WCS Working Paper No. 26
Agrawal, Arun and Kent Redford. (2006) Poverty, Development, and
Biodiversity Conservation: Shooting in the Dark?
WCS Working Paper No. 27
Sickler, Jessica, John Fraser, Sarah Gruber, Paul Boyle, Tom Webler, and Diana
Reiss. (2006) Thinking About Dolphins Thinking.
WCS Working Paper No. 28
Castillo, Oscar, Connie Clark, Peter Coppolillo, Heidi Kretser, Roan McNab,
Andrew Noss, Helder Quieroz, Yemeserach Tessema, Amy Vedder, Robert
Wallace, Joseph Walston, and David Wilkie. (2006) Casting for Conservation
Actors: People, Partnerships and Wildlife.
WCS Working Paper No. 29
Redford, Kent H. and Eva Fearn, eds. (2007) Protected Areas and Human
Displacement: A Conservation Perspective.
WCS Working Paper No. 30
Redford, Kent H., and Eva Fearn, eds. (2007) Ecological Future of Bison in
North America: A Report from a Multi-stakeholder, Transboundary Meeting.
WCS Working Paper No. 31
Smith, Brian D., Robert G. Shore, and Alvin Lopez. (2007) Status and
Conservation of Freshwater Populations of Irrawaddy Dolphins.
WCS Working Paper No. 32
Redford, Kent H. and Eva Fearn, eds. (2007) Protected Areas and Human
WCS Working Paper No. 33
Beckmann, J. P., L. Karasin, C. Costello, S. Matthews, and Z. Smith. (2008)
Coexisting with Black Bears: Perspectives from Four Case Studies Across
North America.
WCS Working Paper No. 34
Painter, M., A.R. Alves, C. Bertsch, R. Bodmer, O. Castillo, A. Chicchón,
F. Daza, F. Marques, A. Noss, L. Painter, C. Pereira de Deus, P. Puertas,
H.L. de Queiroz, E. Suárez, M. Varese, E.M. Venticinque, and R.Wallace.
(2008) Landscape Conservation in the Amazon Region: Progress and Lessons.
WCS Working Paper No. 35
Brodie, Jedediah F. (2008) A Review of American Bison (Bos bison)
Demography and Population Dynamics.
WCS Working Paper No. 36
Redford, Kent H. and Catherine Grippo, eds. (2008) Protected Areas,
Governance, and Scale.
WCS Working Paper No. 37
Estes, Richard D. and Rod East. (2009) Status of the Wildebeest
(Connochaetes Taurinus) in the Wild 1967-2005.
WCS Working Paper No. 38
Olupot, William, Alastair J. McNeilage and Andrew J. Plumptre. (2009) An
Analysis of Socioeconomics of Bushmeat Hunting at Major Hunting Sites in
WCS Working Paper No. 39
O’Brien, Tim. (2010) Wildlife Picture Index: Implementation Manual Version 1.0.
WCS canada conservation reports
WCS Canada Conservation Report #1
BIG ANIMALS and SMALL PARKS: Implications of Wildlife Distribution and
Movements for Expansion of Nahanni National Park Reserve. John L. Weaver.
WCS Canada Conservation Report #2
Freshwater Fish in Ontario's Boreal: Status, Conservation and Potential Impacts
of Development. David R. Browne. 2007.
WCS Canada Conservation Report #3
Carnivores in the Southern Canadian Rockies: Core Areas and Connectivity
Across the Crowsnest Highway. Apps, Clayton D., John L. Weaver, Paul C.
Paquet, Bryce Bateman and Bruce N. McLellan. 2007.
WCS Canada Conservation Report #4
Conserving Caribou Landscapes in the Nahanni Trans-Border Region Using
Fidelity to Seasonal Ranges and Migration Routes. John L. Weaver. 2008.
Wildlife Conservation Society | WORKING PAPER NO. 39
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF