Biological Conservation (2005) 22 (1): 99-112.

http://dx.doi.org/10.1016/j.biocon.2004.07.004
Copyright © 2004 Elsevier Ltd All rights reserved.

Sensitivity Of Conservation Planning To Different

Approaches To Using Predicted Species Distribution
Data
Kerrie A. Wilson, Michael I. Westphal, Hugh P. Possingham and Jane Elith

Abstract

The main role of conservation planning is to design reserve networks to protect biodiversity in situ.
Research within the field of conservation planning has focused on the development of theories and
tools to design reserve networks that protect biodiversity in an efficient and representative manner.
Whilst much progress has been made in this regard, there has been limited assessment of the sensitivity
of conservation planning outcomes to uncertainty associated with the datasets used for conservation
planning. Predicted species distribution data are commonly used for conservation planning because the
alternatives (e.g. survey data) are incomplete or biased spatially. However, there may be considerable
uncertainty associated with the use of predicted species distribution data, particularly given the variety
of approaches available to generate a dataset from such predictions for use in conservation planning.
These approaches range from using the probabilistic data directly to using a threshold identified a priori
or a posteriori to convert the probabilistic data to presence/absence data. We assess the sensitivity of
conservation planning outcomes to different uses of predicted species distribution data. The resulting
reserve networks differed, and had different expected species representation. The choice of approach
will depend on how much risk a conservation planner is willing to tolerate and how much efficiency
can be sacrificed.

Keywords: conservation planning; uncertainty; reserve design; species distribution models

1. Introduction
The increasing pressures exerted on our natural environment by the human population make
conservation areas crucial for the persistence of biological diversity (McNeely, 1994 and
Groombridge and Jenkins, 2002). We use the term ‘conservation area’ to describe any area of
land or sea managed for the persistence of biodiversity and natural processes in situ, through
constraints on incompatible land uses. The term ‘reserve network’ denotes a system of
conservation areas.
Selecting conservation areas in an ad hoc manner or for the protection of particular
species generally results in the conservation of economically marginal land and
unrepresentative reserve networks (Mark, 1985, Henderson, 1992, Pressey, 1993, Pressey,
1994, Pressey and Tully, 1994 and Rouget et al., 2003). In recognition of these problems and
the limited resources available for conservation, research efforts in the field of conservation
planning have focused on the development of principles and tools to configure efficient and
representative reserve networks.
Scoring systems were developed in the 1980s in an attempt to provide an explicit and
rational basis for selecting conservation areas (Margules and Usher, 1981, Terborgh and
Winter, 1983, Purdie et al., 1986, Smith and Theberge, 1986, Smith and Theberge, 1987,
Usher, 1986, Margules et al., 1988 and Pressey and Nicholls, 1989). These systems scored or
rated potential conservation areas against several criteria to provide an overall indication of
their conservation value. However, scoring systems did not identify efficient and
representative reserve networks (Pressey and Nicholls, 1989 and Pressey, 1997).
In a move away from scoring systems, systematic conservation planning techniques were
developed. This commenced with Kirkpatrick’s introduction of the minimum-set, which
involved the identification of a network of complementary conservation areas that efficiently
achieved a set of conservation targets (Kirkpatrick, 1983, Justus and Sarkar, 2002 and
Pressey, 2002). Since then, there has been increasing emphasis on meeting quantitative
conservation targets within a network of complementary conservation areas.
Systematic conservation planning is an evolving discipline at the interface of
biological, mathematical and social sciences (Margules and Pressey, 2000 and Possingham et
al., 2000). Systematic conservation planning techniques offer significant improvements over
ad hoc and scoring approaches, as they are goal-directed, transparent, defensible, flexible and
aim to efficiently meet quantitative targets within a network of representative and
complementary conservation areas. Systematic conservation planning has informed
conservation in both terrestrial (Pressey, 1998 and Cowling and Pressey, 2003) and marine
realms (Ferdana, 2002 and Great Barrier Reef Marine Park Authority, 2003).
Systematic conservation planning requires information on the distribution of
biodiversity within the planning region; however, such information is often lacking.
Therefore, the features used for conservation planning act as surrogates for total biodiversity
(Ferrier et al., 2000). These surrogates have included particular groups of species (Richardson
and Funk, 1999 and Clark and Slusher, 2000), vegetation communities (Woinarski et al.,
1996) and land classifications (Kirkpatrick and Brown, 1994 and Pressey et al., 2000).
When species are used as a surrogate for biodiversity, often the information on the
distribution of species is incomplete, biased spatially (e.g. towards areas that are easily
accessible) or biased toward icon species (Podger et al., 1990, Austin, 1998, Keller and
Scallan, 1999, Polasky et al., 2000 and Funk and Richardson, 2002). Additionally, systematic
surveys for more than a few taxa on a regional scale are uncommon (Haila and Margules,
1996). Despite this, the selection of conservation areas cannot be delayed pending acquisition
of complete survey data; options for conservation may be dramatically reduced in the interim.
Instead, predicted species distribution data can be used (Margules and Nicholls, 1987,
Margules and Stein, 1989, Scott et al., 1993, Richardson and Funk, 1999, Clark and Slusher,
2000 and Jennings, 2000). Predicted species distributions are based on the modelled
relationships between species survey data and mapped environmental information (Guisan
and Zimmermann, 2000). We consider approaches that predict the probability of occurrence
of species in unsurveyed areas (e.g. predicted species distribution data generated using
logistic regression, McCullagh and Nelder, 1989). Such predictions can be generated for most
species. However, difficulties might arise when modeling the potential distribution of species
with ranges that are the product of historical environmental change and complex disturbance
regimes (Kirkpatrick and Brown, 1991).
Predicted species distribution data can exhibit considerable uncertainty (Elith et al.,
2002). For example:
(a) Species survey data may be inaccurately recorded, resulting in false-positive and
false-negative errors;
(b) Errors in environmental data may arise during interpretation and processing of
satellite images or aerial photographs;
(c) Uncertainty may arise from the simplifying assumptions used in the modelling
process.
For example, an environmental variable might be used in a species distribution model but this
variable might be correlated with another unmeasured variable that actually drives the
distribution of the species.
Various approaches are available for generating a dataset from predicted species
distribution data. The most common approach involves converting the probabilities of
occurrence to presence/absence data using a threshold (Li et al., 1997, Manel et al., 1999,
Manel et al., 2001, Fleishman et al., 2001 and Fleishman et al., 2003). An alternative
approach is to use the probabilities of occurrence directly. This avoids losing information by
artificially converting probabilities of occurrence into presence/absence data (Margules and
Nicholls, 1987, Polasky et al., 2000 and ReVelle et al., 2002).
The different approaches to using predicted species distribution data for conservation
planning can be thought of as different ways to formulate the conservation planning problem.
Generally, however, only one problem formulation is assessed and the objectives and
constraints within it are regarded as the only ones applicable. Little attention has been paid to
the sensitivity of conservation planning outcomes to different problem formulations.
Comparatively more attention has been given to the performance of reserve selection
algorithms, including their ability to generate optimal solutions and their speed of operation
(for example, Pressey et al., 1996, Pressey et al., 1997, Csuti et al., 1997 and Kelley et al.,
2002). The choice of approach to using predicted species distribution data might influence
which areas are identified as being important for conservation and their expected
representation of species. Here, the focus is on the different ways to formulate the
conservation planning problem, as opposed to comparing reserve selection algorithms.
Here, we assess the sensitivity of conservation planning outcomes to the different approaches
to using predicted species distribution data. The similarity of reserve network solutions
identified using the different datasets is evaluated. Additionally, the expected representation
of species in the different reserve network solutions is used to indicate which approach may
provide the best outcomes for nature conservation.

2. Methods
2.1. Study area and species survey data
The study region in Victoria, Australia, comprises portions of the Goldfields, Wimmera,
Murray Mallee and Central Victorian Upland bioregions of the North Central Catchment
Management Authority area: a total area of almost two million hectares. The region contains
5149 remnants of native vegetation (predominately belonging to the Box–Ironbark forest
type), which cover 306,107 ha. The survey data held for four plant species: Eucalyptus
tricarpa (a common overstorey species), Pultenaea largiflorens (a common understorey
species), Hibbertia exutacies (a common ground layer species) and Acacia ausfeldii (a rare
and threatened understorey species) was obtained (Department of Sustainability and
Environment, 2002). Survey data for these four species were not available for the entire study
region. Consequently, predictions of species occurrence throughout the study region were
generated by extrapolating the survey data to unsurveyed areas using the modelled
relationship between the known occurrence of each species and environmental variables.
These relationships were derived from a logistic regression model (Agresti, 1996). The
probability of occurrence of each species in unsurveyed areas were generated for each one
hectare grid cell (Wilson, 2003). Grid cells in which the species were known to occur were
allocated a probability of one. The one hectare grid cells are of a size that reasonably
represents the variation in the landscape and vegetation, and the accuracy of the species
survey data.
2.2. Conservation planning objectives and species conservation targets
The aim of systematic conservation planning, formulated in a decision theory framework, is
either to minimise or to maximise the value of an objective function, subject to constraints
that control the choice of areas selected to be part of the reserve network. These formulations,
referred to as the minimum-set and the maximal coverage problems, differ according to the
perceived constraints and objectives.
The objective of the minimum-set problem is to minimise resources expended, such
as the area reserved, subject to the constraint that all features meet their conservation targets
(Kirkpatrick, 1983, Margules and Nicholls, 1987, Margules et al., 1988, Possingham et al.,
1993 and ReVelle et al., 2002). The objective of the maximal coverage problem is to
maximise conservation of features subject to the constraint that the resources expended does
not exceed a specified amount (Satersdal et al., 1993, Church et al., 1996, Haight et al., 2000,
Polasky et al., 2000, Polasky et al., 2001, Arthur et al., 2002, Camm et al., 2002 and ReVelle
et al., 2002). Since there are limits on the amount of land that can be set aside for nature
conservation, it is prudent to select a set of conservation areas that achieves comprehensive
representation or meets targets at minimum cost. We adopted the minimum-set framework;
the objective was to minimise the area reserved whilst meeting areal conservation targets
allocated to each of the four plant species. We recognise that a reserve network based on the
requirements of only four species is unlikely to meet the requirements of all species occurring
in the region. Rather, the data for these four species provide a simple data matrix to evaluate
the different approaches to using species distribution data and to assess the sensitivity of
conservation planning to these approaches.
We employed the areal extent of the estimated original distribution of the Ecological
Vegetation Classes in which the common plant species occur (Muir et al., 1995) as a
surrogate measure of their historical distribution. A conservation target of 15% of the pre-
European extent of each Ecological Vegetation Classes in which the species presently occur
was allocated to each of the common plant species (Table 1). This is an arbitrary target but
follows the target currently sought for Australia’s forest types (JANIS, 1997). The procedure
outlined by Burgman et al. (2001) was employed to set the conservation target for the rare
plant species, A. ausfeldii (Wilson, 2003). This procedure uses expert judgement of life
history parameters and the determination of a target area that results in a specified quasi-
extinction risk (0.1% chance of falling below 50 individuals at least once in the next 50
years). Calculation of the target area took into account deterministic and stochastic threats,
such as eucalyptus oil harvesting and land clearing that may reduce available habitat for this
species within the study region (Wilson, 2003).

Table 1. Conservation targets for the four plant species

Species Target (ha)

Acacia ausfeldii 1596

Eucalyptus tricarpa 51615

Hibbertia exutacies 38842

Pultenaea largiflorens 8862

2.3. Threshold approaches – converting probabilistic data to presence/absence data

First, we assumed that a species was absent from an area if the probability of occurrence was
<0.5 and present in an area if the probability of occurrence was equal to or greater than 0.5.
This threshold was chosen prior to generating the probabilities of occurrence (Li et al., 1997,
Manel et al., 1999, Manel et al., 2001, Fleishman et al., 2001 and Fleishman et al., 2003) and
is referred to hereafter as Threshold Method 1.
The main disadvantage of using a threshold chosen a priori is that unequal group
sizes in the original survey data can influence the probabilities generated from a logistic
regression model, with probabilities biased toward the larger group (Hosmer et al., 1988).
This bias is sensible as the probabilities truly reflect frequencies of presences and absences in
the survey data. For example, the survey data used to generate the species distribution models
had unequal group sizes (that is, more absence observations than presence observations).
Consequently, the predicted probabilities are biased toward zero and their mean value is low
(Table 2). A threshold chosen a priori might be much greater than the probability of
occurrence values predicted from the models. For example, a threshold of 0.5 may mean that
nearly all predictions for rare species are realised as absent. Compared with thresholds
selected a posteriori, a priori thresholds are unable to be adapted to specific datasets and user
needs.
Table 2. Group sizes of the original survey data (the number of presences and absences)
and the mean value of the predicted probabilities. The threshold probabilities for the four
species when using Threshold Method 2 are also provided

# of # of Mean value of Threshold probability

Species
presences absences predictions ± SD using Threshold Method 2

Acacia ausfeldii 61 662 0.09 ± 0.12 0.60

Eucalyptus
185 564 0.30 ± 0.23 0.62
tricarpa

Hibbertia
239 531 0.36 ± 0.26 0.77
exutacies

Pultenaea
202 573 0.32 ± 0.21 0.72
largiflorens

We chose the second threshold a posteriori by trading off the sensitivity and specificity of the
generated datasets (Zweig and Campbell, 1993, Fielding and Bell, 1997 and Manel et al.,
2001). Sensitivity is the proportion of observations correctly predicted as a presence.
Specificity is the proportion of observations correctly predicted as an absence. This approach
is referred to hereafter as Threshold Method 2.
Using Threshold Method 2, a threshold was chosen that ensured the converted
presence/absence data had a desired specificity level. In selecting areas for inclusion in a
reserve network, a conservation planner might be cautious, and wish to ensure that only
highly suitable areas for a species are selected. The threshold probability would therefore
need to be stringent, the sensitivity of the generated dataset would be low and the specificity
would be high. The sensitivity and specificity of the datasets derived using different
thresholds were evaluated by converting the predicted probabilities of occurrence to
presence/absence data and comparing the generated dataset with the original survey data. We
selected a threshold probability so that the sensitivity of the generated dataset was 10% and
the specificity was 90%. The threshold probabilities chosen using this approach are provided
in Table 2.
We chose the final threshold a posteriori using information about the relative costs of
false-positive and false-negative errors, whilst taking into account the prevalence of the
species (the proportion of presence observations in the survey data; Zweig and Campbell,
1993). This approach is referred to hereafter as Threshold Method 3.
For Threshold Method 3, the relative costs of false-positive and false-negative errors were
determined. In conservation terms, the cost of a false-positive error is that associated with
selecting an area to be conserved because a species was predicted to be present in the area,
but in fact was not. The cost of a false-negative error is that associated with excluding an area
from the reserve network on the basis that the species was predicted to be absent from the
area, when it was actually present.
The relative costs of false-positive and false-negative errors can extend from zero to
infinity. For example, a cost ratio of 0.2 occurs when the cost of a false negative error is
greater than the cost of a false positive error. A cost ratio of one occurs when the cost of a
false negative error and the cost of a false positive error are equal. A cost ratio of five occurs
when the cost of a false positive error is greater than the cost of a false negative error. The
latter would be the preferred cost ratio of a cautious conservation planner who wishes to
ensure that only highly suitable habitats for a species are selected for the reserve network.
To choose a threshold that yields false-positive and false-negative errors in the optimal
proportion, one must also incorporate information on the prevalence of the species, or the
proportion of presence observations in the underlying survey data (Table 2 and Table 3). Low
prevalence generally calls for lower thresholds, as the probabilities of occurrence generated
for a rare species will be biased towards zero (Swets, 1988, Zweig and Campbell, 1993 and
Manel et al., 2001).
Table 3. Species prevalence measures, slope and associated threshold probabilities chosen
using Threshold Method 3

Species (1 − p)/p Slope Threshold probability using Threshold Method 3

Acacia ausfeldii 11.5 57.5 0.62

Eucalyptus tricarpa 3 15 0.64

Hibbertia exutacies 2.2 11 0.69

Pultenaea largiflorens 2.9 14.5 0.71

Combining the information on prevalence (p) and the relative costs of false-positive and false-
negative errors allows the calculation of a slope (Table 3):

The appropriate threshold probability is found by moving a line, with the calculated slope,
from the top left of a receiver operating characteristic (ROC) curve (Zweig and Campbell,
1993) to the point where the line and the curve first touch. The ROC curve relates relative
proportions of correctly and incorrectly classified predictions (sensitivity and false-positive
pairs) over a wide and continuous range of threshold probability values. The sensitivity is
plotted on the y-axis and the false-positive rate (or the specificity rate) is plotted on the x-axis.
The sensitivity and specificity values at the point on the ROC curve where the line first
touches is used to determine the appropriate threshold probability, given the relative cost of
false-positive and false-negative errors and the prevalence of the species (Fig. 1). In Fig. 1,
the line with slope 0.75 corresponds to a cost ratio of 1:4 and a prevalence of 0.2. The line
first touches the curve at the point with a sensitivity of 0.85 and specificity of 0.4, which
corresponds to a threshold probability of 0.2. The threshold probabilities chosen for the four
plant species using this approach are provided in Table 3. The threshold chosen for A.
ausfeldii is slightly lower than that for the other species, due to its comparatively low
prevalence.

Fig. 1. The use of a ROC curve and a line with a specified slope to determine a threshold probability for
use on predicted species distribution data. The slope represents the relative cost of false-positive and
false-negative errors and takes into account the prevalence of the species (the proportion of presence
observations in the survey data). The point where the line with this slope first touches the curve is used
to identify the appropriate threshold probability.

We used these three methods of threshold selection to convert the predicted probabilities of
occurrence of the four plant species to presence/absence data. We then scaled the data from
the resolution of grid cells to that of the planning units – the extant remnants of native
vegetation. To achieve this, the presence/absence data were summed across the grid cells
(which represent one hectare) to obtain the predicted number of occurrences of each species
in each planning unit. We take this to represent the predicted area of occurrence of each
species (in hectares) in each planning unit.
2.4. Probabilistic approaches – using the probabilities directly
An alternative approach to using predicted species distribution data in conservation planning
is to use the probabilities of occurrence directly and not convert them to presence/absence
data. We evaluated two approaches to using probabilities of occurrence directly.
First, we used the minimum expected coverage approach (referred to hereafter as Probabilistic
Method 1). The expected area of occurrence (in hectares) of each species in each planning
unit was calculated and the objective was to minimise the area reserved, subject to the
constraint that the expected area of occurrence was greater than the pre-specified targets (ri):

(2)
for each species (i), where pik is the probability that species i exists in grid cell k. The set of
grid cells k in planning unit j is denoted by KJ and J represents the set of planning units j
selected for the reserve network. The control variable xj is equal to zero or one for j = 1, …, J
such that xj equals one if planning unit j is in the reserve network and xj equals zero if
planning unit j is not in the reserve network. As far as we are aware, this is the first time
Probabilistic Method 1 has been used for conservation planning.
Second, we modified the minimum expected coverage approach to generate the
minimum expected threshold coverage approach (referred to hereafter as Probabilistic
Method 2). The thresholds generated using Threshold Method 2 (Table 2), where sensitivity
and specificity were traded, were used to find the grid cells with probabilities of occurrence
greater than the specified threshold allocated to each species. The expected area of occurrence
of each species in each planning unit was calculated, but using only the probabilities of
occurrence greater than the specified thresholds.
In summary, three threshold approaches and two probabilistic approaches were used
to generate a dataset from the predicted species distribution data (Table 4).
Table 4. Summary of the approaches used to generate a dataset from the predicted species
distribution data. These approaches include those that use a threshold to convert the
probabilities of occurrence to presence/absence data and those that use the probabilities of
occurrence directly

Method of generating a dataset

Description
from predicted distribution data

Threshold approaches

1. Convert all probabilities greater than 0.5 to presence/absence

Threshold Method 1
data

2. Sum the converted presence/absence data

1. Find the probability that gives converted presence/absence data

Threshold Method 2
that has the desired specificity

2. Convert all probabilities greater than this threshold to

presence/absence data

3. Sum the converted presence/absence data

1. Find the probability that gives converted presence/absence data

Threshold Method 3 that has the desired relative false-positive and false-negative error
rates

2. Convert all probabilities greater than this threshold to

presence/absence data

3. Sum the converted presence/absence data

Probabilistic approaches

Probabilistic Method 1 1. Sum all probabilities regardless of their value

1. Find the probability that gives converted presence/absence data

Probabilistic Method 2
that has the desired specificity

2. Sum all probabilities greater than this value (without

converting them to presence/absence data)

An alternative approach to using probabilities directly is the minimum threshold coverage

approach. This approach has been used for selecting conservation areas (Margules and
Nicholls, 1987, Williams and Araújo, 2000, Williams and Araújo, 2002, Araújo et al., 2002,
Arthur et al., 2002 and ReVelle et al., 2002). The objective is to minimise the area reserved
subject to the constraint that the overall probability of occurrence of each feature in the
reserve network is greater than a pre-specified target:

(3)

for all species (i), where pij is the probability of occurrence of species i in planning unit j. The
objective function employed for Threshold Methods 1–3 and Probabilistic Methods 1 and 2
was to find the minimum area that meets an area target for each species. The objective
function when using the minimum threshold coverage approach is to find the minimum area
that represents each species with a specified probability (α).
Under this alternative approach, two simplifying assumptions are required to allow
the probability of a species occurring in a reserve network to be expressed as the product of
the probabilities of non-occurrence over all component planning units. The first assumption is
that the probability of a species occurring in a planning unit is independent of the probability
of the species occurring in other planning units. The second assumption is that the probability
of a species occurring in a planning unit is independent of the probability of every other
species occurring in that planning unit. These assumptions are often false because the
distribution of individuals is usually strongly spatially correlated and ecological relationships
often exist between many species. The objective function for the other approaches does not
make these assumptions. In addition, the other approaches avoid the difficulty of setting a
target based on a probability of occurrence. Consequently, we chose not to consider the
minimum threshold coverage approach. Finally, optimising for the altered objective function
is redundant for this dataset. This is not meant as a criticism of the approach. Rather, it
illustrates that finding the minimum area that meets a specified probability of occurrence, can
sometimes be achieved using a simple, exhaustive search. Planning units can simply be added
to the reserve network, starting with the planning units with the highest probability of
occurrence, until the probability of occurrence threshold is met. For example, when
guaranteeing a stringent threshold probability (such as 0.95 for all species), the minimum area
required is approximately 9 hectares based on the Box–Ironbark dataset.
2.5. Conservation planning procedure

We used the simulated annealing algorithm in the Marxan reserve selection software package
(Ball and Possingham, 2000) to find the minimum area required to meet the conservation
targets (Liu and Wang, 1994, Murray and Church, 1996, Csuti et al., 1997, Possingham et al.,
2000 and Arthur et al., 2002). For the simulated annealing procedure, an adaptive schedule
followed by iterative improvement was employed and was configured so that the number of
simulated annealing iterations was 10 million and the number of ‘temperature’ decreases was
10,000. We ran Marxan 100 times to produce 100 solutions. The irreplaceability of each
planning unit was calculated as the proportion of the reserve network solutions that included
each planning unit. Planning units included in all of the reserve network solutions were
considered irreplaceable, as these planning units will be required to meet the conservation
targets for the species they contain. Existing conservation areas were taken into account by
making xj equal one for each planning unit j that was already reserved. For simplicity, we did
not seek to minimise the boundary length of the reserve networks. Generally, stepwise
heuristics have been used to handle probabilistic data in conservation planning (Margules and
Nicholls, 1987, Polasky et al., 2000, Williams and Araújo, 2000 and Araújo et al., 2002). This
is the first time we are aware that simulated annealing has been used to address this problem.
2.6. Assessing the sensitivity of the conservation planning outcomes to the different
approaches to using predicted species distribution data

We investigated the sensitivity of the conservation planning outcomes to the different

approaches for generating datasets from predicted species distribution data by determining:

1. The predicted area of occupancy of each species within the study region.
2. The efficiency of each reserve network solution.
3. The area of irreplaceable planning units in the reserve network solutions and the similarity
of selected planning units.
4. The compactness of each reserve network solution.
5. The expected representation of each species in each reserve network solution.
6. A measure of the performance of different reserve networks is how efficiently targets are
met. Efficiency (which varies from zero to 100) can be calculated via:
 1(X/T)×100, (4)

where X is the area of planning units needed to achieve the targets and T is the total area of
planning units available for conservation (Pressey and Nicholls, 1989). The minimum-set of
planning units needed to meet the conservation targets is the most efficient solution.
Therefore, highly efficient solutions have minimal over-achievement of targets.
The Kappa statistic (Cohen, 1960) provides a measure of the similarity of planning units
selected for the different reserve networks. This statistic assesses the extent that reserve
networks overlap after removing overlap due to chance. The Kappa statistic is:

(5)

where Po is the proportion of planning units that overlap and Pc is the proportion of planning
units that overlap due to chance. The Kappa statistic ranges from a minimum of negative one
to a maximum of one, with a value of one indicating perfect overlap, zero indicating expected
overlap due to chance and negative one indicating no overlap (Czaplewski, 1994).
We used a two by two classification table to calculate the Kappa statistic (Table 5); this cross-
tabulated the area of planning units selected for each reserve network solution.

Table 5. A two by two classification table, where Reserve Network 1 might represent the
network selected using Threshold Method 1 and Reserve Network 2 might represent the
reserve network selected using Probabilistic Method 2 (A + B + C + D = N)

Reserve network 1

Included Not included

Reserve Network 2 Included A B

Not included C D

The Kappa statistic was calculated (after Fielding and Bell, 1997) using the formula:

(6)

We calculated the Kappa statistic for the subset of irreplaceable planning units selected to be
part of each reserve network and for the full set of planning units selected to be part of each
reserve network.
A measure of compactness or degree of spatial clustering of a reserve network is provided by
the ratio of the boundary length of the reserve network and the circumference of a circle of
the same area (Possingham et al., 2000):

(7)

A circle is the most compact shape possible, so this is the ratio of the boundary length to the
theoretical minimum and is a dimensionless measure. Values approaching one resemble the
shape of a circle and are highly compact.
We calculated the probability of at least one occurrence of each species in each planning unit
and summed this across all planning units selected for each reserve network:

(8)

This determined the number of times each species is expected to occur in each reserve
network and therefore provided the expected representation of each species in the reserve
network solutions.

3. Results
3.1. The predicted area of occupancy of each species
First, a comparison is made of the predicted area of occupancy of each species. Using
Threshold Methods 2 and 3, the area where the species is predicted to occur is limited to that
deemed likely to contain the species, according to a chosen threshold. These thresholds are
higher than the 0.5 threshold of Threshold Method 1 and hence the predicted area of
occupancy of each species is less. Using Probabilistic Method 1, the probabilities were
summed across all grid cells, which generally resulted in a greater predicted area of
occupancy than with the other methods. Probabilistic Method 2, where only those
probabilities greater than a specified threshold were summed, resulted in the least predicted
area of occupancy of each species (Fig. 2). Given the small area predicted to be occupied by
the species using Threshold Methods 2 and 3 and Probabilistic Method 2, many of the targets
were unable to be met (Fig. 2).

Fig. 2. The predicted area of occupancy of each species, based on the different methods for generating
a dataset from predicted species distribution data. The conservation target for each species is
highlighted. For many species, the original targets cannot be met, due to insufficient area being
predicted as occupied.

Using Threshold Method 1, the predicted area of occupancy of A. ausfeldii, E. tricarpa and P.
largiflorens was less than with Probabilistic Method 1. The modal probabilities of occurrence
of these three species were about 0.4. When using Threshold Method 1, probabilities <0.5
were converted to zero. Therefore, the predicted area of occupancy of these three species was
deflated and less than with Probabilistic Method 1 where the probabilities were summed. The
modal probability of occurrence for H. exutacies was about 0.6. Therefore, since probabilities
equal to or greater than 0.5 were converted to one when using Threshold Method 1, the
predicted area of occupancy of this species was inflated and greater than with Probabilistic
Method 1.
3.2. Efficiency of the reserve network solutions
The dataset generated using Threshold Method 3 resulted in the reserve network with the
greatest area. Consequently, the reserve network generated on the basis of this dataset was the
least efficient (Table 6). Despite its large area, the reserve network was unable to meet the
original conservation targets for any of the species, due to the small area predicted to be
occupied by each species.

Table 6. Number and mean size (in hectares) of the selected planning units and the area of
the reserve network solutions. The efficiency of the reserve network solutions and the area of
irreplaceable planning units are also detailed

Method of generating Area of

# Mean size of Efficiency Area of
a dataset from reserve
planning planning (area irreplaceable
predicted distribution network
units units (ha) selected) planning units
data (ha)

Threshold Method 1 469 243 114,166 62.7% 81849 (n = 40)

Threshold Method 2 693 201 139,381 54.5% 139,378 (n = 682)

Threshold Method 3 678 220 148,820 51.4% 147,729 (n = 640)

Probabilistic Method 1 1197 97 116,354 62.0% 19098 (n = 42)

Probabilistic Method 2 668 216 144,412 52.8% 144,409 (n = 658)

The dataset generated using Threshold Method 1 resulted in the most efficient reserve
network solution (Table 6). Probabilistic Method 1 met the conservation targets within a
reserve network of comparable size to Threshold Method 1, but required almost twice the
number of planning units (Table 6). Basing the analyses on datasets generated using
Threshold Method 1 and Probabilistic Method 1 also allowed the original targets to be met.
The analyses were repeated using targets that were attainable. The patterns exhibited using
the modified targets were identical to the results obtained using the original targets.
The mean size of the planning units within the study region is 62 ha; planning units range in
size from <1 to 14,711 ha. The reserve network solutions generated using the threshold
approaches and Probabilistic Method 2 were biased towards larger planning units (Table 6).
As the probabilities of occurrence were generally <0.5, smaller planning units were less
valued when a threshold was used on the component probabilities. Larger planning units were
less affected by the use of a threshold because they were more likely to contain at least some
high probabilities of occurrence.
3.3. Similarity of planning units in the reserve network solutions
The reserve network solutions identified using the datasets from Threshold Methods 2 and 3
and Probabilistic Method 2 contained large areas of irreplaceable planning units (Table 6). In
comparison, the reserve network solution identified using Probabilistic Method 1 had the
largest area of replaceable planning units, followed by Threshold Method 1. The irreplaceable
planning units of Probabilistic Method 1 were largely comprised of the existing reserves,
which were locked into the reserve network solutions. When the analyses were repeated using
targets that were attainable, the patterns exhibited were identical.
The irreplaceable planning units selected using the datasets from Threshold Methods 2 and 3
and Probabilistic Method 2 overlapped substantially (Table 7). Those selected using
Threshold Method 1 overlapped to a lesser extent and those selected using Probabilistic
Method 1 overlapped the least. When the analyses were repeated using targets that were
attainable, the patterns exhibited were identical, except that Threshold Method 1 and
Probabilistic Method 1 exhibited perfect overlap. This is because their irreplaceable planning
units comprised only the existing conservation areas, which were locked into the reserve
network solutions.

Table 7. The overlap between irreplaceable planning units in the reserve network solutions
selected using the different datasets. The Kappa statistics was used to assess the overlap
between reserve after removing overlap due to chance

Method of generating a dataset from

Overlap in irreplaceable area
predicted distribution data

Threshold Threshold Threshold Probabilistic

Method 1 Method 2 Method 3 Method 1

Threshold Method 2 0.55

Threshold Method 3 0.49 0.84

Probabilistic Method 1 0.31 0.15 0.14

Probabilistic Method 2 0.52 0.94 0.82 0.14

None of the reserve network solutions identified on the basis of the different datasets
comprised the same set of planning units (Table 8). The set of planning units selected using
Probabilistic Method 1 did not overlap with those selected based on the other datasets (Table
8), with the exception of the reserve network solution identified using Threshold Method 1.
The set of planning units selected using the threshold methods overlap to a moderate extent
(Table 8). The set of planning units selected using Probabilistic Method 2 had considerable
overlap with the reserve network identified using Threshold Methods 2 and 3 (Table 8).
Similar patterns were exhibited when the analyses were repeated using targets that were
attainable.

Table 8. The extent of overlap of planning units in the reserve network solutions selected
using the different datasets. The Kappa statistics was used to assess the overlap between
reserve after removing overlap due to chance

Method of generating a
dataset from predicted Overlap in reserve network area
distribution data

Threshold Threshold Threshold Probabilistic

Method 1 Method 2 Method 3 Method 1

Threshold Method 2 0.61

Threshold Method 3 0.62 0.85

Probabilistic Method 1 0.66 0.49 0.45

Probabilistic Method 2 0.60 0.94 0.83 0.49

3.4. Compactness of the reserve network solutions

The reserve network solution identified using Threshold Method 1 was the most compact
(Table 9). The reserve network solution identified using Probabilistic Method 1 was the least
compact and this is due to the selection of smaller planning units to form the reserve network
solutions compared to the other approaches.

Table 9. Compactness of the reserve network solutions identified using the different datasets

Method of generating a dataset from Boundary Area Compactness of the

predicted distribution data length (km) (ha) reserve network

Threshold Method 1 2,901,491 114,166 24

Threshold Method 2 3,952,810 139,381 30

Threshold Method 3 4,186,473 148,820 31

Probabilistic Method 1 4,139,755 116,354 34

Probabilistic Method 2 4,057,581 144,412 30

3.5. Expected representation of each species in the reserve network solutions

The probability of at least one occurrence of each species in each planning unit was summed
across all planning units selected for the different reserve network solutions. This provided
the number of planning units in the reserve network solutions in which each species is
expected to occur.
Initially, Probabilistic Method 2 provided the greatest expected representation of each species,
closely followed by Threshold Method 2 and Threshold Method 3 (Table 10). However, these
methods also resulted in the largest reserve network solutions: any approximation to a
minimum-set problem that selects a larger area can include more representations of species
simply by including more area. The results were therefore standardised for reserve network
area. This made little difference to the results. Subsequent to standardisation, the approaches
that involved the use of a threshold (Threshold Methods 2 and 3 and Probabilistic Method 2)
had equally great expected species representation (Table 10). Similar patterns were exhibited
when the analyses were repeated using targets that were obtainable.
Table 10. Expected species representation in the reserve network solutions identified using
the different datasets

Method of generating a dataset Mean expected species Mean expected species representation
from predicted distribution data representation (standardised for reserve network area)

Threshold Method 1 108.19 0.0008

Threshold Method 2 568.53 0.004

Threshold Method 3 564.51 0.004

Probabilistic Method 1 45.10 0.0003

Probabilistic Method 2 571.49 0.004

4. Discussion
We have shown that conservation planning outcomes are very sensitive to the uncertainty that
arises from the different methods to generate a dataset from predicted species distribution
data. The reserve network solutions differ in their relative efficiencies, their component
planning units, and their expected representation of species. The method of generating a
dataset from predicted species distribution data for use in conservation planning will reflect
how much risk a planner is willing to tolerate and the amount of efficiency that can be
sacrificed. A planner who is risk-averse would seek greater certainty that the target species
occur in the planning units selected for a reserve network (Terborgh and Winter, 1983,
Williams, 1998 and Elith and Burgman, 2003).
The approaches used to generate a dataset from predicted species distribution data for
use in conservation planning may be grouped into two broad categories: (1) those that convert
the probabilities of occurrence to presence/absence data using thresholds identified a priori or
a posteriori, and (2) those that use the probability of occurrence data directly. Three
thresholds were compared: a threshold of 0.5 selected a priori and two thresholds selected a
posteriori; one of these trades the sensitivity and specificity of the datasets, the other trades
the cost of false-positive and false-negative errors.
A priori methods of threshold selection are commonly used (Li et al., 1997, Manel et
al., 1999, Manel et al., 2001, Fleishman et al., 2001 and Fleishman et al., 2003), despite the
fact that they take no account of the range and spread of predictions, nor of dataset
characteristics and user requirements. Stringent thresholds were chosen a posteriori to ensure
that only potentially suitable habitat was predicted as occupied habitat. Therefore, the a
posteriori chosen thresholds were more risk-averse than the threshold chosen a priori.
The approaches to using probabilistic data directly, which have not been used previously for
conservation planning, employ information on the expected area of occurrence of each
species in each planning unit. The second approach differs from the first in that only the
probabilities of occurrence above a stringent threshold are used in determining the expected
area of occurrence. Therefore, the second approach is more risk-averse.
The various methods of generating a dataset from predicted species distribution data
resulted in different predictions of the area occupied by each species. When the probabilities
of occurrence are summed, the predicted area of occupancy is greater than when a stringent
threshold is used to first convert the probabilities to presence/absence data. This result may
vary for different modelling methods and different species; nevertheless, it is likely to be
partly a consequence of the datasets being biased towards zero (due to the low prevalence of
the species in the survey data). The calculation of the predicted area of occupancy of a species
is also dependent on the resolution of the underlying data. One hectare grid cells were used as
the base unit for predictions, because this was a reasonable representation of the variation in
the landscape and the accuracy of the species survey data. There is, however, a trade-off
between smaller grid cells, which might deliver more accurate predictions of the area
occupied by a species, but with a likely increase in computation time. Larger grid cells might
result in the overestimation of the area occupied by a species by including unsuitable habitat
and hence create more uncertainty in the modelled predictions. To help mitigate these
problems, the sensitivity of the area of occupancy predictions to the resolution of the data
could be assessed and species-specific resolutions for estimating area of occupancy could be
employed (Keith, 1998 and Keith et al., 2000).
The use of a stringent threshold on the predicted species distribution data helps to
ensure that only highly suitable habitat is counted towards the predicted area of occupancy of
each species. When using a low threshold or summing the probabilities without using a
threshold, the predicted area of occupancy can also include areas with a very low probability
of occurrence and is therefore a more risk-neutral approach.
With only a small area predicted to contain each species, the datasets generated using the
stringent thresholds offered less flexibility in designing a reserve network. Consequently, the
reserve networks identified on the basis of these datasets covered a large area and were
comparatively less efficient. The reserve network based on the summed probabilities had the
largest area of replaceable planning units. This reflects the flexibility available to generate
this reserve network solution.
While the reserve network solutions identified using the different datasets overlapped,
they still differed from each other. Most importantly, there was limited overlap in
irreplaceable planning units. Therefore, depending on how predicted species distribution data
is used for conservation planning, different sets of planning units will be selected to be part of
the reserve network, particularly if irreplaceability is used to determine the planning units that
are the immediate priorities for conservation.
 The reserve networks identified using the datasets generated with a stringent
threshold (the more risk-averse approaches) had the greatest expected species representation
when standardised for the area of the reserve network. The expected representation of species
in reserve networks is one procedure for evaluating different reserve network solutions, in
terms of how well they might conserve species (their anticipated adequacy). An alternative,
more data intensive, approach to testing adequacy would be to evaluate the predicted
persistence of the species in the reserve network solutions using population viability analysis.
In addition to having the greatest expected species representation, the reserve network
solutions identified using the more risk-averse approaches also favoured large planning units,
which may provide better outcomes for nature conservation due to reduced boundary length
and edge effects (Fagan et al., 1999). However, the reserve networks based on these datasets
were not the most efficient. Therefore, the method of generating a dataset from predicted
species distribution data for use in conservation planning will reflect a trade-off between the
risk of inadequately conserving biodiversity and the amount of efficiency that can be
sacrificed.
The efficiency of a reserve network has two implications. First, due to competition
between incompatible land uses, limited resources for conservation and financial
considerations (e.g. costs of acquisition, management and opportunity costs to other land
uses) efficiency determines the likelihood of achieving conservation targets. Second,
efficiency determines the degree to which reserve network proposals are defensible in the
light of competing land uses. It is important to recognise, however, that ‘efficiency’ simply
refers to the area of land required to achieve the conservation targets. The use of a lower
threshold to convert probabilities of occurrence to presence/absence data might increase the
predicted area of occupancy of a species and provide greater flexibility for the design of a
reserve network (and hence greater efficiency). However, the use of a low threshold might
also result in areas being falsely predicted as occupied and increase the uncertainty associated
with the use of predicted species distribution data for conservation planning.
Predictions of species distribution are essential for conservation planning because,
even in well-surveyed parts of the world, sample plots cover only a small proportion of the
landscape. Conservation planning has to deal with the whole of the landscape. Used
appropriately, predicted species distribution data can help decision-makers understand how
best to use limited resources to achieve conservation targets. However, it is important that the
uncertainty in predicted distribution data be accounted for in conservation planning given the
sensitivity of the conservation planning outcomes to this source of uncertainty.
Recently, the probability of a species occurring in an area has been equated with the long-
term persistence of a species in that area (Williams and Araújo, 2000 and Araújo et al., 2002).
Probabilities of occurrence may not indicate habitat quality for a species and may not be
related to species persistence. This is because the probabilities are based on static species
locality data that does not take into account population processes. Consequently, in generating
these probabilities, it is assumed that the population is at equilibrium (Guisan and
Zimmermann, 2000). If the population is increasing (for example, recovering) then there will
be many false-negative predictions and the quality of the habitat and persistence of a species
may be underestimated. Alternatively, if the population is declining (for example, confined to
relict areas) many false-positive predictions may result, in which case the quality of the
habitat and persistence of a species is likely to be overestimated. In either circumstance, an
inaccurate assessment of habitat quality will result and estimates of persistence are likely to
be unreliable.
Even if the population of a species is at equilibrium, it may not occupy, let alone
persist, in habitat predicted to have a high probability of occurrence due to behaviour,
intraspecific competitive exclusion, or dispersal dynamics (Van Horne, 1983). In the Box–
Ironbark region, altered fire regimes will influence the distribution of plants. For example, A.
ausfeldii is an obligate seeder that regenerates after high soil temperatures, resulting from fire,
break the dormancy of its seeds (Brown et al., 2003). Therefore, if the frequency of fire is
reduced, A. ausfeldii may not occupy habitat in which it was predicted to have high
probability of occurrence. Further, meta-population theory tells us that due to colonisation-
extinction dynamics, high quality habitat may be unoccupied at the time a species survey is
conducted.
This research has demonstrated that planning outcomes are very sensitive to the
different methods for generating a dataset from predicted species distribution data. A natural
further step is to direct effort towards producing the most reliable predictions for use in
conservation planning, and to find the reserve network that is most robust to the uncertainty in
the predictions. Uncertainty is not only related to the use of the probabilities generated from a
model of species distribution, but also to the modelling method and the underlying data (Elith
et al., 2002). Modelled predictions might be made more reliable, for example, by re-
calibrating the model so it predicts more accurately to new cases (Steyerberg et al., 2001).
Uncertainty might also be reduced by producing more ecologically realistic models (Austin,
2002). Nevertheless, there will always be uncertainty associated with the use of predicted
species distribution data for conservation planning. As a consequence on-ground inspection of
sites nominated for conservation will likely be required.
Acknowledgements
This research was supported by grants to the lead author from the University of Melbourne (Australia), the
Menzies Centre for Australian Studies (Kings College, London, UK), the Australian Federation of University
Women, and the Holsworth Wildlife Research Fund (Australia). We also would like to thank Mark Burgman,
Adrian Newton, Daniel Wilson and Bob Pressey for earlier comments on the manuscript.
References
Agresti, 1996 A. Agresti, An Introduction to Categorical Data Analysis, John Wiley & Sons, New York (1996).
Araújo et al., 2002 M.B. Araújo, P.H. Williams and R.J. Fuller, Dynamics of extinction and the selection of nature reserves,
Proceedings of the Royal Society of London. Series B Biological Sciences 269 (2002), pp. 1971–1980.
Arthur et al., 2002 J.L. Arthur, R.G. Haight, C.A. Montgomery and S. Polasky, Analysis of the threshold and expected coverage
approaches to the probabilistic reserve site selection problem, Environmental Modeling and Assessment 7 (2002), pp. 81–89.
Austin, 1998 M.P. Austin, An ecological perspective on biodiversity investigations: examples from Australian eucalypts, Annals
of the Missouri Botanical Garden 85 (1998), pp. 2–17.
Austin, 2002 M.P. Austin, Spatial prediction of species distribution: an interface between ecological theory and statistical
modeling, Ecological Modelling 157 (2002), pp. 101–118.
Ball and Possingham, 2000 I.R. Ball and H.P. Possingham, Marxan (v 1.8.6): Marine reserve design using spatially explicit
annealing, A manual prepared for the Great Barrier Reef Marine Park Authority (2000).
Brown et al., 2003 J. Brown, N.J. Enright and B.P. Miller, Seed production and germination in two rare and three common co-
occurring Acacia species from south-east Australia, Austral Ecology 28 (2003), pp. 271–280.
Burgman et al., 2001 M.A. Burgman, H.P. Possingham, A.J.J. Lynch, D.A. Keith, M.A. Mccarthy, S.D. Hopper, W.L. Drury,
J.A. Passioura and R.J. Devries, A method for setting the size of plant conservation target areas, Conservation Biology 15 (2001),
pp. 603–616.
Camm et al., 2002 J. Camm, S. Norman, S. Polasky and A. Solow, Nature reserve site selection to maximize expected species
covered, Operations Research 50 (2002), pp. 946–955.
Church et al., 1996 R.L. Church, D.M. Stoms and F.W. Davis, Reserve selection as a maximal covering location problem,
Biological Conservation 76 (1996), pp. 105–112.
Clark and Slusher, 2000 F.S. Clark and R.B. Slusher, Using spatial analysis to drive reserve design: a case study of a notional
wildlife refuge in Indiana and Illinois (USA), Landscape Ecology 15 (2000), pp. 75–84.
Cohen, 1960 J. Cohen, A coefficient of agreement for nominal scales, Educational and Psychological Measurement 20 (1960),
pp. 37–46.
Cowling and Pressey, 2003 R.M. Cowling and R.L. Pressey, Introduction to systematic conservation planning in the Cape
Floristic Region, Biological Conservation 112 (2003), pp. 1–13.
Csuti et al., 1997 B. Csuti, S. Polasky, P.H. Williams, R.L. Pressey, J.D. Camm, M. Kershaw, R. Kiester, B. Downs, R.
Hamilton, M. Huso and K. Sahr, A comparison of reserve selection algorithms using data on terrestrial vertebrates in Oregon,
Biological Conservation 80 (1997), pp. 83–97.
Czaplewski, 1994 Czaplewski, R.L., 1994. Variance approximations for assessments of classification accuracy. Rocky Mountain
Forest and Range Experiment Station, USDA Forest Service, Fort Collins.
Department of Sustainability and Environment, 2002 Department of Sustainability and Environment, 2002. Flora Information
System – May 2002. Viridans Biological Databases, Brighton East, Victoria.
Elith and Burgman, 2003 J. Elith and M.A. Burgman, Chapter 8: Habitat models for PVA In: C.A. Brigham and M.W. Schwartz,
Editors, Population Viability in Plants, Springer-Verlag, New York (2003).
Elith et al., 2002 J. Elith, M.A. Burgman and H.M. Regan, Mapping epistemic uncertainties and vague concepts in predictions of
species distribution, Ecological Modelling 157 (2002), pp. 313–329.
Fagan et al., 1999 W.F. Fagan, R.S. Cantrell and C. Cosner, How habitat edges change species interactions, The American
Naturalist 153 (1999), pp. 165–182.
Ferdana, 2002 Z., Ferdana, Approaches to integrating a marine GIS into The Nature Conservancy’s ecoregional planning process
In: J. Breman, Editors, Marine Geography: GIS for Oceans and Seas, ESRI, Redlands, WA (2002), pp. 151–158.
Ferrier et al., 2000 S. Ferrier, R.L. Pressey and T.W. Barrett, A new predictor of the irreplaceability of areas for achieving a
conservation goal, its application to real-world planning, and a research agenda for further refinement, Biological Conservation
93 (2000), pp. 303–325.
Fielding and Bell, 1997 A.H. Fielding and J.F. Bell, A review of methods for the assessment of prediction errors in conservation
presence/absence models, Environmental Conservation 24 (1997), pp. 38–49.
Fleishman et al., 2001 E. Fleishman, R. Mac Nally, J.P. Fay and D.D. Murphy, Modeling and predicting species occurrence
using broad scale environmental variables: an example with butterflies of the Great Basin, Conservation Biology 15 (2001), pp.
1674–1685.
Fleishman et al., 2003 E. Fleishman, R. Mac Nally and J.P. Fay, Validation tests of predictive models of butterfly occurrence
based on environmental variables, Conservation Biology 17 (2003), pp. 806–817.
Funk and Richardson, 2002 V.A. Funk and K.S. Richardson, Systematic data in biodiversity studies: use it or lose it, Systematic
Biology 51 (2002), pp. 303–316.
Great Barrier Reef Marine Park Authority, 2003 Great Barrier Reef Marine Park Authority, 2003. Representative Areas in the
Marine Park. Available from: http://www.gbrmpa.gov.au/corp_site/key_issues/conversation/rep_areas/?rep_areasoverview. html.
Groombridge and Jenkins, 2002 B. Groombridge and M.D. Jenkins, Global biodiversity: responding to the change In: B.
Groombridge and M.D. Jenkins, Editors, World Atlas of Biodiversity: Earth’s Living Resources in the 21st Century, University
of California Press, Berkeley, CA (2002), pp. 195–223.
Guisan and Zimmermann, 2000 A. Guisan and N.E. Zimmermann, Predictive habitat distribution models in ecology, Ecological
Modelling 135 (2000), pp. 147–186.
Haight et al., 2000 R.G. Haight, C.S. ReVelle and S.A. Snyder, An integer optimization approach to the probabilistic reserve site
selection problem, Operations Research 48 (2000), pp. 697–708.
Haila and Margules, 1996 Y. Haila and C.R. Margules, Survey research in conservation biology, Ecography 19 (1996), pp. 323–
331. Abstract-GEOBASE
Henderson, 1992 N. Henderson, Wilderness and the nature conservation ideal: Britain, Canada and the United States contrasted,
Ambio 21 (1992), pp. 394–399.
Hosmer et al., 1988 D.W. Hosmer, S. Lemeshow and J. Klar, Goodness-of-fit testing for the logistic regression model when the
estimated probabilities are small, Biometrical Journal 30 (1988), pp. 911–924.
JANIS, 1997 JANIS, 1997. Nationally agreed criteria for the establishment of a comprehensive, adequate and representative
reserve system for forests in Australia. Joint ANZECC/MCFFA National Forest Policy Statement Implementation Sub-
committee. National forest conservation reserves: Commonwealth proposed criteria. Commonwealth of Australia, Canberra.
Jennings, 2000 M.D. Jennings, Gap analysis: concepts, methods, and recent results, Landscape Ecology 15 (2000), pp. 5–20.
Justus and Sarkar, 2002 J. Justus and S. Sarkar, The principle of complementarity in the design of reserve networks to conserve
biodiversity: a preliminary history, Journal of Biosciences 27 (2002), pp. 421–435.
Keith, 1998 D.A. Keith, An Evaluation and Modification of World Conservation Union Red List Criteria for Classification of
Extinction Risk in Vascular Plants, Conservation Biology 12 (1998), pp. 1076–1090.
Keith et al., 2000 D.A. Keith, T.D. Auld, M.K.J. Ooi and B.D.E. Mackenzie, Sensitivity analyses of decision rules in World
Conservation Union (IUCN) Red List criteria using Australian plants, Biological Conservation 94 (2000), pp. 311–319.
Keller and Scallan, 1999 C.M.E. Keller and J.T. Scallan, Potential roadside biases due to habitat changes along breeding bird
survey routes, Condor 101 (1999), pp. 50–57.
Kelley et al., 2002 C. Kelley, J. Garson, A. Aggarwal and S. Sarkar, Place prioritization for biodiversity reserve network design:
a comparison of the SITES and ResNet software packages for coverage and efficiency, Diversity and Distributions 8 (2002), pp.
297–306.
Kirkpatrick, 1983 J.B. Kirkpatrick, An iterative method for establishing priorities for the selection of nature reserves: an example
from Tasmania, Biological Conservation 25 (1983), pp. 127–134.
Kirkpatrick and Brown, 1991 J.B. Kirkpatrick and M.J. Brown, Planning for species conservation In: D.A. Saunders, G.W.
Arnold, A.A Burbidge and A.J.M. Hopkins, Editors, Nature Conservation: The Role of Remnants of Native Vegetation, Surrey
Beatty and Sons Pty Ltd in association with CSIRO and CALM, Sydney (1991), pp. 83–89.
Kirkpatrick and Brown, 1994 J.B. Kirkpatrick and M.J. Brown, A comparison of direct and environmental domain approaches to
planning reservation of forest higher plant communities and species in Tasmania, Conservation Biology 8 (1994), pp. 217–224.
Li et al., 1997 W. Li, Z. Wang, Z. Ma and H. Tang, A regression model for the spatial distribution of red-crown crane in
Yancheng Biosphere Reserve, China, Ecological Modelling 103 (1997), pp. 115–121.
Liu and Wang, 1994 C.M. Liu and A.H. Wang, Solving location-allocation problems with rectilinear distances by simulated
annealing, Journal of the Operational Research Society 45 (1994), pp. 1304–1315.
Manel et al., 1999 S. Manel, J.M. Dias, S.T. Buckton and S.J. Ormerod, Alternative methods for predicting species distribution:
an illustration with Himalayan river birds, Journal of Applied Ecology 36 (1999), pp. 734–747.
Manel et al., 2001 S. Manel, H.C. Williams and S.J. Ormerod, Evaluating presence-absence models in ecology: the need to
account for prevalence, Journal of Applied Ecology 38 (2001), pp. 921–931.
Margules and Nicholls, 1987 C.R. Margules and A.O. Nicholls, Assessing the conservation value of remnant habitat islands:
Mallee patches on the western Eyre Peninsula, South Australia In: D.A. Saunders, G.W. Arnold, A.A Burbidge and A.J.M.
Hopkins, Editors, Nature Conservation: The Role of Remnants of Native Vegetation, Surrey Beatty and Sons Pty Ltd in
association with CSIRO and CALM, Sydney (1987), pp. 89–102.
Margules and Pressey, 2000 C.R. Margules and R.L. Pressey, Systematic conservation planning, Nature 405 (2000), pp. 243–
253.
Margules and Stein, 1989 C.R. Margules and J.L. Stein, Patterns in the distribution of species and the selection of nature
reserves: An example from Eucalyptus forest in south eastern New South Wales, Biological Conservation 50 (1989), pp. 219–
238.
Margules and Usher, 1981 C.R. Margules and M.B. Usher, Criteria used in assessing wildlife conservation potential: a review,
Biological Conservation 21 (1981), pp. 79–109.
Margules et al., 1988 C.R. Margules, A.O. Nicholls and R.L. Pressey, Selecting Networks of Reserves to Maximise Biological
Diversity, Biological Conservation 43 (1988), pp. 63–76.
Mark, 1985 A.F. Mark, The botanical component of conservation in New Zealand, New Zealand Journal of Botany 23 (1985),
pp. 789–810.
McCullagh and Nelder, 1989 P. McCullagh and J.A. Nelder, Generalized Linear Models, Chapman and Hall, London (1989).
McNeely, 1994 J.A. McNeely, Protected areas for the 21st Century: working to provide benefits for society, Biodiversity and
Conservation 3 (1994), pp. 3–20.
Muir et al., 1995 A.M. Muir, S.A. Edwards and M.J. Dickins, Description and Conservation Status of the Vegetation of the Box–
Ironbark Ecosystem in Victoria, Department of Conservation and Natural Resources, East Melbourne, Victoria (1995).
Murray and Church, 1996 A.T. Murray and R.L. Church, Applying simulated annealing to location-planning models, Journal of
Heuristics 2 (1996), pp. 31–53.
Podger et al., 1990 F.D. Podger, D.C. Mummery, C.R. Palzer and M.J. Brown, Bioclimatic analysis of the distribution of damage
to native plants in Tasmania by Phytophthora cinnamomi, Australian Journal of Ecology 15 (1990), pp. 281–289.
Polasky et al., 2000 S. Polasky, J.D. Camm, A.R. Slow, B. Csuti, D. White and R. Ding, Choosing reserve networks with
incomplete species information, Biological Conservation 94 (2000), pp. 1–10.
Polasky et al., 2001 S. Polasky, J.D. Camm and B. Garber-Yonts, Selecting biological reserves cost effectively: an application to
terrestrial vertebrate conservation in Oregon, Land Economics 77 (2001), pp. 68–78.
Possingham et al., 1993 H. Possingham, J. Day, M. Goldfinch and F. Salzborn, The mathematics of designing a network of
protected areas for conservation In: D. Sutton, E. Cousins and C. Pearce, Editors, Decision Sciences: Tools for Today, 12th
Australian Operations Research Conference, ASOR, University of Adelaide, Adelaide (1993), pp. 536–545.
Possingham et al., 2000 H. Possingham, I. Ball and S. Andelman, Mathematical methods for identifying representative reserve
networks In: S. Ferson and M. Burgman, Editors, Quantitative Methods for Conservation Biology, Springer-Verlag, New York
(2000), pp. 291–305.
Pressey, 1993 R.L. Pressey, The good and the bad news for conservation planning: procedures for reserve selection are
improving but we still don’t know enough about birds and other fauna In: C.P. Catterall, P.V. Driscoll, K. Hulsman, D. Muir and
A. Taplin, Editors, Birds and their Habitats: Status and Conservation in Queensland, Queensland Ornithological Society, St
Lucia, Queensland (1993), pp. 146–156.
Pressey, 1994 R.L. Pressey, Ad hoc reservations: forward or backward steps in developing representative reserve systems?,
Conservation Biology 8 (1994), pp. 662–668.
Pressey, 1997 R.L. Pressey, Priority conservation areas: towards an operational definition for regional assessments In: J.J.
Pigram and R.C. Sundell, Editors, National Parks and Protected Areas: Selection, Delimitation, and Management, Centre for
Water Policy Research, Armidale, New South Wales (1997), pp. 337–357.
Pressey, 1998 R.L. Pressey, Algorithms, politics and timber: an example of the role of science in a public, political negotiation
process over new conservation areas in production forests In: R.T. Wills and R.J. Hobbs, Editors, Ecology for Everyone:
Communicating Ecology to Scientists, the Public and the Politicians, Chipping Norton, Surrey Beatty and Sons, NSW (1998),
pp. 73–87.
Pressey, 2002 R.L. Pressey, The first reserve selection algorithm – a retrospective on Jamie Kirkpatrick’s 1983 paper, Progress
in Physical Geography 26 (2002), pp. 434–441.
Pressey and Nicholls, 1989 R.L. Pressey and A.O. Nicholls, Efficiency in conservation planning: scoring versus iterative
approaches, Biological Conservation 50 (1989), pp. 199–218.
Pressey and Tully, 1994 R.L. Pressey and S.L. Tully, The cost of ad hoc reservation: a case study in western New South Wales,
Australian Journal of Ecology 19 (1994), pp. 375–384.
Pressey et al., 1996 R.L. Pressey, H.P. Possingham and C.R. Margules, Optimality in reserve selection algorithms: when does it
matter and how much?, Biological Conservation 76 (1996), pp. 259–267.
Pressey et al., 1997 R.L. Pressey, H.P. Possingham and J.R. Day, Effectiveness of alternative heuristic algorithms for identifying
indicative minimum requirements for conservation reserves, Biological Conservation 80 (1997), pp. 207–219.
Pressey et al., 2000 R.L. Pressey, T.C. Hager, K.M. Ryan, J. Schwarz, S. Wall, S. Ferrier and P.M. Creaser, Using abiotic data
for conservation assessments over extensive regions: quantitative methods applied across New South Wales, Australia,
Biological Conservation 96 (2000), pp. 55–82.
Purdie et al., 1986 R.W. Purdie, R. Blick and M.P. Bolton, Selection of a conservation reserve network in the mulga
biogeographic region of south-western Queensland, Australia, Biological Conservation 38 (1986), pp. 369–384.
ReVelle et al., 2002 C.S. ReVelle, J.C. Williams and J.J. Boland, Counterpart models in facility location science and reserve
selection science, Environmental Modeling and Assessment 7 (2002), pp. 71–80.
Richardson and Funk, 1999 K.S. Richardson and V.A. Funk, An approach to designing a systematic protected area system in
Guyana, Parks 9 (1999), pp. 7–10.
Rouget et al., 2003 M. Rouget, D.M. Richardson and R.M. Cowling, The current configuration of protected areas in the Cape
Floristic Region, South Africa–reservation bias and representation of biodiversity patterns and processes, Biological
Conservation 112 (2003), pp. 129–145.
Satersdal et al., 1993 M. Satersdal, J.M. Line and H.J.B. Birks, How to maximise biological diversity in nature reserve selection:
vascular plants and breeding birds in deciduous woodlands, western Norway, Biological Conservation 66 (1993), pp. 131–138.
Scott et al., 1993 J.M. Scott, F. Davis, B. Csuti, R. Noss, B. Butterfield, C. Groves, H. Anderson, S. Caicco, F. d’Erchia, T.C.
Edwards, J. Ulliman and R.G. Wright, Gap Analysis: a geographic approach to protection of biological diversity, Wildlife
Monograph 123 (1993), pp. 1–41.
Smith and Theberge, 1986 P.G.R. Smith and J.B. Theberge, A review of criteria for evaluating natural areas, Environmental
Management 10 (1986), pp. 715–734.
Smith and Theberge, 1987 P.G.R. Smith and J.B. Theberge, Evaluating natural areas using multiple criteria: theory and practice,
Environmental Management 11 (1987), pp. 447–460.
Steyerberg et al., 2001 E.W. Steyerberg, F.E. Harrell Jr., G.J.J.M. Borsboom, M.J.C. Eijkemans, Y. Vergouwe and J.D.F.
Habbema, Internal validation of predictive models: Efficiency of some procedures for logistic regression analysis, Journal of
Clinical Epidemiology 54 (2001), pp. 774–781.
Swets, 1988 J.A. Swets, Measuring the accuracy of diagnostic systems, Science 240 (1988), pp. 1285–1293.
Terborgh and Winter, 1983 J. Terborgh and B. Winter, A method for siting parks and reserves with special reference to
Columbia and Ecuador, Biological Conservation 27 (1983), pp. 45–58.
Usher, 1986 M.B. Usher, Wildlife Conservation Evaluation, Chapman and Hall, London (1986).
Van Horne, 1983 B. Van Horne, Density as a misleading indicator of habitat quality, Journal of Wildlife Management 47 (1983),
pp. 893–901.
Williams, 1998 P.H. Williams, Key sites for conservation: area-selection methods for biodiversity In: G.M. Mace, A. Balmford
and J.R. Ginsberg, Editors, Conservation in a Changing World. Integrating Processes into Priorities for Action, Cambridge
University Press, Cambridge (1998), pp. 211–240.
Williams and Araújo, 2000 P.H. Williams and M.B. Araújo, Using probability of persistence to identify important areas for
biodiversity conservation, Proceedings of the Royal Society of London Series B Biological Sciences 267 (2000), pp. 1959–1966.
Williams and Araújo, 2002 P.H. Williams and M.B. Araújo, Apples, oranges, and probabilities: Integrating multiple factors into
biodiversity conservation with consistency, Environmental Modeling and Assessment Special Issue: Reserve Design Modeling 7
(2002), pp. 139–151.
Wilson, 2003 Wilson, K.A., 2003. Uncertainty and vulnerability in conservation planning. PhD Thesis. Institute of Land and
Food Resources, The University of Melbourne.
Woinarski et al., 1996 J.C.Z. Woinarski, O. Price and D.P. Faith, Application of a taxon priority system for conservation
planning by selecting areas which are most distinct from environments already reserved, Biological Conservation 76 (1996), pp.
147–159.
Zweig and Campbell, 1993 M.H. Zweig and G. Campbell, Receiver-Operating Characteristic (ROC) plots: a fundamental
evaluation tool in clinical medicine, Clinical Chemistry 39 (1993), pp. 561–577.