home
DAVID M. STOMS, FRANK DAVIS, and
SANDY ANDELMAN
University of California, Santa Barbara
Problem Statement
Gap analysis identifies the current level of representation of land
cover types and vertebrate species in designated biodiversity man-
agement areas. That is, it helps define the identity and magnitude
of the gaps. The next step is to work toward filling the gaps, which
typically includes identifying a network of nature reserves. Re-
serve network design is a hard problem, however, because the num-
ber of conservation elements and planning units is large (typically
hundreds of elements and hundreds to thousands of sites). Over the
past 15 years researchers have developed computer-based ap-
proaches to make the reserve selection process more systematic
and explicit. These approaches respond to the perceived need for
reserve siting to be as efficient or cost-effective as possible, given
the competing social and economic demands for land and resources.
They also address the concern that reserve system design should be
repeatable, so that the reserve systems can be readily reevaluated
and modified over time as conditions change and new information
is acquired. These approaches assist planners in sorting through
the large volume of data to identify good initial solutions to this
“hard and wicked” problem. However, at present these models re-
main primarily research tools, beyond the reach of most agencies
and nongovernmental organizations that plan and implement re-
serve networks.
Recently, The Nature Conservancy (TNC) began a new planning
initiative with the aim of developing “portfolios” of conservation
sites for each ecoregion in the U.S., the Caribbean, and Latin
America that collectively conserve viable examples of all native
species and plant communities. Both the use of ecologically de-
fined planning regions and the adoption of biotic representation as
an explicit conservation objective posed many new institutional,
scientific, and technical challenges to TNC, which historically has
operated on a state-by-state basis and has focused on rare and threat-
ened species and plant communities.
Based on discussions with various TNC staff involved with
ecoregional planning efforts, none of the existing reserve selection
approaches, in their current form, was well-suited to TNC’s plan-
ning needs. From TNC’s perspective, the major limitations of cur-
rent tools fall into three general categories:
1. those that require high-end computing power, specialized soft-
ware, and/or a high level of technical GIS expertise;
2. those with insufficient documentation and/or inadequate test-
ing of computer code; or
3. those with overly simplistic decision rules (e.g., “greedy” or
rarity-based heuristics).
In 1998, TNC funded a team of investigators (Frank Davis, Sandy
Andelman, and David Stoms) at the University of California-Santa
Barbara to:
• Develop a conservation siting model for regional conservation
analysis that explicitly incorporates spatial design criteria into
the portfolio design process.
• Test and apply the new modeling approach in a structured deci-
sion process involving local TNC staff to develop hypothetical
conservation portfolios in two different ecoregions.
• Produce a training manual, including worked examples using
TNC data and ecoregions, and train TNC staff in applying the
regional conservation-planning tools.
The result is a “toolbox” that can solve ecoregion-sized planning
problems on a Windows-based personal computer. The heart of the
toolbox is a portfolio design model, written by Ian Ball and Hugh
Possingham from the University of Adelaide in Australia. This soft-
ware, based on a simulated annealing algorithm, has been integrated
with a GIS (ArcView) interface for visualizing potential portfolios
and allocating sites prior to running the model. The model has
already been applied in two TNC ecoregions—the Idaho Batholith
and the Northern Sierra Nevada. GAP data were used in both cases
to set some of the representation goals. In addition, the model pro-
vides options for influencing the spatial configuration of an alter-
native portfolio. The toolbox makes it relatively easy for TNC staff
to explore a range of alternatives and identify the effects of their
choices in representation goals, assumptions about costs, impor-
tance of spatial clustering of selected sites, and so on. A planning
team must still review the initial solutions and modify them using
local knowledge, judgment, and other evidence not considered in
the modeling approach.
The Toolbox
Ecoregional planning entails identifying a set of sites that collec-
tively capture viable examples of all native species and communi-
ties from among a larger set of “planning units” within the ecoregion.
The first step in applying the toolbox is to determine which conser-
vation elements, such as cover types, are to be represented in the
portfolio and at what level of representation. It is also possible to
weight some elements as more important than others.
Representa-
tion goals can be the specified number of population occurrences
of a species or total area of a land cover type that must be included
in the regional portfolio to fill the gaps.
The second step is to delineate and characterize the set of planning
units in an ecoregion. There is no simple answer to what are the
right spatial units to use as planning units. It is a complicated ques-
tion, influenced by the size of the ecoregion, the primary ecologi-
cal processes, data sources and resolution, and political issues. We
have frequently used watersheds, at various hierarchical divisions,
as planning units. Whatever the choice of units, data must be com-
piled about several variables for each unit. Data are required for
the composition of cover types (such as GAP data) or other
biodiversity elements and are optional for costs, length of the bound-
aries of planning units, and preallocations of a starting portfolio
(sites required to be in or out of the portfolio). The compositional
information is used by the toolbox to determine each planning unit’s
potential contribution toward meeting the representation goals. The
cost and boundary information is used in minimizing portfolio cost
and in controlling spatial clustering of the portfolio sites. Costs
can be a function of area or any other units you like.
Defining the starting portfolio is an important consideration, and
generally planners will want to consider several alternatives. For
example, they may consider an alternative that starts by fixing ex-
isting national parks and other formally designated reserves. Ele-
ment occurrences in these areas will automatically be counted to-
wards meeting the representation goals. A different alternative
would be to create a portfolio without assuming any starting re-
serves or to consider existing public reserves plus other planning
units that are known to have significant biological resources and
that planners want to make sure are included in the final portfolio.
This step can be done with a word processor or through the ArcView
selection tools. In each alternative the toolbox will then fill any
remaining gaps.
The model seeks to minimize Total Portfolio Cost as measured by a
weighted sum of actual cost (or area) of the selected set of sites, a
penalty for not meeting representation goals for elements, and a
measure of spatial compactness and connectivity. The actual solu-
tions depend on how site cost is measured, on the target levels and
the penalty cost for each element (these are set separately for each
element), and on how heavily one weights spatial contiguity as an
additional cost factor. The selection procedure uses an iterative
method known as simulated annealing. The model selects a set of
planning units at random and then randomly evaluates the effect of
adding or deleting randomly selected planning units. After a very
large number of iterations, the model converges towards a good
solution. It does not guarantee to find an “optimal” solution. Typi-
cally the model is run multiple times, with a different initial ran-
dom solution, and saves the best solution of those multiple runs.
Another useful feature is to tally the total number of times each
planning unit was selected in a series of runs, which suggests which
units tend to be part of most efficient alternatives.
The graphical user interface is built on the ArcView software but
has been customized for the toolbox. All actions, except for pre-
paring the data input files, are handled through menus, buttons, or
regular ArcView functionality. The reserve selection software pro-
duces output to files that can be displayed as maps, charts, or tables
within the toolbox. The most basic output is the set of planning
units selected for a given alternative. This can be displayed as a
theme in the ArcView project, either as the best solution from mul-
tiple runs or as the total number of times each planning unit was
selected in a series of runs. The summary information from a se-
ries of runs can be viewed as a table, allowing comparison of the
costs, number of planning units selected, total boundary length, and
the number of conservation elements that fell short of their target
levels. The distribution of individual elements can be displayed in
relation to an alternative portfolio to show the spatial pattern of
representation for that element. Conversely, the composition of
individual planning units can be examined (Figure 1) to see what
they contributed (or might contribute if preallocated) to the repre-
sentation goals.
Figure 1. A tabular display of the biodiversity elements and their aerial
extent
in a selected planning unit.
The series of three maps below (Figure 2) shows the effect of ad-
justing a model parameter (BLM) on the amount of spatial contigu-
ity, while representation goals and other parameters are kept the
same. The model tries to minimize the outer perimeter length of
the set of clusters of planning units. For a given total area, this is
accomplished by aggregating planning units into larger, more com-
pact clusters. The upper left map has no clustering enforced; any
apparent clustering is the result of the composition of the planning
units and the representation goals. The upper right and lower maps
have increasing clustering enforced by increasing the value of
BLM
