Three steps to strengthen confidence in connectivity models
ABSTRACT
Maintaining and restoring ecological connectivity is considered a global
imperative to help reverse the decline of biodiversity. To be
successful, practitioners need to be guided by connectivity modeling
research that is rigorous and reliable for the task at hand. However,
the methods and workflows within this rapidly growing field are diverse
and few have been rigorously scrutinized. We propose three procedural
steps that should be consistently undertaken and reported on in
connectivity modeling studies in order to improve rigour and utility:
(1) describe the type of connectivity being modeled, (2) assess the
uncertainty and sensitivity of model parameters, and (3) validate the
model outputs, ideally with independent data. We reviewed the literature
to determine the extent to which studies included these three steps. We
focused on studies that generated novel landscape connectivity outputs
using circuit theory. Among 181 studies meeting our search criteria,
39% communicated the type of connectivity being modeled and 18%
conducted some form of sensitivity or uncertainty analysis (or both).
Only 19% of studies attempted to validate their connectivity model
outputs and only 7% used fully independent data. Our findings highlight
a clear need and opportunity to improve the rigour, reliability, and
utility of connectivity modeling research. At a minimum, researchers
should be transparent about which, if any, of these three steps were
undertaken. This will help practitioners make more informed decisions
and ensure limited resources for connectivity conservation and
restoration are allocated appropriately.
Keywords: biodiversity, circuit theory, Circuitscape, connectivity,
conservation, sensitivity analysis, uncertainty, validation
INTRODUCTION
A 2019 global assessment by the Intergovernmental Science-Policy
Platform on Biodiversity and Ecosystem Services estimated that up to one
million species around the globe are at risk of extinction (Brondizio et
al. 2019). Preventing this large-scale biodiversity decline will require
action to reduce the intensity of its underlying drivers: habitat loss,
degradation, and fragmentation (Newbold et al. 2015). These drivers
collectively exacerbate the problem by impeding species movement, or
ecological connectivity, which is necessary to allow individuals to
access food and water, establish new territories, supplement existing
populations, avoid predators, and to find breeding partners (Hilty et
al. 2020). In addition, reduced connectivity will inhibit species’
ability to shift ranges and adapt to climate change effects (Thomas et
al. 2004; Heller & Zavaleta 2009; Hannah 2011). Consequently, Parties
to the Convention on the Conservation of Migratory Species of Wild
Animals (CMS; 2020), a multilateral environment agreement under the
United Nations, reaffirmed that maintaining and restoring ecological
connectivity is one of their top priorities.
The need to identify areas important for connectivity spurred a
proliferation of connectivity modeling studies (Correa Ayram et al.
2016), while the introduction of free and user-friendly tools such as
Condatis (Wallis & Hodgson 2015), Linkage Mapper (Gallo & Greene
2018), Unicor (Landguth et al. 2012), and Circuitscape (McRae et al.
2008) made modeling connectivity more broadly accessible. Although the
wide interest in, and application of, connectivity modeling is generally
a positive development, it has fostered an abundance of literature that
includes an overwhelming array of modeling methods and workflows (Zeller
et al. 2012), few of which have been rigorously scrutinized (Wade et al.
2015; Zeller et al. 2018).
Given the limited resources available for implementing connectivity
conservation actions, it is crucial that sound and transparent evidence
underpin conservation prioritizations and actions (Elliot et al. 2014;
McClure et al. 2016; Carroll et al. 2020). Furthermore, evaluating and
weighing the evidence would be far easier if studies implement steps
that are considered to be crucial for maximizing model reliability
(Spear et al. 2010; Sawyer et al. 2011; Zeller et al. 2012; Wade et al.
2015; Abrahms et al. 2017; Laliberté & St-Laurent 2020). These steps
include: (1) articulating the type of connectivity being modeled, (2)
evaluating how much model output changes in response to uncertainty in
the input parameters, and (3) validating the resistance surface and
especially the connectivity model outputs. Here we report the results of
a literature review focused specifically on if and how studies
implemented any of these three steps.
Articulating the type of connectivity being modeled
Study goals should dictate, and clearly articulate, the type of
connectivity that needs to be modeled at the outset (Allen & Singh
2016; Diniz et al. 2020). There are several types of connectivity,
including daily foraging, seasonal migrations, dispersal/genetic, or
range shifts, each of which correspond to a type of movement that
operates at different spatial and/or temporal scales (Wade et al. 2015).
For example, daily habitat connectivity pertains to movements by
individuals to meet daily food, water, and shelter needs, whereas range
shift connectivity focuses on movements that enable species to track
habitat that is shifting due to climate change. The relevant spatial and
temporal grains and extents clearly differ between those two types of
connectivity; hence the need to clearly define what type of connectivity
is being modeled and to select model inputs accordingly (Laita et al.
2011; Elliot et al. 2014).
Sensitivity and uncertainty analysis
Evaluating how much model output changes due to uncertainty among input
parameters constitutes “uncertainty analysis” (Beier et al. 2009).
This is technically different from sensitivity analysis, which evaluates
which input parameters have the greatest influence on model output
(Beier et al. 2009). Although the specific goals of uncertainty analysis
and sensitivity analysis differ, in practice they are often implemented
simultaneously. For example, the process of systematically varying the
values of input parameters and quantifying how it affects variation in
outputs can simultaneously serve as both uncertainty and sensitivity
analysis (e.g., Marrec et al. 2020). The key is to articulate how the
parameters in question relate to uncertainty or knowledge gaps. For
simplicity, we will henceforth refer to these analyses collectively as
sensitivity and uncertainty analysis (SUA).
Many authors have highlighted the importance of SUA, in part because
there are many sources of uncertainty to consider in connectivity
modeling, and because of the many decisions and assumptions in the
analysis process (Wade et al. 2015). For example, using cougar GPS
telemetry data, Zeller et al. (2017) showed that connectivity model
outputs, including predicted locations for road crossings, were
sensitive to the number of geospatial layers, the number of classes in
categorical geospatial data (thematic resolution) and the spatial
resolution of the geospatial data used when constructing the resistance
surface. Using circuit theory based approaches to connectivity modeling,
Bowman et al. (2020) showed that the range of cost weights within cost
surface maps influenced the spatial distribution of current density.
This narrowed the range of input cost weights and tended to spread
current across the landscape, whereas current tended to be more
concentrated when a broader range of cost weights were used.
Model validation
Given the many assumptions and sources of uncertainty inherent to
connectivity models, perhaps the most important step in the analysis
process involves validation of inputs (e.g., resistance surface) or more
importantly, outputs (e.g., current density map, potential corridors)
(Zeller et al. 2018; Carroll et al. 2020; Laliberté & St-Laurent 2020).
For example, a study with an objective to help identify specific
linkages for protection or locations along roads for overpasses or
fencing generates an obvious need to validate model predictions with
real-world movement or even presence/absence data (Xu et al. 2019;
Cerqueira et al. 2021). However, if the study’s aim is to assess the
overall degree of connectivity of a landscape, then model predictions
could be validated with inferential data such as genetics and
biogeochemical markers (Wade et al. 2015). In either case, truly
independent data would ideally be used; in practice these data are
costly and time consuming to obtain. As a result, data used for
validation vary in type and degree of independence, with implications
for the reliability of model predictions (Spear et al. 2010; McClure et
al. 2016). In the case of telemetry-based connectivity models, a portion
of individuals, or relocations, could be withheld from the modeling
procedure and used to validate results (but see Roberts et al. 2016).
Methodological shortfalls in the connectivity modeling literature have
been reported for about a decade (Zeller et al. 2012; Wade et al. 2015;
Laliberté & St-Laurent 2020), so we were hoping to see evidence of
these shortfalls being addressed in more recent studies. Thus, in
addition to evaluating if and how studies addressed the preceding
procedural steps, we assessed whether the frequency of studies attending
to these steps has changed over time. We focused our review on studies
of terrestrial mammals (including bats) that used circuit theory in the
time since it was first popularized as an approach to assess
connectivity with the introduction of Circuitscape (McRae et al. 2008).
Circuitscape is one of the most widely used connectivity modeling tools
in recent years, appearing in 80 publications in 2017 alone (Dickson et
al. 2019). We assume that the practices used in this subset of studies
are representative of connectivity modeling studies in general.
METHODS
Literature Search
We used the Web of Science® database (WoS) to find
publications that used circuit theory for landscape connectivity
modeling. We accessed WoS on March 1st, 2022, and
found all publications that had cited McRae et al.’s Using Circuit
Theory to Model Connectivity in Ecology, Evolution, and Conservation(2008) up until Dec 31st, 2021. We refined the
resulting list to only include studies that generated novel landscape
connectivity outputs using circuit theory and that focused on one or
more species of terrestrial mammal, including bats. Studies that modeled
multispecies connectivity for a suite of taxa, mammalian or otherwise,
were accepted so long as at least one terrestrial mammal was included.
Literature Assessment
For each study we recorded whether the researchers stated what type of
connectivity was being modeled, and for those that did, we categorized
the studies according to the following types: foraging, seasonal
migrations, dispersal/genetic, or range shifts. If the type was not
stated, it was recorded as non-specific.
We determined whether some form of SUA was conducted, and if so, we
recorded the sources of uncertainty that were evaluated (e.g.,
resistance values, grid resolution). We also noted whether a factorial
design was used in the SUA. A factorial design enables the simultaneous
assessment of the influence of multiple factors within the SUA,
including any interactions.
We then determined whether model validation was conducted, and if so, we
noted (i) whether validation was conducted on the layers used to derive
connectivity models (henceforth “input layers,”, e.g., habitat layers
or resistance surfaces) and / or on the “output layers” from the
connectivity model itself (i.e., the current density map produced using
circuit theory) (ii) the type of data that were used for validation
(tracking, genetic, camera trap, point occurrence, or expert opinion)
(Table 1). If a study used more than one type of data for validation, we
recorded the type that is deemed to be associated with least uncertainty
according to the hierarchy presented by Wade et al. (2015) (iii) the
degree of independence of the data used for validation (categories: not
independent, partially independent, fully independent), and (iv) the
degree of agreement between model predictions and validation data. The
latter results were taken at face value from authors’ written
interpretation of their validation analysis and generalized as either
“positive” or “negative.” If authors found their models had mixed
ability to predict validation data, or if the results of the validation
were not clearly presented, they were categorized as either “mixed” or
“inconclusive,” respectively. For our purposes, “mixed” included
multispecies studies that found success in modeling movement of one
studied species, but not all.
We cross-tabulated studies according to the type of data used for
validation and the degree of independence of the data used for
validation.
Temporal trends in modeling practices
We first tallied the number of studies that met our criteria in each
publication year. We then qualitatively evaluated whether practices have
changed in time by plotting line graphs of the proportion of studies in
each year that fell into each of the assessment categories. For
instance, we calculated the proportion of studies in each year that
conducted some form of SUA and noted any obvious trend.
RESULTS
Literature search
We found 884 studies that cited McRae et al. (2008) in our literature
search. Of those, 181 met our criteria of having novel landscape
connectivity outputs generated using circuit theory (Figure 1) and that
focused on one or more species of terrestrial mammal. The Supporting
information provides details for each study.