Multistage modelling approach
modelling approach is designed to help determine:
- How suitable tree and bird species habitats
may change under future climates
- The likelihood that modelled future habitats
are colonized by the species
- How much can unmodelled factors like
disturbances and biological characteristics change future outcomes
Our mutistage approach uses two models, DISTRIB
and SHIFT, and one
decision support system, MODFACs.
potential suitable species habitats for 134 tree species under current and future
SHIFT is used to determine the likelihood of the tree species colonizing these new
suitable habitats. DISTRIB+SHIFT
gives the potential colonization by 2100. MODFACs (Modification
factors) which uses a scoring system based on literature survey can be used to look at how outside
disturbances and biological factors might influence the future
distributions of these species. The results of DISTRIB and MODFACs can
be explored via the tree
atlas, with SHIFT results coming soon. The
bird atlas is the result of the bird DISTRIB model that incorporates output from
the tree DISTRIB model to predict the distribution of 147 bird species.
This multi-stage approach can be used to
evaluate the impact of climate change on species habitat distributions
better than a single-model approach. See (Iverson 2011)
for details of
our strategy. Our approach has been used as a decision support
system to plan for better management under current and future
disturbances via regional assessments.
discuss each component of the schematic diagram below. Please refer to citations for more detail.
For trees, response variables are the importance values (IV) of 134
tree species. IVs is a relative measure of abundance which is
calculated from US Forest Service's Forest Inventory and Analysis (FIA)
data for 37 states east of the 100th meridian. It consisted of over
100,000 plots and nearly 3 million tree records. The importance values
were calculated as
where x is the species, BA is the basal area and NS is the number of
stems. AllSpp indicates all species present in the plot. Therefore,
IV is a measure of relative abundance. Its value ranges from 0 to 100 ;
the latter indicating monotypic stands of the species. The plot data is
aggregated to 20 km cells.
For birds, response variables are incidences of 147 bird species from
nearly 1000 Breeding Bird Survey routes. Incidence is calculated as the
number of occurrences recorded/route over 7-10 yrs and aggregated to 20
km cells. The incidences range from 0 to 1.
For the tree
DISTRIB model we used 38
predictors which consist of 7
climate variables, 5 elevation classes, 9 soil classes, 13 soil
properties and 4 landuse variables. The bird DISTRIB model uses a total
of 50 variables, which includes 7 climate variables, 5 elevation
classes and 39 modelled tree IVs. Hence the bird DISTRIB is coupled to
the tree DISTRIB as the habitat of birds depends on the tree species
To model future climates we swap the current climate with future
climates predicted by 3 GCM models (HadCM3 (Hadley), GFDL and PCM)
under two carbon emission scenarios (the harsh but realistic A1FI
(High) and the mild optimistic scenario of B1 (Low) according to IPCC).
This results in the following scenarios:
HadleyCM3 – A1FI (High, “Harsh”)
HadleyCM3 – B1 (Low)
PCM – A1FI (High)
PCM – B1 (Low, “Mild”)
GFDL – A1FI (High)
GFDL – B1 (Low)
Avg. of 3 GCMs – A1FI (High)
Avg. of 3 GCMs – B1 (Low)
Note that Hadley High is the harshest and PCM Low is the mildest
scenario. See the (Iverson 2008)
for details on the data inputs.
The DISTRIB model predicts suitable habitat as depicted by IVs for tree
for bird species under current
(~2000) and future
climates. The DISTRIB species model has served as the core of the
multi-stage approach. The FIA data and the
DISTRIB model provides a basis for depicting the species current
well as how the habitat for these species might change with changing
model is based on decision-tree based ensemble approach called TriMod
which uses a robust technique called RandomForest for prediction. See (Prasad 2006) for details of
the ensemble modelling technique. Using
38 predictors that includes climate, soil, elevation and landuse
variables as discussed in the Data
section, tree DISTRIB models the IVs of 134 tree species
that are mapped as suitable habitats of the species. The bird DISTRIB
uses the modelled
suitable habitat of trees as input,
in addition to climate and elevation variables, to model the incidences
of 147 bird species. See (Matthews 2011a) for details on bird DISTRIB. Future
distributions are modelled by swapping the
current climate with GCM predicted future climates as discussed in the Data section.
Not all species models are equal - we therefore need to know about the "model
reliability" of each modelled species. We therefore rate
for each species,
the reliability of the DISTRIB model into three classes - high, medium
and low - taking into account several model performance factors.
An important caveat when interpreting our DISTRIB model is that we are
predicting potential suitable habitat by year 2100 – not
where the species will be. See (Iverson 2008) for details of the DISTRIB modelling approach.
DISTRIB predicts future suitable habitats, but these habitats may not
be colonized in this century depending on the speed that a tree species
can migrate and the fragmented nature of the landscapes. This is where
SHIFT, a spatially explicit simulation model steps in, to
calculate the likelihood of colonization based on the abundance of a
tree species within its current range (source region), the habitat
quality of the landscape beyond the source range boundary (sink) and
the distance between occupied and unoccupied areas, for a range of
historical migration rates (~ 10 km to > 100 km/century). SHIFT
thus calculates the likelihood of colonization for suitable
sink habitats (in approximately 100 year time period consiting of
multiple generations depending on the tree species). See (Prasad 2013)
for details on
the SHIFT model.
The future suitable habitats predicted by DISTRIB (~ 2100) are
intersected with SHIFT likelihood of colonization (100 years), to
estimate the likehihood of colonization of these future sink habitats
around 2100. This typically shows only a small percentage of the sink
habitats (ie., beyond the current range boundary) available to the tree
species, has any chance (say > 5%) of colonization. Using this
approach we can screen multiple tree speices and assess suitable
migration corridors and patches that could help iin assisted
colonization of valuable forest tree species that may not be able to
keep pace with rapid climate change.
the SHIFT model is not yet implemented in the climate change
atlas. We hope to incorporate it shortly. The diagram below shows the
DISTRIB output and SHIFT model intersected for post oak (Quercus stellata).
It illustrates the nature of DISTRIB and SHIFT and how they are
combined to produce potential colonization likelihood by 2100. Notice
that only a small percentage of the future sink habitats have any
chance of getting colonized.
MODFACs (Modification Factors)
The DISTRIB and the SHIFT models cannot take into account the multitude
of biological and disturbance factors affecting species distribution
(insect outbreaks, fire, etc). We therefore use a
scoring system based on the available literature to
account for these factors. Our scoring system gauges the effect of 9 biological
and 12 disturbance components in modifying the interpretations of the
species response outcomes from the DISTRIB and SHIFT models. This
framework also addresses species model uncertainties in light of
climate change. Forest managers and other end users who have expertise
locally can modify the tabular scoring system locally or regionally, to
obtain customized outputs and inform management relevant decisions. See
An example for red maple, a highly adaptable species, achieving the
highest score of 8.5 is shown below.
We have used
the outputs of the multi-stage approach, mainly DISTRIB and MODFACs to assess management
options for different regions. See (Swanston 2011) for
details on regional assessment.
Prasad, A.M., L.R. Iverson, and A.
Liaw 2006. Newer classification and regression tree
techniques: bagging and random forests for ecological prediction.
Iverson, L. R., A. M. Prasad, S.
N. Matthews, and M. Peters. 2008. Estimating potential habitat for 134
eastern US tree species under six climate scenarios. Forest Ecology and
Prasad, A. M., J. Gardiner, L.
Iverson, S. Matthews, and M. Peters. 2013. Exploring tree species
colonization potentials using a spatially explicit simulation model:
implications for four oaks under climate change. Global Change Biology 19: 2196–2208.
Matthews, S. N., L. R. Iverson,
A. M. Prasad, and M. P. Peters. 2011. Potential habitat changes of 147
North American bird species to redistribution of vegetation and climate
following predicted climate change. Ecography 260:1460-1472.
Matthews, S. N., L. R. Iverson,
A. M. Prasad, M. P. Peters, and P. G. Rodewald. 2011. Modifying climate
change habitat models using tree species-specific assessments of model
uncertainty and life history factors. Forest Ecology and Management 262:1460-1472.
Iverson, L., A. M. Prasad, S.
Matthews, and M. Peters. 2011. Lessons learned while integrating
habitat, dispersal, disturbance, and life-history traits into species
habitat models under climate change. Ecosystems 14:1005-1020.
Swanston, Chris; Janowiak, Maria; Iverson, Louis; Parker, Linda; Mladenoff, David;
Brandt, Leslie; Butler, Patricia; St. Pierre, Matt; Prasad, Anantha; Matthews,
Stephen; Peters, Matthew; Higgins, Dale; Dorland, Avery. 2011. Ecosystem vulnerability assessment and synthesis: a report from the Climate Change Response Framework Project in northern Wisconsin. Gen. Tech. Rep. NRS-82. Newtown Square, PA: U.S. Department of Agriculture, Forest Service, Northern Research Station. 142 p.