Climate-Biome Shifts in Alaska
and Western Canada
Current Results and Final Steps
October 2011
Participants
 Scenarios Network for Alaska Planning (SNAP),
University of Alaska Fairbanks
 EWHALE lab, Institute of Arctic Biology, University of
Alaska Fairbanks
 US Fish and Wildlife Service
 The Nature Conservancy
 Ducks Unlimited Canada
 Government of the Northwest Territories
 Government of Canada
 Other invited experts
Overview

This project is intended to:
◦ Develop climate and vegetation based biomes (based on cluster
analysis) for AK, Yukon, NWT, and areas to the south that may
represent future climatic conditions for AK,Yukon or NWT.
◦ Model potential climate-induced biome shift.
◦ Based on model results, identify areas that are least or most likely to
change over the next 10-90 years.
◦ Provide maps, data, and a written report summarizing, supporting, and
displaying these findings.


This project builds ,and makes use of, work previously conducted
by SNAP, EWHALE, USFWS, TNC, and other partners.
The completed analysis will be used by partners involved in
protected areas, land use, and sustainable land use planning, e.g.
connectivity.
Goals of this meeting
Review Project Goals
Summary of project background
Refresher on modeling methods and data
Update on decisions and progress thus far
Discussion and decisions from group:
◦ Finalizing and processing results
◦ Discuss data delivery and formats
◦ Other issues?
 Timeline for completion





The Scenarios Network for Alaska
and Arctic Planning (SNAP)


SNAP is a collaborative network of the University of
Alaska, state, federal, and local agencies, NGOs, and
industry partners.
Its mission is to provide timely access to scenarios of
future conditions in Alaska for more effective planning
by decision-makers, communities, and industry.
SNAP Projections:
based on IPCC models







SNAP uses data for 5 of 15 models that performed
best for Alaska,Yukon, and NWT
PRISM downscaled to 2 km resolution OR CRU
downscaled to 10 minutes (18.4 km)
Monthly temp and precip from 1900 to 2100
(historical CRU + projected)
5 models x 3 emission scenarios
Available as maps, graphs, charts, raw data
On line, downloadable, in Google Earth, or in
printable formats
No data yet:
◦ Extreme events
◦ Snowpack
◦ Coastal/Oceans
Phase I: Alaska model
Mapped shifts in potential biomes based on current climate envelopes for six Alaskan
biomes and six Canadian Ecozones
7
http://geogratis.cgdi.gc.ca/geogratis/en/collection/detail.do?id=4361
Phase I Results:Potential Change: Current - 2100
(Noting that actual species shifts lag behind climate shifts)
Improvements over Phase I
Extend scope to northwestern Canada
 Use all 12 months of data, not just 2
 Eliminate pre-defined biome/ecozone categories
in favor of model-defined groupings (clusters)

◦ Eliminates false line at US/Canada border
◦ Creates groups with greatest degree of intra-group
and inter-group dissimilarity
◦ Gets around the problem of imperfect mapping of
vegetation and ecosystem types
◦ Allows for comparison and/or validation against
existing maps of vegetation and ecosystems
Other Improvements

Comparison with multiple landcover classification
categories
◦ Eliminates over-reliance on one categorization scheme
◦ Utilizes divergent methods of classification

Recalibration of all SNAP downscaling to
reconcile discontinuities between US and
Canada and errors in A2 and B1 data (not
used in Phase 1)*
* Note that this has caused project delays, but the very
extensive work involved will NOT be charged to project
funders.
Temperature Average Mosaic 2050
Sampling Extent
Cluster analysis





Cluster analysis is the assignment of a set of
observations into subsets so that observations
in the same cluster are similar in some sense.
Clustering is a method of “unsupervised
learning” (the model teaches itself, and finds the
major breaks)
Clustering is common for statistical data analysis
used in many fields
The choice of which clusters to merge or split is
determined by a linkage criterion (distance
metrics), which is a function of the pairwise
distances between observations.
Cutting the tree at a given height will give a
clustering at a selected precision.
Step 1: Create a Dissimilarity Matrix
Distance measure determines how
the similarity of two elements is
calculated.
 Some elements may be close to one
another according to one distance
and farther away according to
another.
 In our modeling efforts, all 24
variables are given equal weight, and
all distances are calculated in “24dimensional space”

(similarity matrix, proximity matrix,
distance matrix get converted
into each other)
Taxicab geometry versus
Euclidean distance:
The red, blue, and yellow lines
have the same length in
taxicab geometry for the same
route. In Euclidean geometry,
the green line has length
6×√2 ≈ 8.48, and is the unique
shortest path.
Methods: Partitioning Around Medoids (PAM)





The dissimilarity matrix describes pairwise distinction
between objects.
The algorithm PAM computes representative objects,
called medoids whose average dissimilarity to all the
objects in the cluster is minimal
Each object of the data set is assigned to the nearest
medoid.
PAM is more robust than the well-known kmeans
algorithm, because it minimizes a sum of dissimilarities
instead of a sum of squared Euclidean distances,
thereby reducing the influence of outliers.
PAM is a standard procedure
Resolution limitations
Data are not available at the same resolution for
the entire area
◦ for Alaska, Yukon, and BC, SNAP uses 19611990 climatologies from PRISM, at 2 km,
◦ for all other regions of Canada SNAP uses
climatologies for the same time period from
CRU, at 10 minutes lat/long (~18.4 km)
◦ In clustering these data, both the difference in
scale and the difference in gridding algorithms
led to artificial incongruities across
boundaries.
Solution to resolution limitations
The solution to both resolution and clustering
limitations was to cluster across the whole
region using CRU data, which is available
for the entire area, but to project future
climate-biomes using PRISM, where
available, to maximize resolution and
sensitivity to slope, aspect, and proximity to
coastlines.
How many clusters?
Choice is mathematically somewhat
arbitrary, since all splits are valid
 Some groupings likely to more closely
match existing land cover classifications
 How many clusters are defensible?
 How large a biome shift is “really” a shift
from the conservation perspective?

Sample cluster analysis showing 5 clusters, based on CRU 10’
climatologies. This level of detail was deemed too simplistic to meet
the needs of end users.
Sample cluster analysis showing 30 clusters, based on CRU 10’
climatologies. This level of detail was deemed too complex to meet
the needs of end users, as well as too fine-scale for the inherent
uncertainties of the data.
0.2
0.18
0.16
Average Sihouette Width
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
25
30
40
50
Number of Clusters Returned
Mean silhouette width for varying numbers of clusters between 3
and 50. High values in the selected range between 10 and 20 occur at
11, 17, and 18.
Eighteen-cluster map for the entire study area. This cluster
number was selected in order to maximize both the
distinctness of each cluster and the utility to land managers
and other stakeholders.
Cluster certainty based on silhouette width. Note that certainty
is lowest along boundaries.
Assessing the clusters
Box plots, rose plots,
and other direct
temp/precip depictions
 Congruence with
existing land cover
classification by modal
values
 Congruence with land
cover classification by
percent

10
JanT
0
-10
-20
-30
1
2
3
4
5
6
7
8
9
AKClustPAM
10
11
12
13
14
15
Rose Plots Showing the “Shape” of clusters
Line Graphs – clear, simple, familiar
NALCMS Land Cover, 2005





North American Land Change
Monitoring System (Canada,
Mexico and US)
Based on monthly composites of
2005 MODIS imagery
250m resolution;19 classes
Separating shrubland, grasslands,
deciduous forests and evergreen
forests, as well as open vs closed
canopies.
Does not distinguish boreal forest,
temperate forest and rain forest,
and defines tundra as either
“grassland” or “sparse.”
AVHRR Landcover, 1995
USGS – NOAA data
13 categories with clear distinctions between
forest, shrubs, and grasslands.
 “Dwarf shrub” and “herbaceous” categories
help define tundra
 Forested areas are defined primarily as
deciduous or needle-leaf.


GlobCover 2009
ESA JRC initiative
 300m MERIS sensor, ENVISAT satellite
 Similar strengths and weaknesses to
AVHRR

Alaska Biomes and Canadian
Ecozones
Nowacki et al. 2001 and Envirnoment Canada
 Differs fundamentally from other 3 classification systems -not based on remote sensing data but rather on a
combination of observed data and interpolated data.
 Based on not only cover type but also functional and
morphological details, e.g. ecosystem function, soil type,
bedrock, wildlife.

1.00
Permanent snow and
ice
0.90
0.80
Water bodies
0.70
0.60
Sparse (<15%)
vegetation
0.50
0.40
Closed to open (>15%)
herbaceous vegetation
(grassland, savannas or
lichens/mosses)
Mosaic grassland (5070%) / forest or
shrubland (20-50%)
0.30
0.20
0.10
0.00
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18
Dominant GlobCover 2009 land cover by cluster number. All land
cover categories that occur in 15% or more of a given cluster are
included.
1.00
Bare Ground
0.90
Cropland
0.80
Grassland
0.70
Open Shrubland
0.60
Closed Shrubland
0.50
Wooded Grassland
0.40
0.30
Woodland
0.20
Mixed Forest
0.10
Deciduous
Broadleaf Forest
0.00
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Dominant AVHRR land cover types by cluster number. All land
cover categories that occur in 15% or more of a given cluster are
included.
1
Montane Cordillera
0.9
Pacific Maritime
0.8
Boreal Cordillera
0.7
Taiga Cordillera
0.6
Prairie
0.5
Boreal PLain
Boreal Shield
0.4
Taiga Shield
0.3
Taiga Plain
0.2
Southern Arctic
0.1
Northern Arctic
Western Tundra
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18
Dominant Alaska biomes and Canadian Ecozones by cluster
number. Alaska biomes are adapted from Nowacki et al. 2001;
Canadian ecozones are defined by Natural Resources Canada. All
categories that occur in 15% or more of a given cluster are included.
Climate-biome characteristics
1. The coldest cliome. Northern Arctic
sparsely vegetated tundra with up to
25% bare ground and ice, with an
extremely short growing season.
2. Cold northern arctic tundra, but
primarily vegetated
3. More densely vegetated arctic tundra
with up to 40% shrubs but no tree
cover
4. Arctic tundra with denser vegetation, a
longer growing season, and more shrub
cover including some small trees
5. Dry sparsely vegetated southern arctic
tundra
6. Northern boreal/southern arctic
shrubland, with an open canopy and
short growing season
7. Northern boreal coniferous woodland,
open canopy
8. Dry boreal wooded grasslands —
mixed coniferous forests and grasses
9. Mixed boreal forest, interior climate
10. Boreal forest with coastal influence
and intermixed grass and tundra
11. Cold northern boreal/subarctic forest
12. More densely forested closed-canopy
boreal
13. Dry, sparsely vegetated northern
boreal
14. Interior boreal, densely forested
15. Warmer boreal zone with mixed
evergreen and deciduous
16. Southern boreal, mixed forest
17. Coastal rainforest, wet, more
temperate
18. Prairie and grasslands
Data selected
from SNAP models
Available climate data from SNAP includes output for each
of the five best-performing GCMs as well as a
composite (mean) of all five models for all months of all
years to 2100, for each of three emission scenarios as
defined by the IPCC: A1B, A2, B1
 Selected three future time periods (2030-2039; 20602069 and 2090-2099). Decadal averages deemed more
useful than single years for capturing trends.
 Chose to provide outputs from all five models plus
composite in A1B, and all three emission scenarios with
composite.
Defining refugia and areas of
greatest change
Decadal results from RandomForest will be
analyzed to determine which grid cells are
projected to remain within the same biome
climate envelope over the time periods.
 Thresholds for what constitutes greatest change

◦ Sites that shift climatically to match non-adjacent
biomes can be interpreted as a proxy for magnitude
of change
◦ Areas that shift to a climate-biome with different
dominant land cover types according to AVHRR,
GlobCover, etc.
Interpreting confidence in refugia/
areas of change
◦ Only consider areas selected as refugia in
the majority (or all?) of the climate
models
◦ RandomForest assigns a ranking value to
each of pixel that can be used to identify
the model confidence
◦ Treat boundary zones (low silhouette
values) between cliomes as belonging to
either cliome
Data and Product Delivery






Formal report
Report submitted to the FWS Journal of Fish &
Wildlife Management or to a peer reviewed
journal
Executive summary
Additional short/simplified/regional versions?
PPT presentations – for what audiences?
Posters, talks, and additional publications?
Timeline for completion
Remaining salary provided by grants will not
be withdrawn until project completion
 Remaining modeling/mapping steps

◦ Finalizing projection results
◦ Defining refugia/areas of change
◦ Hoping to meet current Dec 31 goal

Product creation
◦ Simultaneous with above
◦ Likely to require group feedback, final edits, copy
editing and printing in early 2012