ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69 – 85
www.elsevier.com/locate/isprsjprs
Mapping vegetation spatial patterns from modeled water,
temperature and solar radiation gradients
Caren C. Dymond *, Edward A. Johnson
Department of Biological Sciences, University of Calgary, Calgary, AB, Canada T2N 1N4
Received 15 January 2002; accepted 23 August 2002
Abstract
Maps of current and potential vegetation spatial patterns can be used to assess land cover changes, and aid in ecosystem
management and restoration. The vegetation spatial patterns of subalpine forest species are largely controlled by variation in
temperature, water and solar radiation resources. These fundamental resources were quantified across a 1100-km2 landscape
using biophysical models, a digital elevation model (DEM), and weather station data. Field data of species abundances were
used to define species – habitat relationships and calibrate maximum likelihood classifications of the biophysical gradients. For
comparison, a standard land cover classification of Landsat Thematic Mapper satellite imagery had an overall accuracy 68.3%.
Using the biophysical gradients alone gave a similar 67.4% accuracy. The highest accuracy classification (83.2%) used both
biophysical and spectral data. The biophysical resources were also used to map the presence or absence of four herb and shrub
species that cannot be sensed remotely. These predictions ranged from 60% to 79% accurate. Maps of relative abundance were
less accurate, from 61% to 63.2%. This low result may be due to historical and stochastic events, or simply a small data set. The
spatial pattern of species and communities that are controlled by resources can be predicted using general biophysical models.
The species – habitat relationships can also be used to improve remote sensing products.
D 2002 Elsevier Science B.V. All rights reserved.
Keywords: spatial pattern prediction; species distribution; land cover change; subalpine
1. Introduction
There is a growing need to predict vegetation
spatial patterns for assessing land cover change, for
planning restorations, and for implementing sustainable management. Changes in land cover due to
human influence could be estimated by comparing
* Corresponding author. Current address: Canadian Forest
Service, 5320-122 Street, Edmonton, AB, Canada T6H 3S5. Tel.:
+1-780-435-7223; fax: +1-780-435-7359.
E-mail address: [email protected] (C.C. Dymond).
current patterns with landscape scale predictions of
natural species distributions (Davis and Goetz, 1990;
Iverson et al., 1997). This kind of landscape scale
model could identify restoration sites for rare species
(Wiser et al., 1998). These predicted sites could be
expanded to extirpated species if the model was
general enough to be calibrated on one landscape
and applied to a different one. This ability to predict
the landscape pattern of individual plant species could
also be used to identify existing or potential wildlife
habitat (Dettmers and Bart, 1999). In addition, efficient mapping of herbs, shrubs, and habitats across
landscapes would help bring about the mandate of
0924-2716/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 4 - 2 7 1 6 ( 0 2 ) 0 0 11 0 - 7
70
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
ecosystem management (CCFM, 1998). These needs
have led this study to the development of a landscape
scale model that employs general physical processes
in order to predict species spatial patterns.
One approach to understanding and predicting
vegetation patterns is to relate species abundance to
environmental descriptors. However, application of
this approach is limited to because the gradients are
locally defined. For example, in high relief areas,
species distributions have been correlated to elevation,
topographic indices and aspect in ordination studies
(Whittaker, 1973; La Roi and Hnatiuk, 1980; Peet,
1981), and in vegetation prediction studies (Brzeziecki et al., 1993; Bolstad et al., 1998). These studies
cannot be applied to different areas because elevation
and aspect do not affect plants directly, but are
correlated to climate processes (Whittaker et al.,
1968; Zobel et al., 1976; Wentworth, 1981). These
climate processes will change from one landscape to
the next. For example, the same elevation at two
different latitudes will have different temperatures
and radiation levels, in effect, different levels of the
resources affecting plant physiology. Topographic
indices are similarly limited because they are categorical, so a moisture ranking of 4 is probably not twice
as wet as a ranking of 2 (e.g. Allen and Peet, 1990).
These relative, rather than absolute, gradients make it
difficult to apply the models for large-scale mapping
or compare between research protocols.
The correlation or association of species to each
other has been used map the shrub and herbaceous
layers that cannot be remotely sense through a forest
canopy (Lillesand and Kiefer, 1994). For example,
communities have been mapped using the association
between tree species and understory species (He et al.,
1998). Inaccuracies can occur when the underlying
assumption of similar niche widths is invalid because
different species have wider or narrower niches than
others (Gleason, 1926).
Another approach to mapping vegetation communities depends on their relationship to ecosystem and
gap dynamics models that incorporate biological feedback into models of resource availability. These
models tend to be complex and data intensive (Shugart and West, 1977; Running and Gower, 1991;
Botkin et al., 1972). Part of the intricacy of these
models is the local definition of biotic and abiotic
feedback systems and relationships. The detail
required prohibits the models from being applied to
landscapes where these relationships are not well
understood. The costs, complexity and limited predictive ability associated with these models also
prevent their widespread application in the management and conservation communities (Korzukhin et al.,
1996).
It may be possible to improve vegetation maps
with simpler, more general biophysical models.
Successful predictions of communities with this
method include: local dune communities using
hydrological parameters (Noest, 1994), California
chapparal using climate parameters and regional
north temperate forests using climate parameters
(Lenihan, 1993; Monserud and Tchebakova, 1996).
These studies relied on biophysical process models
that defined species – habitat relationships which
were comparable and transferable to other scales or
sites, unlike environmental descriptors. These biophysical process models have no biological feedback
and parameterization required only simple measurements (e.g. Tchebakova et al., 1994; Ostendorf and
Reynolds, 1998), thus avoiding the complexity of
ecosystem models. The three biophysical resources
that are fundamental to physiological function are
temperature, water and solar radiation (Jarvis and
Leverenz, 1983; Stephenson, 1990; Larcher, 1995).
These resources have been identified as controlling
much of the spatial pattern in subalpine forests at the
landscape scale (Whittaker and Neiring, 1965; Busing et al., 1993; Wentworth, 1981). In addition,
these resources have large variance within subalpine
landscapes, and that variability can be captured with
simple, mechanistic models.
There are many other factors affecting vegetation
patterns. Stochastic, dispersal and historical processes
also affect species distribution (McCune and Allen,
1985; Ricklefs, 1987). Therefore, Therefore, our predictions do not attempt to explain all spatial variability. For example, landscape scale patterns may
depend on climatic variability, but those same climate
models cannot be used to predict variability within a
stand.
In this study, we combined the general nature of
process models with the precision and predictive
power of empirical models (Levins, 1966; Korzukhin
et al., 1996). The purpose was to test the ability of
general species– habitat relationships to map the spa-
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
tial pattern of vegetation across a mountainous landscape. Our objectives were to (1) understand the
relationships between moisture input, temperature
and solar radiation with the current species distribution; (2) map the spatial pattern of the tree, shrub and
herbaceous species using a maximum-likelihood classifier to characterize species– habitat relationships. (3)
The effectiveness of the predictions were tested by
comparing the accuracy of the maps to similar maps
made from remotely sensed observations of the land
cover, or to a null model when direct observations
were not possible. This study was completed in the
Kananaskis River and Spray River watershed in the
southeastern Canadian Rockies. This area provides an
ideal test site because subalpine vegetation is well
known to have strong relationships to physical gradients, (Whittaker, 1973; La Roi and Hnatiuk, 1980;
Peet, 1981).
2. Methods
2.1. Study area geology and climate
The Kananaskis River and Spray Reservoir watersheds are located in the front ranges of the Southern
Canadian Rockies (Fig. 1). The watersheds occupy
1184 km2 of high relief mountains, with the outflow
from the valley at 1350 m a.s.l., and the mountain
peaks between 2400 and 3000 m a.s.l. The structure
of the front ranges is defined by a series of thrust
faults that resulted in parallel ridges of older (Palaeozoic) limestone over younger (Cretaceous) sandstone and shale. Erosional processes on these
bedrock layers formed valleys in the softer sandstone
and shale, leaving the more resistant limestone as
ridges (McMechan, 1988).
Climate in the Kananaskis river watershed is a
transition between the Continental climate and the
Cordilleran climate. The average precipitation regime has a winter maximum (22.3 mm) and a
summer maximum (102.6 mm), with minima in late
winter and again in autumn (Janz and Storr, 1977).
The annual temperature profile follows a typical
pattern for this latitude, with the lowest average
monthly temperatures in January ( 6 jC), and the
highest in July (14 jC) (Kananaskis Field Stations,
1410 m a.s.l.).
71
The climate also changes spatially with elevation
(Storr and Ferguson, 1972). The average annual precipitation increases linearly from 474.5 mm at 1400 m
to 1110.5 mm at 2400 m, based on an 8-year study of a
sub-basin in the watershed. Data from the same study
showed that the average of the July maximum daily
temperature decreases from 21.9 jC at 1400 m to 14.2
jC at 2400 m (Janz and Storr, 1977).
Large-scale climate processes drive the large sized
and high intensity fires which drive population
dynamics of the subalpine forest (Johnson and Larsen,
1991). The stand-replacing fires burned independent
of forest type or elevation, therefore, the same disturbance rate occurs throughout the study area.
Human disturbances have had little impact on most
of the study area (Johnson and Fryer, 1987). Logging
occurred in limited parts of the Kananaskis valley
from 1886 to 1947. This disturbance effectively ended
when the leases were burned by a large fire in 1936.
The logging did not significantly impact species
composition. Development in the study area is
restricted to recreational facilities, hydroelectric dams
and associated reservoirs.
2.2. Subalpine species
The subalpine forest is primarily coniferous and
ranges in elevation from 1350 to around 2350 m.
Lodgepole pine (Pinus contorta Dougl. ex Loud. var.
latifolia Engelm.), white spruce (Picea glauca
(Moench)Voss), white spruce/Engleman spruce
hybrids (Picea engelmanni P. glauca) and trembling
aspen (Populus tremuloides Michx.) are the predominant trees in the lower subalpine. Engelman spruce
(P. engelmanni Parry ex. Engelm.), and subalpine fir
(Abies lasiocarpa (Hook.)Nutt.), with some subalpine
larch (Larix lyallii Parl.) at tree-line, are predominant
in the upper subalpine. The most common understorey
shrubs are soapberry (Shepherdia canadensis (L.)
Nutt.), prickly rose (Rosa acicularis Lindl.) and
willow (Salix sp.). Common herbaceous plants are
asters and erigerons (Compositae), strawberry (Fraagaria virginianum Duchesne), twinflower (Linnaea
borealis L.), heart-leaved arnica (Arnica cordifolia
Hook.), and bearberry (Arctostophylous uva-ursi L.).
Common mosses include feather moss (Pleurozium
schreberi (Brid.) Mitt.), and common lichen include
Peltigera apthosa (L.) Wild. and Cladonia species.
72
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
Fig. 1. The field sites and water ways in the Kananaskis and Spray watershed and the location of the study area in North America.
Steep gully walls were too small to be included in this
study, but are often occupied by dwarf adult trees of
Populus balsamifera (L.), P. contorta or Pinus flexilis
(James).
2.3. Input data
Species abundances were measured at 58 stands
from 1993 to 1995, to calibrate and test the model
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
predictions. Selection criteria for the forest stands
was: (1) as wide a range of species and habitats as
possible, (2) only stands greater than 3 ha, (3) visual
homogeneity in height, diameter, and species composition, (4) lack of standing water, (5) absence of
human disturbance (i.e. no cut stumps and a minimum
of 30 m from trails or roads).
In each stand, vegetation was sampled along a
transect with 15 points which were 15 or 30 m apart.
This transect encompassed about nine pixels on the
TM image. Canopy trees, saplings and shrubs were
sampled with the point centered quarter method
(Cottam et al., 1953). Canopy trees ( z 7 cm diameter
at base) were sampled for frequency, density and
diameter at base (db). A tree is defined as a species
that could grow to the height of the canopy in the
study area. Saplings (trees < 7 cm diameter at base
and >1 m tall) and shrubs (non-trees >1 m tall) were
measured for their density. At each point, herbs (nontree species < 1 m tall) were sampled for their
presence in two 25 25 cm quadrats. Seedlings (tree
species < 1 m tall) were also sampled for frequency
in two 1 1 m quadrats. Importance values were
calculated as: the average of relative frequency,
relative density and relative basal area for canopy
trees; relative densities of understorey sapling trees
and shrubs; and absolute frequencies for herbs and
tree seedling (Greig-Smith, 1983). Voucher specimens were stored in the herbarium at the Kananaskis
Field Stations.
The location of each site was measured with a global
positioning system that was differentially corrected for
an accuracy within 5 m. The stands were on-screen
digitised by their location on a 30-m resolution (or
30 30 m pixel size) Thematic Mapper 4 satellite
image. The image was taken on August 8, 1984, with
a sun elevation of 49.3j and azimuth 138.j. This scene
was used because of previous classification success
and knowledge of no large disturbance in the intervening period (Franklin et al., 1994) The image was
geocorrected to the Universal Trans-Mercator grid
using a previously corrected 25 m digital elevation
model (DEM). The DEM was part of the Alberta
(provincial) Base Mapping Digital Terrain Data System. Independent ground control points confirmed an
error of less than 1 pixel.
The non-forested land cover classes were bare
rock, alpine, lake, or stream. The training and testing
73
pixels characterizing these classes were identified by
visual interpretation of Thematic Mapper images and
air photos. A minimum of 100 pixels per class was
sampled.
2.4. Resource model description
The models of moisture input, temperature and
solar radiation are designed to characterise resource
variables as they are controlled by the terrain and
climate. The equations were simple so they have
easily measured input terms and intuitively understandable behaviour. The modeled resources were
independent of species specific interactions with the
environment, so that definitions of species tolerances
did not depend on a priori knowledge of species
distributions.
2.4.1. Moisture input (mi)
The water input to plants was determined by
climatological scale processes of precipitation and
hillslope scale processes of drainage (Beven and
Kirkby, 1979). This relationship was formalized by
multiplying the annual precipitation (Eq. (2)) by the
wetness index (Eq. (3)) so that the moisture input (Eq.
(1)) was determined by:
mi ¼ wp
ð1Þ
where mi was the moisture input (m3 per meter
contour length per year), w was the wetness index,
and p was the precipitation. The moisture input was a
physically based index which was correlated to the
actual moisture available to the plants.
The moisture input was calculated on an annual
basis because the known relationship between precipitation and elevation in the study area was for an
average year, and other climate data were not available across the elevation range in the study area.
Precipitation (m/year) was calculated from:
p ¼ 415:9 þ 0:636E
ð2Þ
where E was the elevation (m) from the DEM. This
relationship was empirically determined for a subbasin of the Kananaskis (Storr and Ferguson, 1972).
The 8-year study used eleven precipitation gauges at
elevations ranging from 1780 to 2285 m.
74
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
The local hillslope processes distribute the precipitation (Eq. (2)). The wetness was the amount of water
draining into a point, divided by the rate water would
be flowing through that point. For example, when two
sites have the same catchment area, the steeper one
will be drier. The model of hillslope wetness was
computed as:
w ¼ lnðA=tanBÞ
ð3Þ
where B was the slope angle in the site and A, the
catchment area (m2/m), was the size of the area that
collects the water that flows through 1 m of contour
length (Beven and Kirkby, 1979). This model of
hillslope processes has been shown to be effective,
despite its simplicity and the assumption of uniform
ground water loss (Beven et al., 1984; Wolock et al.,
1989). More detailed models that include soil texture
and variation in the water flux within the drainage
area (e.g. O’Loughlin, 1986) were not used because
the required data was not available.
The catchment area values for the wetness model
(Eq. (3)) were based on an accumulation algorithm
where the water from each pixel on the DEM flows
into the lowest elevation pixel of its eight neighbouring pixels (Jenson and Domingue, 1988; PCI, 1996).
The first step was to use this accumulation algorithm
to determine the upslope area of each pixel by
calculating how many pixels drained into each pixel.
This was done by tracing the flow of water upslope
from the pixel in question, until the point where the
pixels in the path are no longer the lowest in a
neighbourhood and water ceases to flow into them.
The second step was to calculate the median of a 3 3
pixel neighbourhood. This step was necessary to
smooth out some of the high variance in accumulation
that is a consequence of all of the flow from a pixel
going into a single neighbouring pixel (Tarboton,
1997). In the third step, the value one was added to
the median, so that each site had at least its own
climate water flux as input. In other words, the
minimum upslope area was one pixel. The last step
was to convert from pixels to meters, by multiplying
the number of pixels by the area of a pixel (900 m2),
and then dividing the product by 30 m, the length of a
pixel edge. The result was the area drained per meter
contour length for a site; assuming that water flows
out of a pixel edge.
The angle of the slope (B) was calculated as the
angle of the plane formed by the vector connecting the
left and right neighbour pixels, and the vector connecting the upper and lower neighbours of the pixel in
question (PCI, 1996).
2.4.2. Temperature
The mean July maximum temperature in degrees
Celcius (TJ) changed with elevation in meters (E) such
that:
TJ ¼ 32:7 0:00771E
ð4Þ
This linear regression relationship was determined
over an 8-year period for a sub-basin of the Kananaskis watershed. The study used seven weather
stations located from 1370 to 2285 m a.s.l. (Janz
and Storr, 1977). This study also found that the
highest temperatures occurred in July, therefore this
represents the optimal growing month.
2.4.3. Solar radiation
The incoming shortwave radiation for an average
July day was calculated using the DEM and Nikolov
and Zeller’s (1992) algorithm for estimating solar
radiation received by mountain slopes. The month
of July was used for consistency with the temperature
variable. The cloudiness estimate used long-term
averages of July relative humidity from the nearby
Banff weather station. The incoming radiation varies
with elevation, aspect and slope angle, which changed
considerably within the study area, as well as with
latitude and date (Nikolov and Zeller, 1992). The
horizon was set at 15j to compensate for the mountain
shadows on the predominantly N –S valley. The solar
incidence angle and solar altitude angle were calculated for each daylight hour of an average July day,
then averaged across the hours. The incoming solar
radiation was subsequently calculated for an average
July day.
2.5. Gradient analyses
Canonical correspondence analyses (CCA) were
run using standard topographical descriptors (elevation, slope tangent and cosine of aspect), and with the
modeled resources to compare the two methods of
defining habitats. CCA is considered the best general
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
purpose direct ordination method because it assumes a
unimodal frequency distribution and is relatively
robust against curvilinear distortion (Fasham, 1977;
Gauch et al., 1977). To reduce distortions caused by
species which were absent in a large number of sites
(see Swan, 1970), species present in < 60% of the
stands were deleted from the analysis. When ordinations were run with species present in < 10% to 60%
of the stands, the patterns of species and stand
distribution along the ordination axes were consistent.
However, when species presence was above or below
these thresholds, the pattern changed markedly.
2.6. Classification of biophysical and spectral data
Two-thirds of the pixels that were sampled in the
field were assigned to classes (defined below), and
used to train a maximum-likelihood classifier (PCI,
1996). This classifier used the distributions of the
training data on the resource gradients or on the
spectral bands to assign the rest of the pixels in the
study area to the classes. Three classifications were
preformed and compared to determine differences in
class discrimination. These classifications were bases
on the resource variables, spectral variables of TM
bands 3,4,5 (Nelson et al., 1984; Horler and Ahern,
1986) and the combination of all six variables.
2.6.1. Class definitions
The forest cover vegetation classes were defined
using measured abundances so that the accuracy of
the map was maximized. If a canopy tree species had
an importance value of >16% then that species was
included in the class name (e.g. the name lodgepole
pine was applied to stands that had only lodgepole
pine with a >16% importance value). The resulting
tree classes were: lodgepole pine, trembling aspen,
lodgepole pine and Engelman/white spruce, and subalpine fir and Engelman/white spruce. The non-forested classes of rock, alpine, deep water and shallow
water largely occupied the remainder of the landscape.
There are roads and small developments in the watersheds, but these represent less than 5% of the area.
Five herb and shrub species were used to test the
predictive ability of the resource gradients beyond
what can be mapped by remote sensing. The species
were chosen for their frequency in the stands. The
high frequency of a species represents a large enough
75
sample to derive signatures for the classification.
From among the common species, a subset was
chosen that represented a variety of growth habits.
Specifically, prickly rose is a large shrub, bearberry is
a small shrub, Arnica is flowering herb, P. apthosa is a
lichen, and feather moss.
Herb and shrub individual species classes were
defined in terms of presence/absence and in terms of
abundance classes. Low abundance was 7% to 38%.
Medium abundance was 38% to 69%. High abundance was importance greater than 69%. Absence
was defined as an importance value less than 7% and
presence was importance greater than 7%. Importance values less than 7% were assumed to be due
more to historical, dispersal, or stochastic factors than
physiological tolerances. The higher class thresholds
were based on adequate sample size to define signatures.
2.6.2. Accuracy assessment
The one-third of pixels with known classes, but not
used to train the classifier, were compared to the
output in order to test the accuracy of the map. The
overall accuracy was expressed as a percent of the test
pixels successfully assigned to the correct class. The
Kappa statistic (KHAT) is a statistical method of
assessing accuracy that takes into account the chance
of random agreement (Congalton and Green, 1999).
We used the KHAT and corresponding 95% confidence intervals as a null model to ensure that the
predictions were significantly better than a random
classification.
3. Results
3.1. Resource model behaviour
The distributions of stands showed the modelled
resources were independent. Linear regressions of
each pair-wise relationship between resources had r2
values less than 0.035. Despite the co-dependance on
the terrain, each resource variable represented a different aspect of the vegetation habitat.
3.1.1. Moisture input
The moisture input distribution among stands
reflected the interplay of the local and landscape
76
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
climate processes (Fig. 2a). Within lower elevations,
the gentle bases of long hillslopes had more moisture
than the steep, short sites. However, the same relationship was not found at high elevations, possibly
because most hillslopes are short, and therefore had
small catchment areas. The slopes at high elevations
had as much moisture input as gentle slopes at low
elevation because of the higher amounts of precipitation (Fig. 2a). The driest sites occurred on some of the
flat slopes because they were on the ridges and
Fig. 2. Relationships between the three modeled resource gradients and three common environmental descriptors. (a) Moisture response to slope
angle for two elevation classes. Only two classes were used here to illustrate the different relationships that occurred. (b) Temperature
correspondence to changing elevation. (c) Temperature correspondence to slope angle. (d) Solar radiation response to slope angle. (e) Solar
radiation response to aspect. (f) Lack of radiation correspondence to elevation.
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
77
Fig. 3. Canonical correspondence analysis of common species and modeled July temperature, moisture input and solar radiation.
therefore had very small catchment areas. These sites
were no sites with less than 6 m3/m moisture input
sampled at high elevations because ridges were generally not forested. The moisture input did not respond
to aspect because the drying associated with south
facing slopes was not incorporated into the model.
3.1.2. Temperature
Unlike the moisture input, July temperature only
varied with landscape processes. As expected from
Eq. (4), the July temperature was directly correlated
with elevation (Fig. 2b). The scatter around the
correlation was due to averaging the temperature
values among the pixels of a stand. There was a
spurious pattern of temperature with slope angle
where the warmest temperatures rarely occur on steep
sites (Fig. 2c). This pattern occurred because steep
slopes were less likely to occur where temperatures
were warmest within the watershed. Hillslopes tended
to be gentler at lower elevations compared to higher
elevations because of erosional and depositional
processes. Temperature did not vary with aspect
because the warming affect associated with south
facing slopes was not incorporated into the model
(Dymond, 1998).
3.1.3. Solar radiation
In contrast to temperature, the effects of slope
angle and slope aspect outweighed the elevational
affect on incoming solar radiation. The variation in
the radiation increased as slope angle increased
because steep, north facing slopes are more shadowed
and steep, south facing slopes intercept more beam
radiation (Fig. 2d). Radiation generally decreased
from south to north facing slopes as the amount of
time and exposure to direct beam radiation decreased
(Fig. 2e). As elevation increased, there is less atmosphere to absorb solar radiation, so more should reach
Table 1
Accuracy assessment for three land cover maps with different types
of data being classified
Spectral
classification
Resource variable
classification
Combination
classification
KHAT a
95% Confidence
interval of KHAT
Overall
accuracy (%)
0.61
0.030
68.3
0.60
0.029
67.5
0.79
0.036
83.2
a
KHAT accuracies are significantly different when their
confidence intervals do not overlap.
78
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
Fig. 4. Land cover classifications from three types of input data. (a) Classification of TM spectral data. (b) Predicted land cover using July
temperature, moisture input, and solar radiation. (c) Classification of modeled resources and TM spectral data.
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
the ground. This effect was not apparent over the
elevation range in this watershed, despite being part of
the algorithm (Fig. 2f).
79
class was mapped to cooler sites, but the accuracy was
low at 21.2%. Since this class represented the transition class, it was not surprising that 44% of the pine/
spruce sites were predicted as lodgepole pine or
3.2. Gradient analysis
Using these modeled resource in a direct ordination
explained slightly more variation in the vegetation
abundance than the traditional topographic descriptors. The CCA using temperature, moisture input and
radiation explained 17.2% of the species variance with
the first axis and 5.7% with the second axis. Using the
slope angle, aspect and elevation explained 16.3%
variance with the first axis and 3.8% with the second
axis.
The species –habitat relationships for canopy trees
can be described from the CCA results (Fig. 3).
Lodgepole pine and Aspen stands occupied higher
temperature habitats. Aspen stands tended to be found
in areas with higher moisture input and solar radiation, whereas pine habitat included the range of
moisture and solar radiation conditions. Spruce and
fir stands occupied lower temperature habitats, with
fir concentrated in lower moisture and solar radiation
habitats, and spruce in higher moisture and higher
solar radiation conditions. Pine and spruce occur
together in habitats with relatively moderate temperatures.
3.3. Land cover maps
The spectral classification of land cover had low
accuracy at mapping vegetation pattern (Table 1). A
large percentage of the forest was mapped to pure
lodgepole pine forests (61.9% accuracy) and the small
area in pine/spruce (18%) and spruce/fir forests (12%
accuracy) (Fig. 4a). While lodgepole pine was abundant in the watersheds, it did not dominate the landscape to such a degree. The most accurate classes
were the broad land cover classes of deep water
(97.3%), alpine (91.5%), and rock (81.6%).
Using the modeled resources produced a similarly
accurate representation of the spatial pattern on the
landscape (Table 1). Trembling aspen occupied the
warmer and higher radiation sites with 83.3% accuracy (Fig. 4b). Lodgepole pine occupied the warm and
lower radiation sites with 61.9% accuracy, but was
confused with the pine/spruce class. The pine/spruce
Fig. 5. (a) The distribution of prickly rose abundance along the
modeled resources. (b) The predicted pattern of prickly rose in the
forested part of the study area.
80
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
and 79.2%). Deep water areas were successfully
mapped 90.2%, and were only confused with shallow
water. The shallow water class was poorly mapped
(39.3%).
Combining the spectral with modeled resources
had the highest accuracy at representing the forest
cover pattern (Table 1). The KHAT accuracy was
significantly better than using the spectral data alone
or the resource gradients alone (Fig. 4c). The combination of spectral and resource variables improved the
distribution of the lodgepole pine and spruce/subalpine fir compared to either the spectral or the resource
variable classification. The aspen and pine/spruce
mapping was poor, at 52% and 36.4%, largely
because of confusion with lodgepole pine. The combination of variables also produced a more accurate
map of spruce/fir, rock, deep water, shallow water and
alpine (80%, 96.8%, 98.2%, 63.7%, 94.2% accuracy).
The most significant change from using spectral data
alone was the assignment of areas to pine/spruce or
spruce/fir rather than pine. The most significant
change from using resource data alone was the assignment of areas to spruce/fir rather than alpine.
3.4. Herb and shrub predicted maps
In addition to forest cover types, understorey
species with narrow niches defined by temperature,
solar radiation and moisture input were accurately
predicted from the resource variables. For example,
sites with high abundance of prickly rose were
clumped where July temperatures were highest, and
radiation levels were high and moderate (Fig. 5a). The
resulting prickly rose presence/absence classification
Table 2
The accuracy assessments of five maps of species presence or
absence
Fig. 6. (a) The distribution of A. cordifolia along the modeled
resources. (b) The predicted pattern of A. cordifolia on the forested
part of the landscape.
spruce/fir. Forests of spruce and subalpine fir occupied the coolest and wettest areas with 68% accuracy.
There was some confusion between spruce/fir stands
and pine/spruce or alpine. The rock and alpine landcover types were otherwise accurately mapped (78.8%
Species
KHAT a
95% Confidence
interval for KHAT
Percent
accuracy (%)
Prickly rose
Bearberry
Peltigera
Feather moss
Arnica
0.57
0.46
0.28
0.30
0.15
0.13
0.14
0.17
0.18
0.17
78.7
71.3
61.7
60.3
62.3
a
When the KHAT confidence interval does not overlap with
zero, then the map was significantly more accurate than a random
map.
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
81
Fig. 7. The predicted pattern of prickly rose relative abundance on the forested landscape.
of the forested part of the landscape was 78.7%
accurate (Fig. 5b). Prickly rose was predicted to be
present at lower elevations on flat and south facing
slopes, where sites tended to have relatively high July
temperatures and high levels of solar radiation.
The species presence/absence pattern could not be
predicted for species without a relationship to the
modeled resource variables. A. cordifolia clearly had
variation in abundance, but there was no obvious
pattern with the modeled gradients (Fig. 6a). The
presence/absence maximum likelihood classification
of A. cordifolia was 62.3% accurate (Fig. 6b). However, this was not significantly better than a null model
(Table 2). The accuracy of presence/absence classifications for a variety of understorey species ranged from
60% to 79% (Table 2). Four of the five classifications
82
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
Table 3
The accuracy assessments of maps of the relative abundance for five
species
Species
KHAT a
95% Confidence
interval for KHAT
Percent
accuracy (%)
Prickly rose
Bearberry
Peltigera
Feather moss
Arnica
0.46
0.40
0.35
0.18
0.11
0.12
0.13
0.15
0.12
0.14
63.2
61.0
61.0
36.0
39.5
a
When the KHAT confidence interval does not overlap with
zero, then the map was significantly more accurate than a random
map.
predicted significantly more of the spatial pattern than a
null model (Table 2). The predictions were successful
across the range of growth forms.
The spatial pattern of relative abundance was also
predictable for some species. For example, the relative abundance classification of prickly rose had
63.2% overall accuracy (Fig. 7). In this prediction,
the highest abundance of prickly rose occurred where
the temperatures were warmest and the solar radiation was highest. Prickly rose was predicted at lower
abundance in some of the cooler areas where radiation levels were high, such as on south facing slopes.
As with presence/absence, herb and shrub abundance
predictions were statistically better than the null
model, except for A. cordifolia (Table 3). The
decrease in accuracy from presence/absence to relative abundance was greater for some species than
others. For example, feather moss accuracy dropped
24.3%, whereas bearberry dropped less than 1%. Part
of the drop in accuracy was due to the limited
amount of ground truth data that was divided among
three classes (absent/low/medium) for bearberry and
four classes (absent/low/medium/high) for Peltigera,
instead of just presence or absence.
4. Discussion
Biophysical variables can explain as much, or more
variation in subalpine species composition than traditional environmental descriptors. Elevation, slope, and
aspect are correlated to subalpine vegetation (e.g.
Whittaker et al., 1968; Allen and Peet, 1990), but by
using some simple process models we can move
beyond descriptions into testing hypotheses about
what controls species patterns and predicting those
patterns. Since the incorporation of additional descriptors has helped explain vegetation variation in the past
(e.g. Zobel et al., 1976; Busing et al., 1993; McAuliffe, 1994), adding models of soil nutrient flow, soil
development, or evapotranspiration may further the
success of using process models. This success will be
limited to the variation in the species spatial patterns
that is controlled by deterministic relationships
(McCune and Allen, 1985; Ricklefs, 1987).
Simple, mechanistic models of temperature, solar
radiation and moisture can be used to improve vegetation maps in mountainous areas. The most accurate
map of subalpine land cover used the combination of
spectral and resource variable data. This result supports the improvement in classifications achieved
using environmental descriptors (Franklin, 1992;
Franklin et al., 1994). The resource models combined
hillslope processes that predict local communities
(e.g. Ostendorf and Reynolds, 1998) and regional
climate processes that predict subcontinental and
global vegetation patterns (e.g. Woodward and Williams, 1987; Tchebakova et al., 1994; Monserud and
Tchebakova, 1996). Improving the accuracy of the
predictions would require more sophisticated environmental models. For example, incorporating the shadowing effect of the surrounding mountains on solar
radiation may improve mapping on north facing
slopes.
Furthermore, using resource variables as one of the
input data sets can reduce the dependence of mapping
processes on spectral data. For example, areas of
cloud, shadow or saturation could be separately mapped using resource variables. The resulting classification could be merged with a spectral and resource
variable classification of the surrounding area.
The herb and shrub classifications showed that
species which have clumped distributions with respect
to the modeled resources were accurately predicted to
be present or absent (accuracy greater than 60% and
KHAT significantly better than random). The relative
abundance of some species was also accurately predicted. This approach to predicting individual species
is uncommon (Iverson and Prasad, 1998; Wiser et al.,
1998). However, there is a growing conservation,
restoration and ecosystem management need to understand and predict individual species and species richness variation across landscapes (Hill and Keddy,
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
1992; Franklin, 1998; Wiser et al., 1998). Improving
the predictive ability for more species will probably
require including more environmental processes, such
as the effect of canopy density on understory radiation, since not all species respond to the same variables (Hill and Keddy, 1992). Species that do not have
clumped distributions with respect to temperature,
radiation and moisture input may have distributions
predictable by other habitat variables that can be
modeled. The prediction of relative abundances would
likely be improved with more ground truth data.
Having more data can produce more reliable signatures for the classification, and better results (Congalton and Green, 1999).
Applying biophysical-based predictions to other
types of vegetation requires knowledge about the
resources controlling the species abundances. Gradient analyses and other studies in a particular area
can help determine which biophysical gradients are
most likely to successfully predict species distributions (e.g. Host et al., 1987; Davis and Goetz, 1990).
These studies will also help determine the scale at
which successful predictions may be possible. For the
Kananaskis and Spray watersheds, a combination of
local and landscape scale processes predicted the
landscape scale vegetation pattern.
5. Conclusion
The first purpose of this study was to test the
ability of general species – habitat relationships to
map the spatial pattern of vegetation across a mountainous landscape. The use of simple biophysical
models improved the maps of the Kananaskis and
Spray watersheds in three ways. First, the accuracy of
the land cover was improved by combining the
resource variables with spectral data. Second, the
presence and absence of sub-canopy shrub and herb
species can be accurately mapped using the resource
variables. Third, ecological or management sensitive
areas, due to rare or exotic species can be mapped
based on the resource variables.
The second purpose of this study was to provide
ecological tools to people mapping large areas. The
incorporation of topographical information is an
established tool to improve remote sensing products
(e.g. Franklin et al., 1994). Replacing the standard
83
topographical measurements with biophysical models
will provide more ecologically based information to
the mapping process. This is especially important
when the study area spans sub-continents or continents and the relationship between species and topography changes.
Acknowledgements
This work was supported by a Natural Sciences
and Engineering Research Council operating grant
and the NSERC Sustainable Forest Management
Network of Centres of Excellence. We thank Tara
Tchir, Kara Webster, Tonia Robb and Fran Dymond
for their assistance in the field and lab. This manuscript was substantially improved with input from
Kathy Gonzales, Volker Radeloff, Michael Collins,
Brian Amiro, Ronald Hall and two anonymous
reviewers.
References
Allen, R.B., Peet, R.K., 1990. Gradient analysis of forests of the
Sangre de Cristo Range, Colorado. Canadian Journal of Botany
68, 193 – 201.
Beven, K.J., Kirkby, M.J., 1979. A physically based, variable contributing area model of basin hydrology. Hydrological Science
Bulletin 24, 43 – 69.
Beven, K.J., Kirkby, M.J., Schofield, N., Tagg, A.F., 1984. Testing
a physically-based flood forecasting model (TOPMODEL) for
three U.K. catchments. Journal of Hydrology 69, 119 – 143.
Bolstad, P.V., Swank, W., Vose, J., 1998. Predicting Southern Appalachian overstory vegetation with digital terrain data. Landscape Ecology 13, 271 – 283.
Botkin, D.B., Janak, J.F., Wallis, J.R., 1972. Some ecological consequences of a computer model of forest growth. Journal of
Ecology 60, 948 – 972.
Brzeziecki, B., Kienst, F., Wildi, O., 1993. A simulated map of the
potential natural forest vegetation of Switzerland. Journal of
Vegetation Science 4 (4), 499 – 508.
Busing, R.T., White, P.S., MacKenzie, M.D., 1993. Gradient analysis of old spruce – fir forests of the Great Smoky Mountains
circa 1935. Canadian Journal of Botany 71, 951 – 958.
CCFM, 1998. Canada Forest Accord. Canadian Council of Forestry
Ministers, Ottawa, ON.
Congalton, R.G., Green, K., 1999. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices. Lewis Publishers,
New York.
Cottam, G., Curtis, J.T., Hale, B.W., 1953. Some sampling characteristics of a population of randomly dispersed individuals.
Ecology 34 (4), 741 – 757.
84
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
Davis, F.W., Goetz, S., 1990. Modeling vegetation pattern using
digital terrain data. Landscape Ecology 4 (1), 69 – 80.
Dettmers, R., Bart, J., 1999. A GIS modeling method applied to
predicting forest songbird habitat. Ecological Applications 9 (1),
152 – 163.
Dymond, C.C., 1998. Redefining species tolerances in a subalpine
forest by using energy and moisture processes. MSc thesis,
Department of Biological Sciences, University of Calgary, Calgary, Alberta, Canada.
Fasham, M.J.R., 1977. A comparison of nonmetric multidimensional scaling, principal components and reciprocal averaging for
the ordination of simulated coenoclines, and coenoplanes. Ecology 58 (3), 551 – 561.
Franklin, S.E., 1992. Satellite remote sensing of forest type and
landcover in the subalpine forest region, Kananaskis valley,
Alberta. Geocarta International 4, 25 – 35.
Franklin, J., 1998. Predicting the distribution of shrub species in
southern California from climate and terrain-derived variables.
Journal of Vegetation Science 9, 733 – 748.
Franklin, S.E., Connery, D.R., Williams, J.A., 1994. Classification of
alpine vegetation using landsat thematic mapper, spot HRV and
DEM data. Canadian Journal of Remote Sensing 20 (1), 49 – 56.
Gauch Jr., H.G., Whittaker, R.H., Wentworth, T.R., 1977. A comparative study of reciprocal averaging and other ordinations.
Journal of Ecology 65, 157 – 174.
Gleason, H.A., 1926. The individualistic concept of the plant association. Bulletin of the Torrey Botanical Club 53, 7 – 26.
Greig-Smith, P., 1983. Quantitative Plant Ecology, 3rd ed. University of California Press, Los Angeles.
He, H.S., Mladenoff, D.J., Radeloff, V.C., Crow, T.R., 1998. Integration of GIS data and classified satellite imagery for regional
forest assessment. Ecological Applications 8, 1072 – 1083.
Hill, N.M., Keddy, P.A., 1992. Prediction of rarities from habitat
variables: coastal plain plants on Nova Scotian lakeshores. Ecology 73 (5), 1852 – 1859.
Horler, D.N.H., Ahern, F.J., 1986. Forestry information content of
Thematic Mapper data. International Journal of Remote Sensing
7 (3), 405 – 428.
Host, G.E., Pregitzer, K.S., Ramm, C.W., Hart, J.B., Cleland, D.T.,
1987. Landform-mediated differences in successional pathways
among upland forest ecosystems in northwestern lower Michigan. Forest Science 33, 445 – 457.
Iverson, L.R., Prasad, A.M., 1998. Predicting abundance of 80 tree
species following climate change in the eastern United States.
Ecological Monographs 68 (4), 465 – 485.
Iverson, L.R., Dale, M.E., Scott, C.T., Prasad, A., 1997. A GISderived integrated moisture index to predict forest composition
and productivity of Ohio forests (U.S.A.). Landscape Ecology
12, 331 – 348.
Janz, B., Storr, D., 1977. The Climate of the Contiguous Mountain
Parks. Project No. 30, Applications and Consulting Division,
Environment Canada, Edmonton.
Jarvis, P.G., Leverenz, J.W., 1983. Productivity of temperate, deciduous and evergreen forests. In: Lange, O.L., Nobel, P.S.,
Osmond, C.B., Zeigler, H. (Eds.), Encyclopedia of Plant Physiology. Physiological Plant Ecology, Chap. 8, vol. 12D Springer-Verlag, New York.
Jenson, S.K., Domingue, J.O., 1988. Extracting topographic structure from digital elevation data for geographic information system analysis. Photogrammetric Engineering and Remote Sensing 54 (11), 1593 – 1600.
Johnson, E.A., Fryer, G.I., 1987. Historical vegetation change in the
Kananaskis Valley, Canadian Rockies. Canadian Journal of Botany 65, 853 – 858.
Johnson, E.A., Larsen, C.P.S., 1991. Climatically induced change in
fire frequency in the southern Canadian Rockies. Ecology 72
(1), 194 – 201.
Korzukhin, M.D., Ter-Mikaelian, M.T., Wagner, R.G., 1996. Process versus empirical models: which approach for forest ecosystem management? Canadian Journal of Forest Research 26,
879 – 887.
Larcher, W., 1995. Physiological Plant Ecology, 3rd ed. SpringerVerlag, New York.
La Roi, G.H., Hnatiuk, R.J., 1980. The Pinus contorta forests of
Banff and Jasper National Parks: a study in comparative synecology and syntaxonomy. Ecological Monographs 50 (1), 1 – 29.
Lenihan, J.M., 1993. Ecological response surfaces for North American boreal tree species and their use in forest classification.
Journal of Vegetation Science 4 (5), 667 – 680.
Levins, R., 1966. The strategy of model building in population
biology. American Scientist 54 (4), 421 – 431.
Lillesand, T.M., Kiefer, R.W., 1994. Remote Sensing and Image
Interpretation, 3rd ed. Wiley, New York.
McAuliffe, J.R., 1994. Landscape evolution, soil formation and
ecological patterns and processes in Sonoran desert bajadas.
Ecological Monographs 64 (2), 111 – 148.
McCune, B., Allen, T.F.H., 1985. Will similar forests develop on
similar sites? Canadian Journal of Botany 63, 367 – 376.
McMechan, M.E., 1988. Geology of Peter Lougheed Provincial
Park, Rocky Mountain Front Ranges, Alberta. Map (scale
1:50,000). Open file 2057 Geological Survey, Ottawa.
Monserud, R.A., Tchebakova, N.M., 1996. A vegetation model for
the Sayan Mountains, southern Siberia. Canadian Journal of
Forest Research 26, 1055 – 1068.
Nelson, R.F., Latty, R.S., Morr, G., 1984. Classifying northern
forests using Thematic Mapper simulator data. Photogrammetric
Engineering and Remote Sensing 50 (5), 607 – 617.
Nikolov, N.T., Zeller, K.F., 1992. A solar radiation algorithm for
ecosystem dynamic models. Ecological Modelling 61, 149 – 168.
Noest, V., 1994. A hydrology-vegetation interaction model for predicting the occurrence of plant species in dune slacks. Journal of
Environmental Management 40 (2), 119 – 128.
O’Loughlin, E.M., 1986. Prediction of surface saturation zones in
natural catchments by topographical analysis. Water Resources
Research 22 (5), 794 – 804.
Ostendorf, B., Reynolds, J.F., 1998. A model of arctic tundra vegetation derived from topographic gradients. Landscape Ecology
13, 187 – 201.
PCI, 1996. Version 6.1 EASI/PACE software. Richmond Hill.
Peet, R.K., 1981. Forest vegetation of the Colorado Front Range.
Vegetatio 45, 3 – 75.
Ricklefs, R.E., 1987. Community diversity: relative roles of local
and regional processes. Science 235, 167 – 171.
Running, S.W., Gower, S.T., 1991. FOREST-BGC, a general model
C.C. Dymond, E.A. Johnson / ISPRS Journal of Photogrammetry & Remote Sensing 57 (2002) 69–85
of forest ecosystem processes for regional applications. II: Dynamic carbon allocation and nitrogen budgets. Tree Physiology
9, 147 – 160.
Shugart Jr., H.H., West, D.C., 1977. Development of an Appalachian deciduous forest succession model and its application to
assessment of the impact of the Chestnut Blight. Journal of
Environmental Management 5, 161 – 179.
Stephenson, N.L., 1990. Climatic control of vegetation distribution: the role of the water balance. American Naturalist 135
(5), 649 – 670.
Storr, D., Ferguson, H.L., 1972. The distribution of precipitation in
some mountainous Canadian watersheds. Proceedings, Geilo
Symposium—Distribution of Precipitation in Mountainous
Areas. WMO/OMM No. 326, vol. 1. World Meterological Organization, Geneva, Switzerland, pp. 243 – 263.
Swan, J.M.A., 1970. An examination of some ordination problems
by use of simulated vegetation data. Ecology 51 (1), 89 – 102.
Tarboton, D.G., 1997. A new method for the determination of flow
directions and upslope areas in grid digital elevation models.
Water Resources Research 33 (2), 309 – 319.
Tchebakova, N.M., Monserud, R.A., Nazimova, D.I., 1994. A Siberian vegetation model based on climatic parameters. Canadian
Journal of Forest Research 24, 1597 – 1607.
Wentworth, T.R., 1981. Vegetation on limestone and granite in the
Mule Mountains, Arizona. Ecology 62 (2), 469 – 482.
85
Whittaker, R.H., 1973. Direct gradient analysis: techniques. In:
Whittaker, R.H. (Ed.), Handbook of Vegetation Science. Ordination and Classification of Communities, vol. 5. Dr. W. Junk,
The Hague.
Whittaker, R.H., Neiring, W.A., 1965. Vegetation of the Santa Catalina Mountains, Arizona: II. A gradient analysis of the south
slope. Ecology 46, 429 – 452.
Whittaker, R.H., Buoi, S.W., Neiring, W.A., Havens, Y.H., 1968. A
soil and vegetation pattern in the Santa Catalina Mountains,
Arizona. Soil Science 105 (6), 440 – 450.
Wiser, S.K., Peet, R.K., White, P.S., 1998. Prediction of rare-plant
occurrence: a southern Appalachian example. Ecological Applications 8 (4), 909 – 920.
Wolock, D.M., Hornberger, G.M., Beven, K.J., Campbell, W.G.,
1989. The relationship of catchment topography and soil hydraulic characteristics to lake alkalinity in the northeastern
United States. Water Resources Research 25, 829 – 837.
Woodward, F.I., Williams, B.G., 1987. Climate and plant distribution a global and local scales. Vegetatio 69, 189 – 197.
Zobel, D.B., McKee, A., Hawk, G.M., 1976. Relationships of environment to composition, structure, and diversity of forest communities of the central western Cascades of Oregon. Ecological
Monographs 46, 135 – 156.