Content uploaded by Jennifer Read
Author content
All content in this area was uploaded by Jennifer Read
Content may be subject to copyright.
A Robust Technique for Mapping
Vegetation Condition Across a Major
River System
S. C. Cunningham,
1
* R. Mac Nally,
1
J. Read,
2
P. J. Baker,
1
M. White,
3
J. R. Thomson,
1
and P. Griffioen
4
1
Australian Centre for Biodiversity, School of Biological Sciences, Monash University, Melbourne, Victoria 3800, Australia;
2
School of Biological Sciences, Monash University, Melbourne, Victoria 3800, Australia;
3
Arthur Rylah Institute for Environmental
Research, Department of Sustainability and Environment, Heidelberg, Melbourne, Victoria 3084, Australia;
4
acroMap,
37 Gloucester Drive, Heidelberg, Melbourne, Victoria 3084, Australia
ABSTRACT
Ecologists need to develop tools that allow char-
acterization of vegetation condition over scales that
are pertinent to species’ persistence and appropri-
ate for management actions. Our study shows that
stand condition can be mapped accurately over the
floodplain of a major river system (ca 100,000 ha of
forest over 1600 km of river)—the Murray River in
southeastern Australia. It demonstrates the value of
using quantitative ground surveys in conjunction
with remotely sensed data to model vegetation
condition over very large spatial domains. A com-
parison of four modelling methods found that stand
condition was best modelled using the multivariate
adaptive regression spline (MARS) method
(R
2
= 0.85), although there was little difference
among the methods (R
2
= 0.77–0.85). However, a
subsequent validation survey of condition at new
locations showed that use of artificial neural net-
works had substantially higher predictive power
(R
2
= 0.78) than the MARS model (R
2
= 0.28). This
discrepancy demonstrates the value of using sev-
eral modelling approaches to determine relation-
ships among vegetation responses and environ-
mental variables, and stresses the importance of
validating ecological models with predictive sur-
veys conducted after model building. The artificial
neural network was used to produce a stand con-
dition map for the whole floodplain, which pre-
dicted that only 30% of the area containing
Eucalyptus camaldulensis stands is currently in good
condition. There is a downstream decline in stand
condition, which is related to more extreme de-
clines in flooding, due to water harvesting, and
drier climate found in the Lower Murray region.
Rigorous surveying and modelling approaches,
such as those used here, are necessary if vegetation
health is to be effectively monitored and managed.
Key words: Eucalyptus camaldulensis; floodplains;
neural networks; regression splines; regression
trees; remote sensing; river regulation; validation;
vegetation condition.
INTRODUCTION
Much has been made of adverse effects of loss and
fragmentation of natural vegetation (Sala and
others 2000). Substantial work has also been
undertaken to relate biodiversity decline to degra-
dation of remnant vegetation (Saunders and others
Received 18 February 2008; accepted 19 October 2008;
published online 22 November 2008
Author Contribution: SCC wrote the paper and was involved in all parts
of the research; RM conceived and designed the study, and contributed to
analysis and writing; JR designed the study, and contributed to writing;
PJB designed the study and contributed to writing; MW contributed to
design and modelling; JRT contributed to modelling; PG contributed new
methods and modelling.
*Corresponding author; e-mail: shaun.cunningham@sci.monash.edu.au
Ecosystems (2009) 12: 207–219
DOI: 10.1007/s10021-008-9218-0
2008 Springer Science+Business Media, LLC
207
1991), but this typically is described for relatively
small locations (hectare scale, Parkes and others
2003). There has been much less work on linking
loss and degradation of vegetation to declines in
biota over very extensive spatial extents (land-
scapes to regions). Ecologists need to develop tools
that allow characterization of vegetation condition
over scales that are pertinent to species’ persistence
and over which management actions are or need to
be undertaken.
Extensive dieback has been reported in many
forest types since the 1960s (Huettl and Mueller-
Dombois 1993). Forest dieback is characterized by a
progressive reduction in the crowns of individual
trees that leads to widespread mortality. Forest
dieback affects the quantity and quality of forest
and water resources, the flora and fauna dependent
on these ecosystems, and beyond into associated
terrestrial and aquatic ecosystems. Causes of forest
dieback include changes in chemical composition
of the atmosphere (McLaughlin and Percy 1999),
soils (Czerniakowski and others 2006) and water
(Ohrui and Mitchell 1997), climate change (Bour-
que and others 2005), succession (Mueller-Dom-
bois 2006), and impacts of exotic insects (Bonneau
and others 1999).
Potential indicators of forest condition include
compositional (for example, species richness),
structural (for example, shrub cover) and func-
tional (for example, recruitment, Noss 1999) attri-
butes. Various assessment procedures have been
devised to estimate tree condition and these in-
clude a range of indicators, such as crown vigor,
crown size and foliage density (for example, Innes
1990; USDA Forest Service 2004). Many of these
assessments are based on subjective classifications
of crown condition, which are difficult to use in
temporal and statistical comparisons (Ferretti
1997). However, more objective measures such as
percentage live basal area, plant area index and
crown vigor can provide reliable indicators of stand
condition (Cunningham and others 2007).
Ground surveys are time consuming, costly and
geographically limited, whereas remote sensing,
which uses satellite or airborne data, provides wide
spatial coverage of an area at high resolution
(<1 m to 1 km) and at a lower cost per unit area.
Consequently, there has been increased use of re-
mote sensing in assessment of forest condition.
Remote sensing research in forests has often fo-
cused on estimation of leaf area index (LAI), which
is an indicator of forest productivity and water use
(for example, Nemani and others 1993). Assess-
ments of forest dieback using remote sensing have
concentrated on insect defoliation in boreal and
coniferous forests of Northern America (for exam-
ple, Bonneau and others 1999; Hall and others
2003) and rarely on other types of dieback or on
angiosperms (for example, Coops and others 2004).
Most studies relate forest dieback to vegetation
indices that are ratios of red and near infrared
reflectance bands, with Normalized Difference
Vegetation Index (NDVI, Rouse and others 1973)
being the most widely used.
Forest dieback has been modelled from remotely
sensed data with varying success (R
2
= 0.2–0.9, for
example, Diem 2002; Pontius and others 2005).
Models showing strong relationships between for-
est dieback and remotely sensed data (R
2
> 0.75)
were based on very limited surveys (N < 20, for
example, Hall and others 2003) and, therefore, are
unlikely to predict well beyond the original sites.
Previous forest dieback models were derived using
simple statistical methods (for example, linear
regression and correlation) and were not validated
with subsequent surveys, and some were not even
related to ground surveys (for example, Maselli
2004). Validation surveys comparing actual to
predicted conditions at new locations provide the
most compelling demonstration of the usefulness of
models (Mac Nally and Fleishman 2004; Thomson
and others 2007).
Floodplain ecosystems are of great ecological and
economic importance due to their high biodiversity
and productivity (Tockner and Stanford 2002).
Regulation of rivers using dams, channels and le-
vees has led to the decline of floodplain forests
throughout the world (for example, Busch and
Smith 1995). Rivers in semi-arid landscapes are
among the most susceptible to fluctuations in water
availability through decreases in precipitation, in-
creases in water extraction and changes to the sea-
sonality of flows. The Murray River, in southeastern
Australia, is the country’s longest river (2520 km),
providing water for extensive irrigation farming,
human consumption and power generation.
Increasing regulation of the Murray River since the
1920s has significantly reduced peak flows, and re-
duced frequency (35–62% of historical) and dura-
tion (40–84% of historical) of floods large enough to
connect anabranches to the river (MDBC 2005a).
Climate change is predicted to further reduce water
availability in southeastern Australia by raising
temperatures and decreasing annual rainfall
(CSIRO, BOM 2007), thereby increasing the sever-
ity of droughts (Cai and Cowan 2008).
River red gum (Eucalyptus camaldulensis Dehnh.)
is the dominant floodplain tree along the Murray
River and has the widest natural distribution of all
eucalypts, occupying watercourses throughout
208 S. C. Cunningham and others
mainland Australia (Boland and others 1984).
Qualitative surveys of the health of E. camaldulensis
forests suggest a decline in stand condition has
occurred in the lower Murray River over the past
20 years (Margules and Partners 1990; MDBC
2005b). Here, we present the first work to quantify
the decline of a keystone tree species by building a
robust map of stand condition over a very large
spatial domain (ca 1600 km of river, ca 100,000 ha
of forest) using an extensive ground survey
(N = 140 sites), remotely sensed data, a range of
advanced modelling methods and a predictive val-
idation survey (N = 42 new sites).
METHODS
Study Area
The study area included forests and woodlands of
Eucalyptus camaldulensis in Victoria, Australia, on the
floodplains of the Murray River from the Hume
Dam (36º06¢ S 147º01¢ E) to the South Australian
border (34º01¢ S 141º00¢ E), the lower Ovens River
downstream of Wangaratta and the lower Goulburn
River downstream of Shepparton (Figure 1). A GIS
layer of the current extent of E. camaldulensis on
these floodplains (30 9 30 m pixels) was deter-
mined by cropping a GIS layer of the inferred dis-
tribution of ecological vegetation classes (EVCs)
prior to European settlement (EVC1750) with a
layer of the extant distribution of woody vegetation
(Tree25), and only including EVCs that are domi-
nated by E. camaldulensis. EVCs are a classification of
vegetation types used by land managers in Victoria,
Australia, based on plant communities (for exam-
ple, life form, genera, vegetation structure), geo-
morphology and climate (DNRE 1997). Tree25 is a
layer derived from a combination of digital classifi-
cation and visual interpretation of SPOT panchro-
matic imagery to determine the presence of woody
vegetation. All patches of E. camaldulensis within
15 km of the Murray River in Victoria, the lower
Ovens River or the lower Goulburn River were in-
cluded. The study area comprised 103,550 ha of E.
camaldulensis forest and woodland. This included
open forests (10–30 m tall, 30–45% projective foli-
age cover) and woodlands (10–30 m tall, 20–25%
projective foliage cover (Specht 1981)) with
shrubby, sedgy and grassy understoreys (Margules
and Partners 1990). The climate across this area is
temperate and covers a wide range of annual pre-
cipitation (270–715 mm y
-1
), annual evaporation
(120–1790 mm y
-1
), mean annual maximum
temperatures (22.1–24.5C) and mean annual
minimum temperatures (8.9–15.2C, BOM 2007).
Site Selection
A random-stratified approach was used to select
sites covering a range of stand conditions. First,
sites were restricted to public land, which provided
an adequate coverage of the study area, to mini-
mize access problems. Second, the area was divided
into the five Bioregions (Victorian Riverina, Upper
Murray Fans, Lower Murray Fans, Robinvale Plains
and Murray Scroll Belt), which are defined by dif-
ferences in climate, geomorphology, lithology and
biodiversity (Environment Australia 2000). This
ensured good spatial coverage of the extensive
survey area and an ability to standardize data
within different ecological units. Third, selection
was restricted within each Bioregion to the pre-
dominant E. camaldulensis communities (EVCs) that
covered the largest area. EVCs were added
sequentially, starting with the EVC with the largest
area, until 65% or more of the total forest area
within the Bioregion was included. This ensured
that both riverine and floodplain stands were se-
lected and, therefore, included stands with a range
of flooding frequencies and potential stand condi-
tions. Finally, sites were selected within each EVC
according to a classification of reflectance values
from Landsat7 imagery from 2004 to ensure a
range of reflectance values and potentially a range
of stand conditions were included.
To classify the Landsat7 imagery from 2004,
10,000 pixels were randomly selected from each
Bioregion using Hawth’s Tools (http://www.spatia-
lecology.com). For each Bioregion, principal com-
ponents analysis (PCA) was used to determine the
relationship of Landsat7 data (six reflectance bands)
among these pixels. PCA loadings from the primary
axis were used to produce a combined reflectance
layer. The reflectance layer was then divided into the
25%, 25–75% and 75% quartiles of reflectance
values to ensure sites with low and high reflectance
values were well represented. GIS layers of the 45
stratification groups [Bioregions (5) 9 EVC (2–
4) 9 reflectance quartile (3)] were produced.
Twenty-eight point locations were selected from
each Bioregion (a total of 140 sites), with equal
numbers of sites randomly selected from each
stratification group. Each point location was at the
center of a 30 9 30 m pixel of the Landsat7 imagery.
Stand Condition Assessment
Previous research showed that percentage live ba-
sal area, plant area index and crown vigor were the
best indicators of condition in E. camaldulensis
stands (Cunningham and others 2007). Percentage
Mapping Vegetation Condition Across Regions 209
live basal area is the percentage of a stand’s basal
area that is live trees. Plant area index (PAI) is the
area of leaves and stems per unit ground area
without adjustment for clumping of canopy com-
ponents. Crown vigor is the percentage of the po-
tential crown, which is determined by the extent of
the existing branching structure that contains foli-
age. Epicormic growth, which was estimated as the
percentage of live foliage that contained epicormic
shoots, was included to discriminate between per-
sistent crown and potentially transient growth re-
sponses to recent watering events. Crown vigor and
epicormic growth were measured from 30 repre-
sentative trees. We measured these indicators in a
survey of 140 sites. Details of measurement pro-
cedures for these indicators are given in Cunning-
ham and others (2007). The condition of stands of
E. camaldulensis was surveyed between June and
October 2006, during the wettest months of the
year when the canopy should be at a maximum. A
0.25 ha plot was established at each point location.
Most plots were 50 9 50 m but rectangular plots
(width = 20–40 m) were used to assess linear
stands along watercourses.
Environmental Predictor Variables
Twenty-five environmental variables were assem-
bled as GIS layers, with a 30 9 30 m pixel resolu-
tion, using ArcInfo (ESRI, Redlands, California).
Environmental variables included mean annual
precipitation, mean annual temperature, combined
radiometric data for thorium and potassium (used
for mapping soil types, Cook and others 1996), a
tree density layer (Tree9) and Landsat7 data. Tree9
was derived from the tree cover layer (Tree25) by
calculating the number of 10 9 10 m pixels in the
3 9 3 array centered on a 30 9 30 m pixel that
contained woody vegetation. Landsat7 data layers,
which included six spectral bands (0.45–0.52, 0.52–
0.60, 0.63–0.69, 0.76–0.90, 1.55–1.75, 2.08–
2.35 lm), were obtained for 1989, 1991, 1992,
1995, 1998, 2000, 2002, 2004 and 2005. GIS layers
of the Normalized Difference Vegetation Index
(NDVI, Rouse and others 1973) were calculated for
each year. Layers of mean values and standard er-
rors were calculated for each band and NDVI from
the above nine years of data. The Landsat7 layers
used in the modelling included the six spectral
bands and NDVI for 2005 (7 layers) and for the
means and standard errors over the nine years of
data (14 layers). Estimates of flooding history were
not included because hydrological models and
digital elevation models with the resolution (sub-
metre) necessary in areas of low topography were
not available for the majority of the floodplain.
Model Building
A two-stage approach was taken to determine how
stand condition assessed during the ground survey
could be best modelled using spatial variation in
the above environmental (predominantly remotely
sensed) data, which had coverage across the study
area. First, artificial neural networks were used to
determine which combination of the condition
indicators was best modelled by the environmental
data and, therefore, was the best stand condition
index to map the floodplain. Second, we deter-
mined whether this stand condition index could be
better modelled from the environmental data by
other advanced modelling methods.
Feed-forward, multilayer perceptron artificial
neural networks learned by a back-propagation
Figure 1. Distribution of
Eucalyptus camaldulensis forests
on the Victorian Murray River
floodplain showing the extent of
the Bioregions (/).
210 S. C. Cunningham and others
algorithm (ANN, Rumelhart and others 1986) were
used to determine the best stand condition indicator
because they are a powerful technique for finding
patterns in ecological data, which rarely meet para-
metric statistical assumptions and commonly in-
volve nonlinear relationships. Artificial neural
networks make no prior assumptions about the form
of relationship between input variables and the dis-
tributions of the data (O
¨
zesmi and others 2006).
An artificial neural network was built for each
condition indicator using the Neural Networks
module within Statistica (StatSoft Inc., Tulsa,
Oklahoma). PAI was standardized by Bioregional
maxima to detrend the natural downstream decline
in PAI owing to the decline in productivity associ-
ated with reduced rainfall and increased evapora-
tion. Compound indices were also calculated to
determine if these were better modelled by the
environmental data than the individual indicators.
A series of compound indices (for example, PAI
+ crown vigor) was calculated by adding indicators
in order of how well they were modelled by the
environmental data. Each condition indicator
contributed a maximum of five points to the total
score of a compound index.
For each condition indicator, the modelling
procedure involved the following steps. Input data
(25 environmental variables) and output data
(condition indices) were standardized to unit
maxima. The dataset was divided into training (70
sites), selection (35 sites) and test data (35 sites).
The training data were used to build the artificial
neural network, the selection data were used to
check for over-fitting (that is, the model is too
specific to the idiosyncrasies of the training data)
and the test data were used to test the final fit of the
model. Models were built from 200 random starts
and the 50 models with the best statistical fit
(highest R
2
values) were retained. Selection of the
best model was based on parsimony, with a model
being considered the best fit when increasing
numbers of variables did not increase the amount
of variance explained by more than 2%. An upper
limit of 14 variables was set because it is recom-
mended to have at least ten times more samples
than there are variables to avoid overly complex
models (Burham and Anderson 2002).
In the past decade, modelling approaches based
on machine learning developed by computer sci-
entists, including artificial neural networks, have
been used successfully to predict the nonlinearities
and interactions of ecological data (for example,
Leathwick and others 2005; Moisen and others
2006). Three of these recent modelling advances
were used to assess if they could better model the
chosen stand condition index from the environ-
mental variables than by artificial neural networks.
First, boosted regression trees (BRT) were used be-
cause these select relevant environmental variables
and can also model interactions among variables.
BRTs overcome the inherent inaccuracies in seek-
ing a single parsimonious model by constructing an
ensemble of regression trees, which relate values of
a response (leaves) to its predictors through a series
of binary decisions or branches (Friedman 2001).
Second, Bayesian additive regression trees (BART)
modelling was used, which uses Bayesian model
averaging to determine the average regression tree.
BART has the qualities of BRT but also allows full
assessment of prediction uncertainty, and is
thought to be effective at finding simple models to
explain multivariate data (Chipman and others
2006). Finally, multivariate adaptive regression
spline (MARS) modelling combines linear regres-
sion, construction of splines and binary, recursive
partitioning to fit nonlinear functions with a series
of linear segments. MARS is related to regression
methods, such as generalized additive models, but
allows quicker fitting and simultaneous analysis of
multiple responses (Friedman 1991).
Model Validation
Given that internal validation procedures generally
are optimistic (Mac Nally 2000), a predictive vali-
dation survey was performed in April 2007 to assess
the ability of the models to predict stand condition
beyond the original survey sites. We chose to vali-
date the model in stands from two contrasting areas:
Gunbower Island in the Middle Murray (average
rainfall of 395 mm y
-1
,36
o
45¢ S 144
o
15¢ E) and
Wallpolla Island in the Lower Murray (285 mm y
-1
,
34
o
15¢ S 141
o
45¢ E). The artificial neural network
model, which predicts stand condition scores (SCS)
calculated from PAI, percentage live basal area and
crown vigor, was used to produce stand condition
maps for these two areas. Each condition indicator
contributed up to five points to a total score of 15
points. These maps were classified into areas of high
(SCS = 10.1–15.0), medium (SCS = 5.1–10.0) and
low (SCS = 0.2–5.0) conditions using ArcGIS (ESRI,
Redlands, California). In each area, seven sites were
randomly selected from each condition class using
Hawth’s Tools, giving 42 validation sites across the
two areas. The condition assessment procedure was
the same as the original survey except for the
exclusion of epicormic growth due to its poor pre-
dictive power. Eighteen sites from the original sur-
vey were resurveyed, with nine sites in each region,
to determine if the stand condition had declined
Mapping Vegetation Condition Across Regions 211
during the six months since the original survey. The
resurveyed sites were chosen to cover the full range
of stand condition within an area.
Model Evaluation
We evaluated and compared model fit using con-
ventional and unconventional R
2
relating modelled
to observed stand conditions. This evaluation was
done for four sets of models: (1) the artificial neural
networks built for the various condition indicators,
(2) the four models of the stand condition score
(ANN, BART, BRT, MARS) for the model-building
data, (3) the same four models but for resurveyed
data and (4) the same four models for the valida-
tion sites (previously unmeasured). For each of the
datasets, we fitted conventional (‘‘free’’) linear
regressions: P
i
= a + bO
i
, where P and O are the
predicted and observed values, a is the intercept
and b is the slope. Model fit can be computed using
the conventional R
2
approach (sums of deviations
from fitted values divided by sums of deviations
from
P). However, in the most useful models,
a ” 0 and b ” 1 (that is, there is perfect agree-
ment between the observed and predicted values, P
and O). Therefore, we also constrained regression
parameters to these values and recomputed pseu-
do-R
2
values, which need not lie between 0 and 1
but will be very poor fits for values of 0 or below.
We fitted free and constrained regressions using
Bayesian models in the WinBUGS software pack-
age (Spiegelhalter and others 2003). Regression
parameters were assigned uninformative normal
priors for the free regressions, for which we com-
puted conventional R
2
values. For the constrained
regressions, we assigned high-precision (SD =
0.001) priors with mean a ” 0 and mean b ” 1
and computed the pseudo-R
2
values. Within each
set of comparisons, we simultaneously fitted the
four model regressions (all share the same O-values)
and constructed 95% credible intervals of differ-
ences in model fit (either R
2
or pseudo-R
2
values).
These were then used to rank the relativities among
models in relation to their fits.
Map Analysis
The artificial neural network that predicted stand
condition score was used because of its superior
performance (see Results) to produce a stand con-
dition map for E. camaldulensis forests of the Vic-
torian Murray River floodplain. Stand condition
scores of less than 0.2 were excluded from analysis
because they were considered to be noise and often
coincided with the edges of water bodies. For
analysis, the map was classified into five condition
classes: good (SCS = 12.1–15.0), declining (SCS
= 9.1–12.0), poor (SCS = 6.1–9.0), degraded (SCS
= 3.1–6.0) and severely degraded (SCS = 0.2–3.0)
using ArcGIS.
Percentage of the total area included in each
condition class was determined from the number of
pixels in each class. The resurveyed and validation
datasets were combined (N = 50) to determine the
field accuracy of model predictions. Bayesian
models were used to determine the 25–75% cred-
ible interval for the predictions of stand condition.
The equations for the 25% and 75% credible limits
were used to produce the ‘worst case’ and ‘best
case’ predictions for stand condition across the
study area.
Multiple linear regression was used to determine
if stand condition was related to water availability.
It was assumed that water availability would be
related to rainfall and flooding frequency, which
decreases with distance downstream and is likely to
decline with distance from permanent water. We
used ArcGIS to create (1) a ‘permanent water’ layer
from a layer that contained all water bodies shown
on 1:100,000 topographic maps and (2) a distance
from the Hume Dam layer. A random selection of
25,000 points from the study area was created
using Hawth’s Tools and values for the condition
index, distance from the Hume Dam and distance
from permanent water were extracted for each
point. Multiple linear regression was performed on
this dataset, with stand condition score as the re-
sponse variable, and distance from the Hume Dam,
distance from permanent water and the interaction
between these variables as potential predictor
variables.
RESULTS
Modelling
The 140 sites selected using the random-stratified
approach covered a representative range of stand
condition. Stand condition was negatively skewed
across the study area, with half of the sites having
more than 90% live basal area, a third of stands
having 80–100% of their potential crown and two-
thirds of stands having less than 11% epicormic
growth in their crown. Values of plant area index
(PAI) were normally distributed between 0 and
2m
2
m
-2
.
The four stand condition indicators were mod-
elled using artificial neural networks (ANN) with
different degrees of success using the 25 environ-
mental variables (Table 1). The best ANN among
the single-indicator models was for the PAI data
212 S. C. Cunningham and others
(R
2
= 0.72). The PAI model used six environmental
variables and the variable contributing most to this
model was the Normalized Difference Vegetation
Index for 2005. In contrast, the best ANN for epi-
cormic growth used 13 environmental variables
and could not fit the data (R
2
= 0.09). Stand con-
dition indices, which were calculated from several
stand condition indicators, were more reliably
modelled by the environmental variables than the
individual stand condition indicators (Table 1). The
best ANN (R
2
= 0.77) was for a stand condition
index calculated from PAI, crown vigor and per-
centage live basal area. This model was found to be
most sensitive to reflectance in the near infrared
(0.76–0.90 lm) and far infrared (2.08–2.35 lm)
measured during the most recent year of available
data, 2005. The inclusion of epicormic growth in a
compound index using all stand condition indica-
tors resulted in a poorer model (R
2
= 0.72).
Regression parameters for predicted and ob-
served condition for the model-building dataset
suggest that all four modelling methods fitted the
data well (R
2
= 0.77–0.85), with multivariate
adaptive regression splines (MARS) fitting best and
ANN worst (Table 2). However, all models deviated
substantially from a perfect agreement between
observed and predicted values (a = 0 and b = 1),
with slopes well below unity. When regression
parameters were constrained (pseudo-R
2
), the
performance of ANN increased markedly relative to
Bayesian additive regression trees (BART) and
boosted regression trees (BRT). For resurvey sites,
MARS again performed better than the other three
modelling approaches for both free and uncon-
strained regressions (Table 2). Both MARS and
ANN had slopes exceeding unity (with inter-
cepts <0) whereas slopes for BART and BRT fell
well below 1.
Given the success of MARS with model-building
and resurvey data, a surprising outcome was the
poor performance of MARS (and BART, BRT) for
predicting conditions of new sites—slopes for
MARS, BART and BRT were less than 0.4 and
intercepts exceeded 5.8 (Table 2). ANN performed
exceedingly well for the new sites, with R
2
similar
to the initial model-building data (ca 0.78), the
pseudo-R
2
being similar to the free regression, the
slope being close to unity and the intercept close to
zero.
Map Characteristics
The artificial neural network model that predicted
stand condition score was used to produce a stand
condition map across the whole Victorian Murray
River floodplain (Figure 2). Any interpretation of
the map is contingent on two qualifications. First,
the map has been classified into five condition
classes and all classes other than good condition
should be considered suboptimal. Second, the map
does not distinguish, particularly downstream of
Robinvale, between areas that have poor condition
due to altered watering regimes and those that may
be in poor condition due to the natural limitations
of the environment (for example, infrequent
flooding, low rainfall, low soil nutrients).
The map predicted that across the Victorian
Murray River Floodplain only 30.1% of E. camal-
dulensis stands are currently in good condition
(Table 3). The remaining stands are in declining
Table 1. Results of Linear Regression Analyses of ‘Predicted’ Stand Condition Against ‘Observed’ Condition
for Artificial Neural Network Models Relating Stand Condition Indicators and Compound Indices to Envi-
ronmental Variables
Model Intercept Slope R
2
Pseudo-R
2
*
Condition indicators
PAI 0.28 ± 0.04 0.71 ± 0.04 0.720 ± 0.023 0.601 ± 0.031
V 1.31 ± 0.16 0.66 ± 0.04 0.675 ± 0.006 0.500 ± 0.005
LBA 0.15 ± 0.06 0.85 ± 0.07 0.654 ± 0.051 0.632 ± 0.052
E 0.67 ± 0.03 0.16 ± 0.04 0.086 ± 0.136 -3.287 ± 0.620
Compound indices
PAI + V 1.81 ± 0.24 0.73 ± 0.04 0.744 ± 0.004 0.639 ± 0.001
PAI + V + LBA 2.55 ± 0.38 0.75 ± 0.04 0.771 ± 0.003 0.689 ± 0.002
PAI + V + LBA + E 3.67 ± 0.56 0.73 ± 0.04 0.716 ± 0.004 0.623 ± 0.003
Relativity PVL > PV > PAI
PVLE > LBA V > E
PVL > PV LBA
> PVLE > PAI > V > E
Condition indicators included plant area index (PAI), crown vigor (V), % live basal area (LBA) and epicormic growth (E). Values are means ± standard deviations. See
Methods for explanation of pseudo-R
2
. *Intercept constrained to 0 and slope constrained to 1.
Mapping Vegetation Condition Across Regions 213
(53.5%), poor (10.6%), degraded (4.1%) or se-
verely degraded (1.7%) condition. The ‘worst case’
prediction (25% credible limit) for stand condition
is that no stands are in good condition and that the
majority are declining (73.5%) whereas the ‘best
case’ prediction (75% credible limit) for stand
condition is that 60.2% of stands are in good con-
dition.
There is a general trend of decreasing stand
condition with increasing distance downstream of
the Hume Dam (Figure 2). In the Middle Murray,
only 22% of stands in Barmah Forest are in good
condition, 22% are declining and 3% are in poor
condition (Figure 2A). In Barmah Forest, good
condition stands are generally close to the Murray
River whereas declining and poor stands are fur-
ther inland on the floodplain. In the Lower Mur-
ray, 21% of stands on Lindsay Island are in good
condition, 45% are declining and 34% are poor to
severely degraded (Figure 2B). In this region,
stands are naturally close to the river and ephem-
eral creeks due to the lower frequency of floods
compared to upstream. Good condition stands tend
to be restricted to the river channel on Lindsay Is-
land, whereas declining to severely degraded stands
are found along the river channel and anabranches
and on the floodplain.
Multiple linear regression of predicted stand
condition against distance from the Hume Dam and
distance from permanent water found a weak
relationship (F = 6816, P < 0.001, N = 25 000,
R
2
= 0.35) among these variables:
Stand condition ¼15:09 0:02 0:01 0:00
Hume 0:36 0:01water
where Stand condition is on a 15-point scale, Hume is
the linear distance from the Hume Dam (km), water
is the distance from permanent water (km) and
uncertainties are standard errors.
DISCUSSION
Modelling
Condition of E. camaldulensis stands was modelled
successfully from remotely sensed and spatially
modelled (for example, rainfall) environmental
variables using artificial neural networks. Plant area
index (PAI) was the condition indicator best mod-
elled by the environmental variables (Table 1). This
is not surprising because the model used the Nor-
malized Difference Vegetation Index (NDVI), which
is a measure of the amount of vegetation, and PAI,
Table 2. Results of Linear Regression Analyses of ‘Predicted’ Stand Condition Against ‘Observed’ Condition
for the Four Models of the Original Survey, Resurvey and Validation Datasets
Model Intercept Slope R
2
Pseudo-R
2
*
Original survey (model building, N = 140 sites)
ANN 2.54 ± 0.38 0.75 ± 0.03 0.771 ± 0.003 0.690 ± 0.002
BART 3.38 ± 0.32 0.67 ± 0.03 0.798 ± 0.003 0.612 ± 0.003
BRT 3.13 ± 0.32 0.70 ± 0.03 0.806 ± 0.003 0.658 ± 0.003
MARS 2.02 ± 0.31 0.80 ± 0.03 0.846 ± 0.002 0.797 ± 0.002
Relativity MARS > BRT
> BART > ANN
MARS > ANN
> BRT > BART
Resurvey (N = 18 resurveyed sites)
ANN -2.75 ± 1.87 1.22 ± 0.19 0.691 ± 0.039 0.678 ± 0.001
BART 1.63 ± 1.03 0.79 ± 0.11 0.744 ± 0.032 0.702 ± 0.004
BRT 2.30 ± 1.35 0.74 ± 0.14 0.607 ± 0.047 0.564 ± 0.003
MARS -1.27 ± 1.14 1.08 ± 0.12 0.824 ± 0.022 0.817 ± 0.002
Relativity MARS > BART
>
ANN > BRT
MARS > BART
> ANN > BRT
Validation survey (N = 42 new sites)
ANN 0.62 ± 0.63 0.88 ± 0.07 0.782 ± 0.011 0.771 ± 0.002
BART 5.87 ± 0.48 0.38 ± 0.05 0.546 ± 0.023 -1.115 ± 0.007
BRT 6.29 ± 0.52 0.35 ± 0.06 0.456 ± 0.027 -1.461 ± 0.007
MARS 6.01 ± 0.76 0.36 ± 0.08 0.278 ± 0.035 -0.772 ± 0.004
Relativity ANN MARS
> BART > BRT
ANN MARS
> BART > BRT
Modelling methods included artificial neural networks (ANN), Bayesian additive regression trees (BART), boosted regression trees (BRT), and multivariate adaptive regression
splines (MARS). Values are means ± standard deviations. See Methods for explanation of pseudo-R
2
. *Intercept constrained to 0 and slope constrained to 1.
214 S. C. Cunningham and others
Figure 2. Condition map for Eucalyptus camaldulensis stands along the Victorian Murray River floodplain predicted from
the neural network model, with insets from (A) Lindsay Island and (B) Barmah Forest. Permanent water bodies are
indicated in blue.
Mapping Vegetation Condition Across Regions 215
which is similar to leaf area index, is a measure of
the amount of canopy. Leaf area index of forests has
been modelled successfully (R
2
= 0.7–0.9) from re-
motely sensed data (for example, Coops and others
1997; Heiskanen 2006) but success differs widely
among forest types (R
2
= 0.5–0.8, for example,
Eklundh and others 2003). Epicormic growth was
not modelled well by the environmental variables
(R
2
= 0.23) because it was unrelated to stand con-
dition (r = -0.06, P = 0.48, N = 140). Epicormic
growth is an indicator of recent stress in eucalypts,
so it appears that many stands of good condition are
experiencing water stress that is mitigated by
watering events (rainfall or flooding) or other
environmental stresses such as structural damage,
disease or insect defoliation (Burrows 2002). The
best model of stand condition of E. camaldulensis was
built using a condition index calculated from three
indicators (PAI, crown vigor and percentage live
basal area) previously shown to be reliable, objec-
tive indicators of stand condition (Cunningham and
others 2007). Reflectance in the near infrared and
far infrared measured during 2005, which are sen-
sitive to vegetation structure and vegetation mois-
ture, respectively, was found to explain the most
variance in this model of stand condition.
Stand condition, estimated using the above in-
dex, was modelled from the environmental vari-
ables with varying success using a range of
contemporary statistical methods (Table 2). Multi-
variate adaptive regression splines (MARS) were
marginally best at modelling the original stand
condition data. Recent statistical methods (for
example, MARS and artificial neural networks)
have been shown to outperform more commonly
used modelling methods (for example, logistic
regression and generalized additive models) for
predicting species distributions (Elith and others
2006; Mun
˜
oz and Felicisimo 2004). However, dif-
ferences among contemporary methods such as
those used here (MARS, artificial neural networks
and boosted regression trees) are often small (for
example, Elith and others 2006; Moisen and Fres-
cino 2002).
The real test of predictive power is how well a
model predicts samples beyond the original dataset,
especially samples collected after model construc-
tion (Mac Nally and Fleishman 2004). The artificial
neural network was a much better predictor of
stand condition in the validation dataset than the
other modelling methods. We found that models
with marginally better fits of the original survey
data (for example, MARS) were substantially worse
at predicting stand condition in the validation
survey (Table 2). This emphasizes the importance
of validating models with predictive surveys. This
suggests all methods besides artificial neural net-
works were over-fitting the models by incorporat-
ing information representative of all stands as well
as idiosyncrasies of the original survey data. It was
surprising that artificial neural network modelling
had the highest predictive power because one of its
major criticisms is over-fitting datasets (Paruelo and
Tomasel 1997). Over-fitting may have been pre-
vented here by restricting the number of predictor
variables to a tenth of the number of samples and
by the use of the quasi-independent ‘selection’
dataset, which was from the original survey data
but only used to test the model and not to build the
model.
To our knowledge, this stand condition model for
E. camaldulensis is the only forest dieback model to
have been rigorously validated. The majority of
forest dieback models are based on single surveys,
which often have few sample plots (n < 20, for
example, Hall and others 2003), and some were not
tested against field measurements (for example,
Maselli 2004). Although many forest diebacks have
been modelled successfully using remotely sensed
data (for example, Pontius and others 2005), the
predictive power of these models cannot be known
without a subsequent survey. We are very confi-
dent of the predictions of our stand condition
model for E. camaldulensis due to the high predictive
Table 3. Predictions of the Artificial Neural Network for Stand Condition Across the Victorian Murray
Floodplain
Stand condition Mean
prediction (%)
Worst case
(25% CI)
Best case
(75% CI)
Good 30.1 0 60.2
Declining 53.5 73.5 31.1
Poor 10.6 18.4 7.1
Degraded 4.1 6.0 1.7
Severely degraded 1.7 2.1 0
CI = credible interval.
216 S. C. Cunningham and others
power (R
2
= 0.78) shown by the validation con-
ducted in new sites.
Map Characteristics
The stand condition map predicts that only 30.1%
of E. camaldulensis stands are currently in good
condition along the Victorian Murray River flood-
plain (Figure 2, Table 3). Past surveys of the con-
dition of E. camaldulensis forests have shown an
apparently substantial decline in stand condition in
the lower Murray River (downstream of Swan Hill)
over the past 20 years (Margules and Partners
1990; MDBC 2005b). These surveys used qualita-
tive assessments of crown condition from aerial
photographs and field observations at point loca-
tions. Our study is the first rigorous quantifica-
tion—stand condition assessed by an extensive,
quantitative survey and predicted by a validated
model—of the extent of the crisis faced by this
floodplain ecosystem.
The cause of the decline in condition of E. cam-
aldulensis stands along the Victorian Murray River
is complex but is ultimately due to river regulation.
The biomass of E. camaldulensis forests is consider-
ably higher than would be predicted from the
rainfall of their distribution, suggesting the forests
are dependent on additional water from ground-
water and surface flooding. This is supported by the
increased growth rate of E. camaldulensis after
flooding (Bacon and others 1993) and the pre-
dominant use of groundwater by trees growing
next to ephemeral streams or on floodplains
(Mensforth and others 1994; Thorburn and others
1993). Regulation of flows along the Murray River
has decreased the frequency and extent of floods,
which has lowered groundwater depths and re-
duced leaching of salts drawn up from naturally
saline groundwater, and has raised river levels near
dams, which has consequently raised adjacent
groundwater depths (Jolly 1996). Declines in con-
dition of E. camaldulensis have been observed in
association with these declines in water availability
and quality (DEH 2005;MDBC2005b).
The map shows a downstream decline in stand
condition (that is, condition decreases from east to
west) of E. camaldulensis forests and woodlands
(Figure 2). This trends supports previous observa-
tions of substantial declines in stand condition in
the Lower Murray (for example, Margules and
Partners 1990). The downstream decline in stand
condition probably is related to the larger declines
in flooding frequency in the Lower Murray (from
76 to 35 y century
-1
) compared with the Middle
Murray (from 92 to 57 y century
-1
,MDBC2005a).
The substantially lower rainfall (270 mm y
-1
) and
higher evaporation (1790 mm y
-1
) in the Lower
Murray compared with upstream areas (ca
700 mm y
-1
and 1400 mm y
-1
, respectively)
exacerbates water stress in stands of this region.
Climate change is predicted to further reduce water
availability by raising temperatures (1–1.5C and
consequently increasing evaporative demand) and
decreasing annual rainfall (2–5%) in southeastern
Australia by 2030 (CSIRO, BOM 2007). Therefore,
the substantial declines currently observed in the
Lower Murray may progress upstream in the near
future as flooding frequencies are further reduced
by climate change.
In many areas, the map shows that good condi-
tion stands are found near permanent water (for
example, Murray River, Figure 2A, B). A weak
relationship was found between predicted stand
condition and distance downstream from the Hume
Dam and distance from permanent water. Al-
though these two variables probably are good
indicators of flooding frequency, the relationship
explained only 35% of the predicted variation in
stand condition. At the local scale, stand condition
is also likely to be related to other variables such as
topography, soil type, soil salinity, groundwater
depth and salinity, and disturbance history (for
example, logging and grazing).
The map suggests that current watering regimes
(rainfall and flooding) are insufficient to maintain
the majority of E. camaldulensis stands in good
condition. If the condition of these forests is to be
maintained and potentially improved, the fre-
quency of floods needs to approach that of unreg-
ulated rivers. In recent years, there have been
artificial floods through overbank pumping and
opening regulators in limited areas. Our map pro-
vides a tool for land managers to prioritize available
water allocation and to monitor success of watering
programs.
We have developed a powerful tool for mapping
stand condition over very large areas, but we need
to establish the process-based links to stand health
(for example, physiological stress, productivity,
nutrient cycling). We are also in an excellent po-
sition to establish whether stand condition is con-
sistently linked to effects on biodiversity. Are biotic
components consistently altered in stands of com-
parable condition? Poor condition clearly is asso-
ciated with reduced leaf area, which may affect
occurrence of herbivores and understorey plants.
Other processes such as flowering may also be re-
duced, affecting recruitment and capacity to sustain
populations of nectar-feeding animals. If such links
can be established, then remotely sensed data,
Mapping Vegetation Condition Across Regions 217
which provides knowledge of vegetation condition,
may also provide important information on biodi-
versity status.
CONCLUSIONS
Our study shows that stand condition can be
mapped accurately over the floodplain of a major
river system (ca 100,000 ha). It reinforces the value
of using ground surveys in conjunction with re-
motely sensed data to model vegetation condition
across landscapes and regions. It demonstrates the
value of using several modelling approaches to
determine relationships between vegetation re-
sponses and environmental variables. The study
also illustrates the importance of validating eco-
logical models with a predictive survey, which is
rarely done (for example, Fleishman and others
2002). If the health of ecosystems is to be moni-
tored and managed effectively across large geo-
graphic areas, such rigorous modelling approaches
are necessary.
ACKNOWLEDGEMENTS
This research was funded by an ARC linkage Grant
(LP0560518, which was partially funded by the
Victorian Department of Sustainability and Envi-
ronment (DSE) and four Catchment Management
Authorities (Mallee CMA, North Central CMA,
Goulburn-Broken CMA, and North East CMA)).
We thank Rachael Nolan for assistance with field-
work. Environmental data layers were supplied by
the DSE’s Corporate Geospatial Data Library and
the Australian Greenhouse Office (Landsat7 data).
This is publication No. 146 from the Australian
Centre for Biodiversity: Analysis, Policy and Man-
agement.
REFERENCES
Bacon PE, Stone C, Binns DL, Leslie DJ, Edwards DW. 1993.
Relationships between water availability and Eucalyptus cam-
aldulensis growth in a riparian forest. J Hydrol 150:541–61.
Boland DJ, Brooker MIH, Chippendale GM, Hall N, Hyland
BPM, Johnston RD, Kleinig DA, Turner JD. 1984. Forest trees
of Australia. Melbourne: CSIRO Publications, p 687.
BOM. 2007. Climate data online: Australian Government Bu-
reau of Meteorology. http://www.bom.gov.au/climate/avera-
ges/. Accessed 24 Oct 2007.
Bonneau LR, Shields KS, Civco DL. 1999. A technique to iden-
tify changes in hemlock forest health over space and time
using satellite image data. Biol Invasions 1:269–79.
Bourque CPA, Cox RM, Allen DJ, Arp PA, Meng FR. 2005.
Spatial extent of winter thaw events in eastern North Amer-
ica: historical weather records in relation to yellow birch de-
cline. Glob Change Biol 11:1477–92.
Burham KP, Anderson DR. 2002. Model selection and multi-
model inference: a practical information-theoretic approach.
New York: Springer, p 488.
Burrows GE. 2002. Epicormic strand structure in Angophora,
Eucalyptus and Lophostemon (Myrtaceae)—implications for fire
resistance and recovery. New Phytol 153:111–31.
Busch DE, Smith SD. 1995. Mechanisms associated with decline
of woody species in riparian ecosystems of the southwestern
United States. Ecol Monogr 65:347–70.
Cai W, Cowan T. 2008. Evidence of impacts from rising tem-
perature on inflows to the Murray-Darling Basin. Geophys
Res Lett 35:L07701.
Chipman HA, Geogre EI, McCulloch RE. 2006. BART: Bayesian
Additive Regression Trees. Chicago: Graduate School of
Business, University of Chicago, p 31.
Cook SE, Corner RJ, Groves PR, Grealish GJ. 1996. Use of air-
borne gamma radiometric data for soil mapping. Aust J Soil
Res 34:183–94.
Coops N, Delahaye A, Pook E. 1997. Estimation of eucalypt
forest leaf area index on the South Coast of New South Wales
using Landstat MSS data. Aust J Bot 45:757–69.
Coops NC, Stone C, Culvenor DS, Chisholm L. 2004. Assessment
of crown condition in eucalypt vegetation by remotely sensed
optical indices. J Environ Qual 33:956–64.
CSIRO, BOM. 2007. Climate change in Australia. Melbourne:
CSIRO, p 148.
Cunningham SC, Read J, Baker PJ, Mac Nally R. 2007.
Quantitative assessment of stand condition and its relation-
ship to physiological stress in stands of Eucalyptus camaldul-
ensis (Myrtaceae) in southeastern Australia. Aust J Bot
55:692–9.
Czerniakowski B, Crnov R, Smith IW, Luck JE. 2006. Soil
properties associated with the tree decline ‘Mundulla Yel-
lows’. Plant Soil 285:197–206.
DEH. 2005. Floristic vegetation and tree health mapping, River
Murray floodplain, South Australia. Adelaide: South Austra-
lian Government Department of Environment and Heritage.
Diem JE. 2002. Remote assessment of forest health in southern
Arizona, USA: evidence for ozone-induced foliar injury.
Environ Manage 29:373–84.
DNRE. 1997. Victoria’s biodiversity: directions in management.
Melbourne: Victorian Department of Natural Resources and
Environment.
Eklundh L, Hall K, Eriksson H, Ardo
¨
J, Pilesjo
¨
P. 2003. Investi-
gating the use of Landsat thematic mapper data for estimation
of forest leaf area index in southern Sweden. Can J Remote
Sens 29:349–62.
Elith J, Graham CH, Anderson RP, Dudik M, Ferrier S, Guisan A,
Hijmans RJ, Huettmann F, Leathwick JR, Lehmann A, Li J,
Lohmann LG, Loiselle BA, Manion G, Moritz C, Nakamura M,
Nakazawa Y, Overton JM, Peterson AT, Phillips SJ, Richard-
son K, Scachetti-Pereira R, Schapire RE, Soberon J, Williams
S, Wisz MS, Zimmermann NE. 2006. Novel methods improve
prediction of species’ distributions from occurrence data.
Ecography 29:129–51.
Environment Australia. 2000. Revision of the Interim Biogeo-
graphic Regionalisation of Australia (IBRA) and the develop-
ment of version 5.1—summary report. Canberra: Department
of Environment and Heritage. p 37.
Ferretti M. 1997. Forest health assessment and monitor-
ing—issues for consideration. Environ Monit Assess 48:45–72.
218 S. C. Cunningham and others
Fleishman E, Mac Nally R, Fay JP. 2002. Validation tests of
predictive models of butterfly occurrence based on environ-
mental variables. Conserv Biol 17:806–17.
Friedman JH. 1991. Multivariate adaptive regression splines.
Ann Stat 19:1–141.
Friedman JH. 2001. Greedy function approximation: a gradient
boosting machine. Ann Stat 29:1189–232.
Hall RJ, Fernandes RA, Hogg EH, Brandt JP, Butson C, Case BS,
Leblanc SG. 2003. Relating aspen defoliation to changes in leaf
area derived from field and satellite remote sensing data. Can
J Remote Sens 29:299–313.
Heiskanen J. 2006. Estimating aboveground tree biomass and
leaf area index in a mountain birch forest using ASTER sa-
tellite data. Int J Remote Sens 27:1135–58.
Huettl RF, Mueller-Dombois D. 1993. Forest decline in the
Atlantic and Pacific region. Berlin: Springer-Verlag, p 366.
Innes JL. 1990. Assessment of tree condition. London: Her
Majesty’s Stationery Office, p 96.
Jolly ID. 1996. The effects of river management on the hydrol-
ogy and hydroecology of arid and semi-arid floodplains. In:
Anderson MG, Walling DE, Bates PD, Eds. Floodplain pro-
cesses. Chichester: John Wiley and Sons. p 577–609.
Leathwick JR, Rowe D, Richardson J, Elith J, Hastie T. 2005.
Using multivariate adaptive regression splines to predict the
distribution of New Zealand’s freshwater diadromous fish.
Freshw Biol 50:2034–52.
Mac Nally R. 2000. Regression and model-building in conser-
vation biology, biogeography and ecology: the distinction
between—and reconciliation of—‘predictive’ and explana-
tory’ models. Biodivers Conserv 9:655–71.
Mac Nally R, Fleishman E. 2004. A successful predictive model
of species richness using ‘indicator’ species. Conserv Biol
18:646–54.
Margules and Partners. 1990. Riparian vegetation of the River
Murray. Report prepared by Margules and Partners Pty. Ltd.,
P. & J. Smith Ecological Consultants and Department of
Conservation Forests and Lands. Canberra: Murray-Darling
Basin Commission. p 187.
Maselli F. 2004. Monitoring forest conditions in a protected
Mediterranean coastal area by the analysis of multiyear NDVI
data. Remote Sens Environ 89:423–33.
McLaughlin S, Percy K. 1999. Forest health in North America:
some perspectives on actual and potential roles of climate and
air pollution. Water Air Soil Pollut 116:151–97.
MDBC. 2005a. The Living Murray Foundation Report on the
significant ecological assets targeted in the First Step Decision.
Canberra: Murray-Darling Basin Commission, p 327.
MDBC. 2005b. Survey of River Red Gum and Black Box Health
along the River Murray in New South Wales, Victoria and
South Australia—2004. Canberra: Murray-Darling Basin
Commission, p 24.
Mensforth LJ, Thorburn PJ, Tyerman SD, Walker GR. 1994.
Sources of water used by riparian Eucalyptus camaldulensis
overlying highly saline groundwater. Oecologia 100:21–8.
Moisen GG, Frescino TS. 2002. Comparing five modelling
techniques for predicting forest characteristics. Ecol Model
157:209–25.
Moisen GG, Freeman EA, Blackard JA, Frescino TS, Zimmer-
mann NE, Edwards TCJ. 2006. Predicting tree species presence
and basal area in Utah: a comparison of stochastic gradient
boosting, generalized additive models, and tree-based meth-
ods. Ecol Model 199:176–87.
Mueller-Dombois D. 2006. Long-term rain forest succession and
landscape change in Hawai’i: the ‘Maui Forest Trouble’
revisited. J Veg Sci 17:685–92.
Mun
˜
oz J, Felicisimo AM. 2004. Comparison of statistical meth-
ods commonly used in predictive modelling. J Veg Sci 15:285–
92.
Nemani R, Pierce L, Running S, Band L. 1993. Forest ecosystem
processes at the watershed scale—sensitivity to remotely-
sensed leaf-area index estimates. Int J Remote Sens 14:2519–
34.
Noss RF. 1999. Assessing and monitoring forest biodiversity: a
suggested framework and indicators. For Ecol Manag
115:135–46.
Ohrui K, Mitchell MJ. 1997. Nitrogen saturation in Japanese
forested watersheds. Ecol Appl 7:391–401.
O
¨
zesmi SL, Tan CO, O
¨
zesmi U. 2006. Methodological issues in
building, training and testing artificial neural networks in
ecological applications. Ecol Model 195:83–93.
Parkes D, Newell G, Cheal D. 2003. Assessing the quality of
native vegetation: the ‘habitat hectares’ approach. Ecol
Manage Restoration 4:S29–38.
Paruelo JM, Tomasel F. 1997. Prediction of functional charac-
teristics of ecosystems: a comparison of artificial neural net-
works and regression models. Ecol Model 98:173–86.
Pontius J, Hallett R, Martin M. 2005. Using AVIRIS to assess
hemlock abundance and early decline in the Catskills, New
York. Remote Sens Environ 97:163–73.
Rouse JW, Haas RH, Schell JA, Deering DW. 1973. Monitoring
vegetation systems in the Great Plains with ERTS. In: Third
Earth Resources Technology Satellite-1 Symposium, 10–14
December 1973. Washington, DC, United States of America:
NASA. p 309–17.
Rumelhart DE, Hinton GE, Williams RJ. 1986. Learning repre-
sentations by back-propagating errors. Nature 323:533–6.
Sala O, Chapin F, Armesto J, Berlow E, Bloomfield J, Dirzo R,
Huber-Sanwald E, Huenneke L, Jackson R, Kinzig A, Leemans
R, Lodge D, Mooney H, Oesterheld M, Poff N, Sykes M,
Walker B, Walker M, Wall D. 2000. Global biodiversity sce-
narios for the year 2100. Science 287:1770–4.
Saunders DA, Hobbs RJ, Margules CR. 1991. Biological conse-
quences of ecosystem fragmentation: a review. Conserv Biol
5:18–32.
Spiegelhalter D, Thomas A, Best N. 2003. WinBUGS version 1.4
user manual. Bayesian inference using Gibbs sampling.
Cambridge: MRC Biostatistics Unit, Institute for Public Health.
USDA Forest Service. 2004. Forest Inventory and Analysis South-
ern Research Station Field Guide. Version 1.71, Phase 3. Volume
2: field data collection procedures for Phase 3 plots. Washington:
United States Department of Agriculture, Forest Service.
Specht RL. 1981. Major vegetation formations in Australia. In:
Keast A, Ed. Ecological biogeography of Australia. The Haag:
Junk. p 163–298.
Thomson J, Mac Nally R, Fleishman E, Horrocks G. 2007. Pre-
dicting bird species distributions in reconstructed landscapes.
Conserv Biol 21:752–66.
Thorburn PJ, Hatton TJ, Walker GR. 1993. Combining mea-
surements of transpiration and stable isotopes of water to
determine groundwater discharge from forests. J Hydrol
150:563–87.
Tockner K, Stanford JA. 2002. Riverine flood plains: present
state and future trends. Environ Conserv 29:308–30.
Mapping Vegetation Condition Across Regions 219
