Content uploaded by Aaron Greenville
Author content
All content in this area was uploaded by Aaron Greenville on Jun 11, 2015
Content may be subject to copyright.
On the validity of visual cover estimates for time series
analyses: a case study of hummock grasslands
Vuong Nguyen .Aaron C. Greenville .Chris R. Dickman .
Glenda M. Wardle
Received: 12 December 2014 / Accepted: 9 May 2015
ÓSpringer Science+Business Media Dordrecht 2015
Abstract Changes in vegetation cover are strongly
linked to important ecological and environmental
drivers such as fire, herbivory, temperature, water
availability and altered land use. Reliable means of
estimating vegetation cover are therefore essential for
detecting and effectively managing ecosystem
changes, and visual estimation methods are often used
to achieve this. However, the repeatability and re-
liability of such monitoring is uncertain due to biases
and errors in the measurements collected by observers.
Here, we use two primary long-term monitoring
datasets on spinifex grasslands, each established with
different motivations and methods of data collection,
to assess the validity of visual estimates in detecting
meaningful trends. The first dataset is characterised by
high spatial and temporal coverage but has limited
detail and resolution, while the second is characterised
by more intensive sampling but at fewer sites and over
a shorter time. Using multivariate auto-regressive
state-space models, we assess consistency between
these datasets to analyse long-term temporal and
spatial trends in spinifex cover whilst accounting for
observation error. The relative sizes of these observa-
tion errors generally outweighed process, or non-
observational errors, which included environmental
stochasticity. Despite this, trends in the spatial dynam-
ics of spinifex cover were consistent between the two
datasets, with population dynamics being driven
primarily by time since last fire rather than spatial
location. Models based on our datasets also showed
clear and consistent population traces. We conclude
that visual cover estimates, in spite of their potential
uncertainty, can be reliable provided that observation
errors are accounted for.
Keywords Monitoring Observation error
Spinifex State-space models Time series Wildfire
Visual cover
Introduction
Monitoring programmes that measure changes in
vegetation cover provide important information needed
to assess the status, trends and dynamics of any study
system, and are therefore crucial for evidence-based
policy management (Eyre et al. 2011; Dickman and
Wardle 2012; Dickman et al. 2014). Vegetation cover
has been demonstrated to have strong relationshipswith
Communicated by Sarah M. Emery.
V. Nguyen (&)A. C. Greenville C. R. Dickman
G. M. Wardle
Desert Ecology Research Group, School of Biological
Sciences, The University of Sydney, Sydney, NSW 2006,
Australia
e-mail: pngu4751@uni.sydney.edu.au
V. Nguyen A. C. Greenville C. R. Dickman
G. M. Wardle
Long-Term Ecological Research Network, Terrestrial
Ecosystem Research Network, Canberra, Australia
123
Plant Ecol
DOI 10.1007/s11258-015-0483-7
global environmental drivers such as atmospheric CO
2
and global temperatures (Braswell et al. 1997;Zeng
et al. 1999), water balance and availability (Joffre and
Rambal 1993), changes in land management practices
(Vicente-Serrano et al. 2004) and fire return intervals
and fire intensity (Eckhardt et al. 2000). Increased
vegetation cover in Australian desert systems, for
example, provides the necessary ground fuel required
for large wildfires to establish and spread (Gill 1975;
Greenville et al. 2009;Nanoetal.2012).
Monitoring of grasslands often relies on visual
inspection of fixed plots to estimate cover (i.e. cover as
a percentage of the total plot area), and is a quick and
non-intensive method to use (e.g. Tischler et al. 2013;
Dickman et al. 2014). However, visual methods are
highly susceptible to observation error, raising con-
cerns about the reliability and repeatability of visual
estimates (Helm and Mead 2004; Wintle et al. 2013).
Observation errors include traditional sampling error
resulting from differences between the sampled
population and the overall population, and measure-
ment error resulting from differences between ob-
served estimates and true values (Staples et al. 2004;
Flesch 2014). These errors are less problematic when
sampling designs are standardised as differences in
sampling across sites or changes in sampling proce-
dures over the duration of the study can result in
variable observation errors over space and time.
Process error on the other hand encompasses all non-
observational error and includes variation resulting
from demographic and environmental stochasticity.
Observation errors in grasslands monitoring might
arise due to the fine texture of grasses and potentially
large areas over which cover is estimated, causing
difficulty in obtaining accurate visual estimates and
creating errors that are large, unknown or observer-
specific (Sykes et al. 1983; Bennett et al. 2000). Cover
estimates in spinifex grasslands (Triodia spp.) may be
potentially more robust, as the hummocks of spinifex
are discrete and constitute the dominant life form over
large areas so that estimates can focus on this species
rather than on overall vegetative cover, although
observation error will still be present. Cheal (2008),
for example, found large discrepancies in estimates of
spinifex cover from 16 experienced observers, ranging
from 20 to 60 % on a 10 % point scale, with no
relationship between general ecological experience of
the observer (rather than task-specific experience) and
cover estimates. In addition to the lack of observer
agreement in visual estimates, factors contributing to
observation errors can include plot size, plant mor-
phology, distribution and incorrect identification
(Kennedy and Addison 1987; Klimes
ˇ2003).
Addressing observer biases would vastly improve
the quality of visual estimates, but may not always be
possible when working within the constraints of
project costs, and requires careful consideration of
the trade-offs involved. Errors may be reduced by
taking group averages rather than relying on a single
expert (Klimes
ˇ2003; Burgman et al. 2011), or by
making replicated observations (Dennis et al. 2010;
Knape et al. 2011), but this can be costly and draw
resources away from improving spatial or temporal
coverage. Thus, in monitoring, a choice will usually
occur between sampling a few sites with high accuracy
and therefore gaining statistical power to detect small-
scale dynamical changes, and sampling with more
extensive temporal and spatial coverage to improve
overall knowledge of populations in the study system
(Morris et al. 1999). Similarly, trade-offs in sampling
can occur when determining appropriate spatial and
temporal coverage. Long-term studies are crucial for
understanding ecological processes and are necessary
to capture rare disturbance events such as fires that
drive many ecosystems (Turner et al. 2003; Linden-
mayer et al. 2012). Monitoring populations from
multiple sites is also important to identify whether
populations are independent, with different population
dynamics, or are correlated and have shared environ-
mental drivers and thus are more susceptible to
extinction (Morris et al. 1999; Warton and Wardle
2003; Ward et al. 2010). Populations that exhibit
unique, independent dynamics may require more
spatial replication to properly manage and understand
the dynamics governing each population, whereas a
correlated population might need more emphasis on
temporal replication since rare, catastrophic events
will impact all populations similarly. A careful
balance is therefore required to ensure that adequate
replication occurs on both scales to detect meaningful
changes in population abundances.
While many studies have investigated the repeata-
bility and reliability of visual cover estimates and
quantified observation errors (Sykes et al. 1983;
Klimes
ˇ2003; Helm and Mead 2004; Gray and Azuma
2005), we instead wish to comment on the usefulness
of such data for discerning real trends and processes.
Our study makes use of two primary long-term
Plant Ecol
123
datasets on the coverage of Triodia basedowii ob-
tained from the Simpson Desert, central Australia,
which were set up with different motivations and
methods of data collection, and ran for different
lengths of time. The first places greater emphasis on
spatial and temporal coverage at the expense of
accuracy and detail (Dickman et al. 2014). The other
begins with highly intensive sampling and investigates
fewer sites over a relatively shorter period, but has
greater resolution at each site (Wardle and Dickman,
unpublished data). Using a multivariate auto-regres-
sive state-space (MARSS) approach to account for
observation error, we compare and assess the utility of
monitoring data for making informed decisions by first
analysing long-term trends in spinifex cover and
whether our two independent datasets produce similar
results. By comparing the changes in cover over time
revealed by each dataset, we assess data reliability,
and whether observed trends are substantially influ-
enced by the sampling strategy.
Using the MARSS models, we also test several
hypotheses regarding spatial relationships in T. base-
dowii populations. Re-establishment of spinifex cover
post-fire will vary depending on the amount of rainfall
(Griffin et al. 1983;Nanoetal.2012; Nano and Pavey
2013); a diverse community of herbs and grasses can
emerge during the recovery period, supplying more
palatable food sources for livestock grazing compared
to persistent stands of long-unburnt spinifex (Dickman
et al. 2014). The post-fire recovery period might
therefore produce distinct spinifex population dynamics
and responses to rainfall depending on the time since
last fire. Additionally, this rainfall can be spatially
restricted, producing isolated pockets of productivity
(Letnic and Dickman 2005,2006), such that population
dynamics depend more on the study location rather than
fire history. The interaction between the post-fire
recovery and rainfall could further distinguish spinifex
population dynamics such that study sites impacted by
the same fire may have different population dynamics
due to spatially variable rainfall regimes. Finally, we
compare the relative sizes of observation errors com-
pared to process errors as estimated by the MARSS
models. By using our two independent datasets, we aim
to determine whether visual cover estimates are capable
of detecting meaningful trends despite concerns re-
garding their reliability, or whether observation error is
sufficiently large to result in misleading and contradic-
tory conclusions.
Methods
Study region
Two long-term datasets (hereafter referred to as the
‘wide view’ dataset, and the ‘high resolution’ dataset;
Table 1) were obtained from sampling across four
pastoral stations in the Simpson Desert, central
Australia: Carlo Station, Tobermorey Station, Cravens
Peak and Ethabuka Reserves (Fig. 1), covering a
combined area of 8000 km
2
. The study region is
mostly composed of dune fields, with remaining areas
consisting of clay pans, rocky outcrops and gibber flats
(Dickman et al. 2014). Vegetation is dominated by
lobed spinifex (T. basedowii) with occasional small
stands of gidgee trees (Acacia georginae), other
Acacia shrubs and mallee eucalypts (Frank et al.
2012; Tischler et al. 2013; Wardle et al. 2015). Long-
term annual rainfall for this part of the desert averages
ca. 250 mm, with decreasing gradients in annual
rainfall occurring from north to south and east to west
that partition the study region into three broad
geographical areas (Greenville et al. 2012).
Wide view dataset
The wide view dataset was initially created to track
small mammal and lizard populations with the inten-
tion of long-term monitoring to capture important
boom-bust dynamics associated with sporadic, heavy
rainfall, and responses to major fire events (Dickman
et al. 1999; Letnic and Dickman 2005; Dickman et al.
2010,2011; Greenville et al. 2012). Spinifex cover and
seeding were estimated visually as additional covari-
ates as they are important components of the habitat
and diet of small mammals (Dickman et al. 2011).
Live-trapping and wide view data sampling were
carried out from 1990 to 2013 at one site and at eight
other sites from 1995 to 2013 (Fig. 1), each site
containing 2–12 grids spaced 0.5–2 km apart in
randomly chosen positions along access tracks
(Table 1). Sampling was conducted at irregular inter-
vals from 2 to 6 times a year and missing values were
included in the time series when sites were not
sampled. As sampling did not begin at most sites (8 of
9) until 1995, we use data from 1995 onwards. We also
constructed models using data only from 2004
onwards (‘truncated wide view’ dataset) to match the
temporal scale of the high-resolution dataset (see
Plant Ecol
123
below), allowing us to compare state predictions and
temporal trends in spinifex cover assuming that data
collection for both datasets began at the same time.
Spinifex cover was measured visually as percentage
cover (5 % point scale) in a 2.5 m radius around six
points on each sampling grid (Table 1). Due to the
irregularity of sampling intervals and the need to
reflect divergent response times, data were aggregated
by year to account for the time needed for spinifex
growth to respond to rainfall whilst simplifying model
construction and subsequent comparisons with the
high-resolution dataset. Fire treatment allocation was
ad hoc and done retrospectively, whereby a site was
labelled as burnt if most sampling grids experienced a
fire during the study period (six of nine sites). In such
cases, grids that were unaffected by fire for that site
were discarded for analyses. Likewise, for sites
labelled as unburnt, grids that did experience a fire
were discarded.
High-resolution dataset
The high-resolution dataset was created to monitor
vegetation abundance and diversity with a particular
focus on the effects of a wildfire that spread through
the Simpson Desert in 2001–2002 (Greenville et al.
2009). Subsequently, fire treatment was incorporated
into a stratified design, whereby each site consisted of
two 1-ha grids with one grid positioned in an area
burnt during the 2001–2002 wildfire, while the other
was placed in an unburnt area. Vegetation surveys
were carried out from 2004 to 2013 at four sites
(Fig. 1). Within each grid, 15 5 95 m plots were
placed in a 3 95 arrangement where the five plots on
each row were placed on the swale, the middle, and the
crest of the linear dunes and spaced randomly with a
minimum separation of 5 m along a 100-m transect,
giving a total of 120 plots across the eight grids
(Table 1). Data were not previously aggregated and
each plot was treated as a replicate sample of its
corresponding grid. Intensive sampling was carried
out in the first three years of data collection
(2004–2006), with surveys being conducted quarterly.
However, from 2007 onwards, surveys were only
conducted 1–2 times a year at unequal time intervals.
As with the wide view dataset, missing values were
included when sites were not sampled. Spinifex cover
was measured in m
2
rather than as a percentage, and
models for both datasets are presented as estimates of
cover area.
Model description
Data were analysed using MARSS models (Holmes
et al. 2012a,b). These models were used to calculate
true state estimates for spinifex cover accounting for
process and observation error, estimate missing val-
ues, and to investigate various spatial hypotheses.
Maximum likelihood parameters and state estimates
for these models are achieved via recursive Kalman-
Filter and Expectation–Maximisation algorithms until
the models reach convergence (Shumway and Stoffer
Table 1 Comparison of the main features of the wide view and high-resolution datasets of spinifex cover obtained from sampling at
multiple sites in the Simpson Desert, central Australia
Design feature Wide view dataset High-resolution dataset
Time series length Long (1995–2013) Medium (2004–2013)
Spatial coverage High Medium
Number of sites 9 4
Blocks/grids 2–12 2
Accuracy/detail Low High
Plots 6 per grid, circular 2.5 m radius 15 per grid, square, 5 95m
Spinifex estimation method Percentage cover, 5 % point scale Area cover, 0.5 m
2
point scale
Number of observers High ([10, with varying experience) Low (four trained personnel)
Data analysis Aggregated by site and year Not aggregated
Fire treatment Site level, ad hoc assignment Grids within sites
Plant Ecol
123
2006). Conceptually, the state-space model partitions
population models into observed (data) and unob-
served (true state) components. Let ndenote the
number of discrete survey sites and mdenote the
number of hypothesised populations. The MARSS
model is then denoted by
Xt¼Xt1þuþCctþwt;wtMVN 0;QðÞð1Þ
Yt¼Zxtþaþvt;vtMVN 0;RðÞ ð2Þ
where Eq. (1) represents the model for the true states,
X
t
is the mtrue states at time t,uis the trend parameter,
Cis the covariate effect, c
t
is the covariate value at
time t, and w
t
is the process error assumed to be from a
multivariate normal (MVN) distribution with mean
zero and variance–covariance matrix Q(Holmes et al.
2007; Hinrichsen and Holmes 2009). Equation (2)
represents the model for the observed states, where Y
t
is the nobserved estimates at time t,Zis a
n9mmatrix of 0’s and 1’s denoting population
structure, pairing up each of the nobservations to one
of the mhypothesised states, and is used to explore
various spatial hypotheses (see Fig. 2for more
details), ais the mean linear difference between
survey sites measuring the same sub-population with
respect to the first site, and v
t
is the observation error
assumed to be MVN with mean zero and variance–
covariance matrix R.
For the wide view dataset, we erected and tested the
following hypotheses: (1) Individual site model (9
states): Each of the nine sampling sites act as
independent, uncorrelated populations (Z is a 9 99
Fig. 1 Location of study sites across Carlo Station, Tober-
morey Station, Cravens Park and Ethabuka Reserve, Simpson
Desert, central Australia. Circles indicate sites that belong to the
wide view dataset, the square indicates a site belonging to the
high-resolution dataset and triangles belong to both. Red fills are
sites that were retrospectively labelled as burnt for the wide view
dataset and indicate sites that experienced a fire over the summer
of 2001–2002 and are designated as burnt sites for the duration
of the time series (1995–2013), while blue indicates unburnt
sites. (Color figure online)
Plant Ecol
123
diagonal matrix, whereby each site is paired up with its
own corresponding state); (2) Geographical region
model (3 states): Sites can be grouped by geographical
area into northern, southern and western populations
partitioned by rainfall (Z is a 9 93 matrix whereby
for each row, a 1 in the first, second or third column
pairs that observation with the northern, southern or
western population, respectively); (3) Wildfire model
(2 states): Sites can be grouped into burnt and unburnt
populations based on their time since last fire, with
particular reference to the 2001–2002 wildfires (Z is a
992 matrix whereby for each row, a 1 in the first or
second column pairs that observation with the burnt or
unburnt population, respectively). The Cc
t
parameters
were used to model the immediate effect of the
2001–2002 wildfires, whereby c
t
was set to 1 for 2002
and 0 for the remaining years; and (4) Complete
homogeneity (1 state): Populations from all nine sites
act as a single population following the same trajec-
tory with shared model parameters (Z is a 9 91
column vector of 1’s). For the high-resolution dataset,
we tested similar hypotheses with slight variations due
to the different designs: (1) Complete heterogeneity (8
states): Each of our grids acts as independent,
uncorrelated populations. This can be likened to an
interaction effect between fire history and site; (2)
Individual site model (4 states): Each of the four sites
acts as an independent population with no significant
effect of fire history on the trajectory or parameters;
(3) Wildfire model (2 states): Populations on burnt
grids follow the same trajectory with shared pa-
rameters regardless of site and similarly for unburnt
grids; and (4) Complete homogeneity: There is no
effect of site or fire history, and all grids in the study
area follow the same trajectory with shared pa-
rameters. For each of the hypotheses, we estimate a
unique trend parameter ufor each hypothesised state.
The variance–covariance matrix Qfor the process
errors was modelled as unconstrained (i.e. non-zero
covariances) for all hypotheses except the ‘‘Individual
site model’’ for the wide view dataset, in which Qwas
modelled as a diagonal matrix due to convergence
issues. These convergence issues resulted in degener-
ate variance estimates and is generally caused by a
Fig. 2 Flowchart
describing the process of
testing spatial hypotheses
using MARSS models
Plant Ecol
123
lack of data relative to the number of parameters being
estimated (Holmes et al. 2012b). A unique observation
error was estimated for each site in the high-resolution
dataset, while only a single error term was estimated
for the wide view dataset due to similar issues. For
both datasets, the variance–covariance matrix Rwas
modelled with zero covariance. The best fitting model
was identified as having the lowest Akaike Informa-
tion Criterion adjusted for small sample sizes (AICc),
whereby DAICc\2 points suggests no substantial
difference between models, and DAICc[8 points is
considered weakly supported (Burnham and Anderson
2002).
Results
Spatial hypotheses and state predictions
Multivariate auto-regressive state-space models for
the wide view and truncated wide view datasets show
the wildfire model to have the greatest support
(DAICc[8 compared to the next lowest AICcval-
ues), with the individual site model having the least
support (Table 2). MARSS models for the high-
resolution dataset show similar results, with the
wildfire model having the greatest support despite a
shorter time series which did not include the
2001–2002 wildfires (Table 2). Thus, there is agree-
ment between both datasets and T. basedowii can be
grouped reliably into burnt and unburnt populations,
with trends in spinifex population dynamics better
explained by time since last fire than by spatial
location.
True state estimates produced by the wildfire model
for the wide view dataset indicate gradually increasing
cover from 1995 onwards across all sites until the
2001–2002 wildfires, creating a distinct division
between the burnt and unburnt populations (Fig. 3).
Following the fire, state estimates for both datasets
show similar temporal trends in spinifex cover in
which burnt populations remained low throughout the
remainder of the monitoring period, with only a slight
hint of a recovery by the end of the monitoring period
(Figs. 3,4). Meanwhile, spinifex cover in unburnt
populations decreased gradually, dipping to a mini-
mum in 2008 as a result of an extended drought period,
followed by a relatively rapid recovery that followed
heavy rains in 2010 and led to an outpacing of growth
compared to that in the burnt populations. While
truncating the wide view dataset to match the high-
resolution dataset in terms of time series length still
retains the population structure results (Table 2), this
model was unable to disentangle observation error
from the process error, resulting in poor state estimates
for the burnt population (Fig. 5).
Comparison of maximum likelihood parameter
estimates
In both the wide view and high-resolution datasets, the
observation error made up a larger component of the
error term compared to process error, Q(Table 3).
Observation errors were also found to be spatially
variable, differing across and within sites. For exam-
ple, the high-resolution dataset shows that the Main
Camp site has substantially higher observation error
terms compared to the other sites. In addition, both
datasets suggest that observation error in unburnt sites
is larger relative to that in burnt sites. In contrast, while
both datasets suggest positive spinifex growth (l)in
burnt and unburnt populations, the relative growth
Table 2 Comparison of model performance investigating sub-
population structure of spinifex at sites in the Simpson Desert,
central Australia, using wide view, truncated wide view (from
2004 onwards) and high-resolution datasets of spinifex cover
(see Methods for description of hypotheses), with the best fit-
ting model given by the lowest AICc
States AICc
Wide view dataset
Wildfire 2 438.4703
Geographical region 3 487.1419
Complete homogeneity 1 487.8002
Independent site 9 539.2734
a
Truncated wide view dataset
Wildfire 2 197.2785
Complete homogeneity 1 206.0595
Geographical region 3 223.1979
Independent site 9 245.1594
a
High-resolution dataset
Wildfire 2 6672.06
Complete homogeneity 1 6688.90
Independent site 4 6923.79
Complete heterogeneity 8 7171.68
a
Qwith covariance set to 0
Plant Ecol
123
rates differ between burnt and unburnt populations.
Spinifex growth in burnt sites was estimated to be
substantially higher compared to unburnt sites for the
wide view dataset, while the reverse was true for the
high-resolution dataset. However, confidence inter-
vals in both cases overlap zero suggesting that this
growth is non-significant. In general, confidence
intervals for the wide view dataset are much larger
when compared to the high-resolution dataset, despite
the longer time series.
Discussion
Ecological monitoring programmes are critical for
understanding ecological processes, provision of long-
term data, supporting evidence-based decision mak-
ing, and adaptive conservation management (Speller-
berg 2005; Wintle et al. 2010; Eyre et al. 2011;
Lindenmayer et al. 2012,2014). Visual cover esti-
mates are cheap, rapid and are prevalent in monitoring
programmes of vegetation cover. That these cover
estimates in forests (Helm and Mead 2004) have
previously been shown to be non-reproducible and
highly susceptible to error is therefore a serious issue,
and has implications for management decisions based
upon such inputs. More positively, the use of feedback
and calibration of observer results does improve
reliability of visual estimation (Wintle et al. 2013).
Despite these widely acknowledged concerns, our
study demonstrates through the use of two empirical
datasets that monitoring data can still uncover
1995 2000 2005 2010
2468
Main Camp
1995 2000 2005 2010
23456789
Carlo
1995 2000 2005 2010
02468
Field River South
1995 2000 2005 2010
12345678
Kunnamuka Swamp East
Spinifex cover (m2
per plot)
1995 2000 2005 2010
02468
Shitty Site
1995 2000 2005 2010
1234567
South Site
1995 2000 2005 2010
2468
Field River North
1995 2000 2005 2010
34567
Tobermorey East
Years
1995 2000 2005 2010
34567
Tobermorey West
Fig. 3 State predictions (solid line) and their 95 % confidence
intervals (dashed lines) for cover of spinifex from the best fitting
model (wildfire) for the wide view dataset obtained from nine
sites in the Simpson Desert, central Australia. Filled points
indicate actual observations. Red represents state predictions
and observations from burnt sites, while blue represents those
from unburnt populations. (Color figure online)
Plant Ecol
123
conclusive trends. Although the wide view and high-
resolution datasets began with different motivations,
and subsequently employed varying study designs and
collection methods, monitoring data from these
datasets were mostly consistent in terms of population
structure and temporal trends. Models constructed
from both datasets, for example, suggested that
population dynamics for spinifex can be grouped by
time since last fire into distinct burnt and unburnt
populations, and we found generally little support for
models suggesting site by site differences in the
response to fire. Thus, the impact of the major
2001–2002 wildfire, and the subsequent temporal
dynamics driven by rainfall events and periods of
drought, were so extreme as to override any potential
spatial heterogeneity that might have existed, for
2004 2006 2008 2010 2012
02468
2004 2006 2008 2010 2012
02468
Main Camp
2004 2006 2008 2010 2012
02468
2004 2006 2008 2010 2012
02468
Field River South
2004 2006 2008 2010 2012
02468
Spinifex cover ( m2
per plot)
2004 2006 2008 2010 2012
02468
Carlo Shitty
2004 2006 2008 2010 2012
02468
2004 2006 2008 2010 2012
02468
South Site
Years
Fig. 4 State predictions (solid lines) with their 95 % confi-
dence intervals (dashed lines) for cover of spinifex from the best
fitting model (wildfire) for the high-resolution dataset obtained
from four sites in the Simpson Desert, central Australia. Small
circles indicate observations from each of 15 replicate plots,
while large filled points indicate the means of these plots. Each
of the four sites contains state predictions and observations from
burnt (red) and unburnt (blue) populations. (Color figure online)
Plant Ecol
123
example, as a result of spatially variable rainfall
regimes or pre-fire vegetation composition (Wardle
et al. 2013). Trends such as the relatively slow post-
fire recovery in the burnt populations following the
2001–2002 wildfires and the impact of the drought on
the unburnt populations are also consistent and
reasonably identifiable between the two datasets. A
comparison of the relative growth rates between burnt
and unburnt sites showed conflicting results; however,
this was likely because estimation for the high-
resolution dataset only captures the period following
the 2001–2002 wildfires in which the population
experiences comparatively slow recovery, resulting in
lower growth estimate compared to the wide view
dataset. Thus, we conclude that the monitoring data in
both cases had enough power to detect these large
environmental drivers despite the substantial obser-
vation errors.
Concerns regarding observer bias and errors when
making visual cover estimates are not without good
reason as errors can be potentially quite large (Helm
and Mead 2004; Cheal 2008; Vittoz et al. 2010). These
errors can be so large as to exceed actual year to year
variation in cover (Klimes
ˇ2003), and our results
suggest this is the case with estimates of spinifex
cover, albeit over the period of least change in cover
post-fire rather than including the fire itself. While one
might expect that such substantial errors would
provide unreliable and misleading results (Kennedy
and Addison 1987), these errors here did not appear to
hinder our ability to discern consistent trends from the
two datasets. Having replicate observations in the
high-resolution dataset vastly improved the model’s
capacity to disentangle both process and observation
error, and additionally allowed estimates of observa-
tion error to be calculated for each sampling location.
We note that this was not possible when the high-
resolution data were aggregated, as the models
encountered convergence issues in the algorithms, a
common problem when data are insufficient (Dennis
et al. 2006; Holmes et al. 2012a). Similar issues
occurred with the wide view dataset despite the longer
time series, particularly as there were more sites for
which observation error needed to be estimated and
many missing values for some sites. Consequently,
only a single observation error term was calculated for
the nine sites. However, our high-resolution dataset
suggests it may not be appropriate to assume that all
sites have equal observation errors as they can differ
substantially both between and within sites. For
example, we found observation errors to be larger in
unburnt sites, likely due to greater variation in cover
between plots, higher cover overall and higher process
error resulting in bigger changes over time. Combined
with the higher frequency of large, mature hummocks
of spinifex which vary greatly in shape and distribu-
tion in unburnt sites, and increased likelihood of
overhanging the plot, these factors all add to the
difficulty of providing reliable estimates of cover.
In contrast, spinifex cover at burnt sites remained
low throughout the study and consisted mostly of
small, re-establishing seedlings post-fire (Rice and
Westoby 1999), resulting in lower estimates for
observation error at burnt sites. Kennedy and Addison
2004 2006 2008 2010 2012
02468
Wide View Dataset
2004 2006 2008 2010 2012
01234567
High Resolution Dataset
Spinifex cover ( m2
per plot)
Years
Fig. 5 Comparison of state predictions (solid lines) averaged
across sites and 95 % confidence intervals (dashed lines) for
cover of spinifex derived from wildfire models for the wide view
dataset and the high-resolution dataset obtained from multiple
sites in the Simpson Desert, central Australia. The wildfire model
for the wide view dataset was constructed using only data from
2004 onwards to allow comparison with the high-resolution
dataset, thus assuming that data collection for both datasets
began at the same time. Observations (points) for the high-
resolution dataset are presented as site means. Red represents
mean state predictions and observations for the burnt, while blue
represents those from unburnt populations. (Color figure online)
Plant Ecol
123
(1987) compared observation errors between species
and found larger errors for those with low cover, but
this result was driven by detection errors where small,
infrequently distributed species were seen in one
sampling unit, but not the next. Conversely, Sykes
et al. (1983) suggested that observation errors are less
likely to occur at the lower and higher extremities and
most likely to occur in between, and our results for
spinifex cover seem to follow this pattern. Variation in
the distribution and amount of cover between sites can
therefore lead to some sites being more difficult to
estimate cover than others. Thus, observation error can
be better accounted for by focusing on fewer sites with
greater replication than with the approach taken in the
wide view dataset that emphasised greater spatial and
temporal coverage (Dennis et al. 2010). In addition,
the potential size of observation errors might some-
times be anticipated; for example, if a site is known to
have consistently lower cover or species diversity,
monitoring resources could be distributed more effi-
ciently. Higher diversity sites can complicate visual
estimates as there is more heterogeneity in the possible
sources of error, and cover estimates for individual
plant species are more susceptible to errors than
overall estimates of vegetation cover (Klimes
ˇ2003).
In contrast, study systems dominated by a single
species, as is the case in our study, would be less
vulnerable to complications associated with estimat-
ing cover for multiple, low abundance species, and
estimating cover for only the dominant life form can
be quite informative of the study system. Thus, visual
estimates may be more appropriate for grassland study
systems compared to other environments such as
woodlands where species diversity may be higher.
Investigating the extent of spatial heterogeneity and
determining whether populations exhibit unique
population trajectories can be particularly informative
(Ward et al. 2010). Since spinifex population dynam-
ics were shared across sites and driven predominantly
by time since last fire, the increased spatial coverage
observed in the wide view dataset may not have been
necessary if one is purely interested in monitoring
spinifex cover. However, the longer time series
provided this monitoring programme with an ad hoc
opportunity to observe the impacts of a major wildfire.
Wildfires are known to be a major driver of Australian
desert ecosystems (Gill 1975; Griffin et al. 1983), and
though fire cannot occur without consecutive seasons
of high rainfall that promote plant growth and hence
ground fuel, such events remain difficult to anticipate,
with fire return intervals ranging from 20 to 50 years
(Greenville et al. 2009; Nano et al. 2012). While it
cannot be known exactly when wildfire events might
occur, identifying potentially important drivers a
priori is the key to a well-designed monitoring
programme (Wintle et al. 2010). In contrast, while
the targeted monitoring approach of the high-resolu-
tion dataset is expected to have more power to detect
Table 3 Maximum likelihood parameter estimates from the best fitting models using wide view and high-resolution datasets on
spinifex cover sampled at multiple sites in the Simpson Desert, central Australia
Parameter Wide view High resolution
Burnt Unburnt Burnt Unburnt
Growth rate (l) 0.179 (-0.067, 0.414) 0.056 (-0.486, 0.458) 0.059 (-0.001, 0.119) 0.086 (-0.102, 0.273)
Effect of fire (C)-5.645 (-0.676, -4.022) – –
Process error (Q) 0.119 (0.000, 0.598) 0.907 (0.001, 2.064) 0.032 (0.004, 0.059) 0.265 (0.035, 0.495)
Observation error (R)
All 1.467 (0.908, 1.671) – –
Main Camp – – 2.635 (2.187, 3.083) 2.795 (2.285, 3.304)
Field River – – 0.570 (0.469, 0.670) 0.613 (0.499, 0.727)
Carlo Shitty – – 0.865 (0.717, 1.013) 1.460 (1.204, 1.716)
South Site – – 0.328 (0.269, 0.386) 1.262 (1.040, 1.484)
The immediate reduction in spinifex cover due to the wildfire in 2001–2002 is captured by the Cparameter and is estimated only for
the wide view dataset as it covered the time period in which it occurred. Observation error (R) was estimated for each grid in the high-
resolution dataset, while only a single, shared observation error term was estimated for the wide view dataset due to convergence
issues. Bootstrapped 95 % confidence intervals are given in parentheses
Plant Ecol
123
meaningful trends relating to time since last fire, the
actual event itself may not always be captured, and is a
testament to the value of long-term monitoring
(Franklin 1989; Dickman and Wardle 2012; Linden-
mayer et al. 2012). Interestingly, even when the wide
view dataset was truncated to be of similar length to
the high-resolution dataset but without the additional
sampling intensity, the fire signal was still strong
enough to be detected despite not being incorporated
into the study design. However, this shortened time
series had insufficient data from each site to properly
disentangle process and observation errors, and re-
sulted in poor state estimates for the burnt population.
If funding uncertainties result in researchers being
unable to commit to long-term monitoring pro-
grammes, it may therefore be prudent to ensure that
short-time series are of high quality, either through
achieving sufficient replication over short periods to
estimate observation errors, or through use of solu-
tions that reduce impact such as taking mean values
(Klimes
ˇ2003; Wintle et al. 2013).
Visual cover estimates are by no means perfect, but
are rapid, cheap, non-destructive and commonly
available, and therefore realistically are more likely
to be carried out over long time periods and over large
spatial scales than intensive repeat sampling. The use
of remote sensing technologies and computer-image
analysis provide an attractive alternative for reducing
observation error (Bennett et al. 2000; Booth and
Tueller 2003) but will introduce observation errors of
other kinds, including the choice of post-process
algorithms (Kennedy et al. 2014). Furthermore, this
does not reduce the need for field-based approaches to
assist with image interpretation and obtain other
important measurements to provide a complete picture
of the study system (Lathrop Jr et al. 2014). While it is
important to acknowledge the presence of errors and
biases, visual estimates can generally be relied upon to
detect meaningful changes (Vanha-Majamaa et al.
2000; Irvine and Rodhouse 2010), and our results
provide some vindication for their continued use in
management and monitoring studies. However, this
conclusion depends on observer error being accounted
for. There are several options for reducing observation
error such as taking group averages rather than relying
on a single expert (Klimes
ˇ2003; Burgman et al. 2011),
bias correction factors (Sykes et al. 1983) or active
calibration feedback to evaluate performance (Wintle
et al. 2013), and how one wishes to address this issue
should be considered before deciding to implement
sampling based on visual cover estimates.
Acknowledgments We thank land managers in the Simpson
Desert for access to their properties, especially Bush Heritage
Australia, Bobby Tamayo and David Nelson for much logistical
assistance with all aspects of the field programmes and data
collection, Chin-Liang Beh and other members of the Desert
Ecology Research Group for additional assistance and many
volunteers for help in the field and laboratory. We also thank the
Australian Research Council and the Long-Term Ecological
Research Network for provision of funding.
References
Bennett LT, Judd TS, Adams MA (2000) Close-range vertical
photography for measuring cover changes in perennial
grasslands. J Range Manag 53:634–641
Booth DT, Tueller PT (2003) Rangeland monitoring using re-
mote sensing. Arid Land Res Manag 17:455–467
Braswell BH, Schimel DS, Linder E, Moore B (1997) The re-
sponse of global terrestrial ecosystems to interannual
temperature variability. Science 278:870–873
Burgman MA, McBride M, Ashton R, Speirs-Bridge A, Flander
L, Wintle B, Fidler F, Rumpff L, Twardy C (2011) Expert
status and performance. PLoS ONE 6:e22998
Burnham KP, Anderson DR (2002) Model selection and mul-
timodel inference. Springer Science and Business Media
Inc, New York
Cheal D (2008) Repeatability of cover estimates? Ecol Manag
Restor 9:67–68
Dennis B, Ponciano JM, Lele SR, Taper ML, Staples DF (2006)
Estimating density dependence, process noise, and obser-
vation error. Ecol Monogr 76:323–341
Dennis B, Ponciano JM, Taper ML (2010) Replicated sampling
increases efficiency in monitoring biological populations.
Ecology 91:610–620
Dickman CR, Wardle GM (2012) Monitoring for improved
biodiversity conservation in arid Australia. In: Linden-
mayer D, Gibbons P (eds) Biodiversity monitoring in
Australia. CSIRO Publishing, Collingwood
Dickman CR, Letnic M, Mahon PS (1999) Population dynamics
of two species of dragon lizards in arid Australia: the ef-
fects of rainfall. Oecologia 119:357–366
Dickman CR, Greenville AC, Beh C-L, Tamayo B, Wardle GM
(2010) Social organization and movements of desert ro-
dents during population ‘‘booms’’ and ‘‘busts’’ in central
Australia. J Mammal 91:798–810
Dickman CR, Greenville AC, Tamayo B, Wardle GM (2011)
Spatial dynamics of small mammals in central Australian
desert habitats: the role of drought refugia. J Mammal
92:1193–1209
Dickman CR, Wardle GM, Foulkes J, de Preu N (2014) Desert
complex environments. In: Lindenmayer D, Burns E,
Thurgate N, Lowe A (eds) Biodiversity and environmental
change: monitoring, challenges and direction. CSIRO
Publishing, Collingwood
Plant Ecol
123
Eckhardt H, Wilgen B, Biggs H (2000) Trends in woody
vegetation cover in the Kruger National Park, South Africa,
between 1940 and 1998. Afr J Ecol 38:108–115
Eyre TJ, Fisher A, Hunt LP, Kutt AS (2011) Measure it to better
manage it: a biodiversity monitoring framework for the
Australian rangelands. Rangel J 33:239–253
Flesch AD (2014) Spatiotemporal trends and drivers of
population dynamics in a declining Sonoran Desert
predator. Biol Conserv 175:110–118
Frank ASK, Dickman CR, Wardle GM (2012) Habitat use and
behaviour of cattle in a heterogeneous desert environment
in central Australia. The Rangeland Journal 34:319–328
Franklin JF (1989) Importance and justification of long-term
studies in ecology. In: Likens GE (ed) Long-term studies in
ecology. Springer, New York
Gill AM (1975) Fire and the Australian flora: a review. Aust For
38:4–25
Gray AN, Azuma DL (2005) Repeatability and implementation
of a forest vegetation indicator. Ecol Indic 5:57–71
Greenville AC, Dickman CR, Wardle GM, Letnic M (2009) The
fire history of an arid grassland: the influence of antecedent
rainfall and ENSO. Int J Wildland Fire 18:631–639
Greenville AC, Wardle GM, Dickman CR (2012) Extreme cli-
matic events drive mammal irruptions: regression analysis
of 100-year trends in desert rainfall and temperature. Ecol
Evol 2:2645–2658
Griffin G, Price N, Portlock H (1983) Wildfires in the central
Australian rangelands. J Environ Manag 17:311–323
Helm DJ, Mead BR (2004) Reproducibility of vegetation cover
estimates in South-Central Alaska Forests. J Veg Sci
15:33–40
Hinrichsen R, Holmes EE (2009) Using multivariate state-space
models to study spatial structure and dynamics. In: Cantrell
S, Cosner C, Ruan S (eds) Spatial ecology. CRC/Chapman
Hall, Boca Raton
Holmes EE, Sabo JL, Viscido SV, Fagan WF (2007) A statistical
approach to quasi-extinction forecasting. Ecol Lett
10:1182–1198
Holmes EE, Ward EJ, Scheuerell MD (2012a) Analysis of
multivariate time-series using the MARSS package. Ver-
sion 3.4. http://cran.r-project.org/web/packages/MARSS
Holmes EE, Ward EJ, Wills K (2012b) MARSS: multivariate
autoregressive state-space models for analyzing time series
data. R J 4:11–19
Irvine KM, Rodhouse TJ (2010) Power analysis for trend in
ordinal cover classes: implications for long-term vegeta-
tion monitoring. J Veg Sci 21:1152–1161
Joffre R, Rambal S (1993) How tree cover influences the water
balance of Mediterranean rangelands. Ecology 74:570–582
Kennedy K, Addison P (1987) Some considerations for the use
of visual estimates of plant cover in biomonitoring. J Ecol,
151–157
Kennedy RE, Andre
´foue
¨t S, Cohen WB, Go
´mez C, Griffiths P,
Hais M, Healey SP, Helmer EH, Hostert P, Lyons MB
(2014) Bringing an ecological view of change to Landsat-
based remote sensing. Front Ecol Environ 12:339–346
Klimes
ˇL (2003) Scale-dependent variation in visual estimates
of grassland plant cover. J Veg Sci 14:815–821
Knape J, Jonze
´n N, Sko
¨ld M (2011) On observation distributions
for state space models of population survey data. J Anim
Ecol 80:1269–1277
Lathrop RG Jr., Haag SM, Merchant D, Kennish MJ, Fertig B
(2014) Comparison of remotely-sensed surveys vs. in situ
plot-based assessments of sea grass condition in Barnegat
Bay-Little Egg Harbor, New Jersey USA. J Coast Conserv
1–10
Letnic M, Dickman CR (2005) The responses of small mammals
to patches regenerating after fire and rainfall in the Simp-
son Desert, central Australia. Austral Ecol 30:24–39
Letnic M, Dickman CR (2006) Boom means bust: interactions
between the El Nin
˜o/Southern Oscillation (ENSO), rainfall
and the processes threatening mammal species in arid
Australia. Biodivers Conserv 15:3847–3880
Lindenmayer DB, Likens GE, Andersen A, Bowman D, Bull
CM, Burns E, Dickman CR, Hoffmann AA, Keith DA,
Liddell MJ, Lowe AJ, Metcalfe DJ, Phinn SR, Russell-
Smith J, Thurgate N, Wardle GM (2012) Value of long-
term ecological studies. Austral Ecol 37:745–757
Lindenmayer D, Burns E, Thurgate N, Lowe AJ (2014) The
value of long-term research and how to design effective
ecological research and monitoring. In: Lindenmayer D,
Burns E, Thurgate N, Lowe AJ (eds) Biodiversity and
environmental change. CSIRO Publishing, Collingwood
Morris W, Doak DF, Groom M, Kareiva P, Fieberg J, Gerber L,
Murphy P, Thomson D (1999) A practical handbook for
population viability analysis. Nat Conserv
Nano CE, Pavey CR (2013) Refining the ‘pulse-reserve’model
for arid central Australia: seasonal rainfall, soil moisture
and plant productivity in sand ridge and stony plain habitats
of the Simpson Desert. Austral Ecol 38:741–753
Nano CEM, Clarke PJ, Pavey CR (2012) Fire regimes in arid
hummock grasslands and Acacia shrublands. In: Bradstock
RA, Williams JE, Gill AM (eds) Flammable Australia: fire
regimes, biodiversity and ecosystems in a changing world.
CSIRO Publishing, Collingwood
Rice B, Westoby M (1999) Regeneration after fire in Triodia R.
Br Aust J Ecol 24:563–572
Shumway RH, Stoffer DS (2006) Time series analysis and its
applications. Springer, New York
Spellerberg IF (2005) Monitoring ecological change. Cam-
bridge University Press, Cambridge
Staples DF, Taper ML, Dennis B (2004) Estimating population
trend and process variation for PVA in the presence of
sampling error. Ecology 85:923–929
Sykes JM, Horrill AD, Mountford MD (1983) Use of visual
cover assessments as quantitative estimators of some Bri-
tish woodland taxa. J Ecol 71:437–450
Tischler M, Dickman CR, Wardle GM (2013) Avian functional
group responses to rainfall across four vegetation types in
the Simpson Desert, central Australia. Austral Ecol
38:809–819
Turner MG, Collins SL, Lugo AL, Magnuson JJ, Rupp TS,
Swanson FJ (2003) Disturbance dynamics and ecological
response: the contribution of long-term ecological re-
search. Bioscience 53:46–56
Vanha-Majamaa I, Salemaa M, Tuominen S, Mikkola K (2000)
Digitized photographs in vegetation analysis-a comparison
of cover estimates. Appl Veg Sci 3:89–94
Vicente-Serrano SM, Lasanta T, Romo A (2004) Analysis of
spatial and temporal evolution of vegetation cover in the
Spanish central Pyrenees: role of human management.
Environ Manag 34:802–818
Plant Ecol
123
Vittoz P, Bayfield N, Brooker R, Elston DA, Duff EI, Theurillat
JP, Guisan A (2010) Reproducibility of species lists, visual
cover estimates and frequency methods for recording high-
mountain vegetation. J Veg Sci 21:1035–1047
Ward EJ, Chirakkahl H, Gonza
´lez-Sua
´rez M, Aurioles-Gamboa
D, Holmes EE, Gerber L (2010) Inferring spatial structure
from time-series data: using multivariate state-space
models to detect metapopulation structure of California sea
lions in the Gulf of California, Mexico. J Appl Ecol
47:47–56
Wardle GM, Pavey CR, Dickman CR (2013) Greening of arid
Australia: new insights from extreme years. Austral Ecol
38:731–740
Wardle GM, Greenville AC, Frank ASK, Tischler MK, Emery
NJ, Dickman CR (2015) Ecosystem risk assessment of
Georgina gidgee woodlands in central Australia (Qld, NT,
SA). Austral Ecol (in press)
Warton DI, Wardle GM (2003) Site-to-site variation in the de-
mography of a fire-affected perennial, Acacia suaveolens,
at Ku-ring-gai Chase National Park, New South Wales,
Australia. Austral Ecol 28:38–47
Wintle BA, Runge MC, Bekessy SA (2010) Allocating
monitoring effort in the face of unknown unknowns. Ecol
Lett 13:1325–1337
Wintle BC, Fidler F, Vesk PA, Moore JL (2013) Improving
visual estimation through active feedback. Methods Ecol
Evol 4:53–62
Zeng N, Neelin JD, Lau K-M, Tucker CJ (1999) Enhancement
of interdecadal climate variability in the Sahel by vegeta-
tion interaction. Science 286:1537–1540
Plant Ecol
123