Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR

Page created by Annette King
 
CONTINUE READING
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Centre for Geo-Information, WUR, Wageningen
      Thesis Report GIRS-2000-28-MB

Productivity-biodiversity patterns –
 a study using multitemporal Landsat
 TM NDVI data for the Alice Springs
       region, central Australia.
                  Maaike Bader
                 November 2000

                                      WAGENINGEN UR
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia: Contents

Contents

FOREWORD............................................................................................................. 4

ABSTRACT .............................................................................................................. 6

1      INTRODUCTION................................................................................................ 8

1.1       Biodiversity - productivity .......................................................................................... 8

1.2       Biodiversity – remote sensing ..................................................................................... 8

1.3       Productivity - remote sensing ..................................................................................... 9

1.4       Expected patterns ...................................................................................................... 10

2      METHODS ....................................................................................................... 12

2.1       Study area .................................................................................................................. 12

2.2     Image processing ....................................................................................................... 13
   2.2.1    Data used ............................................................................................................. 13
   2.2.2    Effect of atmospheric correction. ........................................................................ 13
   2.2.3    Calculation of NDVI ........................................................................................... 15
   2.2.4    Spatial variation................................................................................................... 17
   2.2.5    Resampling to lower spatial resolution ............................................................... 17

2.3       Fieldwork ................................................................................................................... 18

2.4       Statistical analysis...................................................................................................... 19

3      EXAMPLES OF VEGETATION TYPES AND THEIR NDVI SIGNAL ............... 22

3.1       Alice Springs .............................................................................................................. 22

3.2       Acacia scrubland ....................................................................................................... 24

3.3       Sand dunes ................................................................................................................. 25

3.4       Mountain ranges........................................................................................................ 26

3.5       Open grassland .......................................................................................................... 27

3.6       Spinifex ....................................................................................................................... 28

4      RESULTS ........................................................................................................ 29

4.1       Species......................................................................................................................... 29

4.2       Class distinction in the field...................................................................................... 29

4.3       NDVI values ............................................................................................................... 29
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia: Contents

4.4       Groundcover .............................................................................................................. 29

4.5       Comparing NDVI classes .......................................................................................... 30

4.6       Correlations species number – NDVI measures per plot and per transect.......... 32

4.7       Correlations per landsystem..................................................................................... 33

4.8       Spatial heterogeneity ................................................................................................. 34

4.9       Vegetation types......................................................................................................... 35

5      DISCUSSION................................................................................................... 38

5.1       Vegetation cover ........................................................................................................ 38

5.2       Relationship productivity- species richness ............................................................ 38

5.3       Species......................................................................................................................... 40

5.4       Habitats for other taxa.............................................................................................. 41

5.5     Data characteristics ................................................................................................... 41
   5.5.1    Influences on NDVI ............................................................................................ 41
   5.5.2    Spatial accuracy................................................................................................... 42
   5.5.3    Behaviour of variation......................................................................................... 43
   5.5.4    Measuring spatial variation ................................................................................. 43
   5.5.5    Scale .................................................................................................................... 44

6      CONCLUSION & RECOMMENDATIONS........................................................ 45

7      REFERENCES................................................................................................. 46

APPENDIX 1........................................................................................................... 51

APPENDIX 2........................................................................................................... 55

APPENDIX 3........................................................................................................... 57
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia: Foreword

1   Foreword
This report is the result of a 4-month thesis research at the Commonwealth Scientific
and Industrial Research Organization (CSIRO) Centre for Arid Zone Research
(CAZR) in Alice Springs, Australia. The thesis is part of my MSc courses in Geo-
Information Science and Biology at the University of Wageningen, The Netherlands.
The project ran from halfway July to halfway November 2000, and was supervised at
CAZR by Graham Griffin, senior scientist, and supported from Wageningen by
Steven de Jong, professor of Remote Sensing.

The supposedly arid desert of central Australia has been amazingly green this year.
And now, halfway into November, there is no sign of the heat that I have been
predicted for this time of the year. Instead there are cloudy skies, impressive
thunderstorms and running rivers.
I could rave on endlessly about the great mountain ranges, beautiful flowers and
birds, cute wallabies, and wonderful weather of the Alice Springs region, but this will
do.

I would like to thank my housemates, Lara, Shrike and Mel, for putting up with me.
Many thanks to all staff at CAZR for support, advice and bush walks, and in particular
Graham Griffin, for having confidence in what I was doing and helping out where
necessary.

I dedicate this report to my friend Klaartje, who died in a car accident while I was
away, and shouldn’t have.

Maaike Bader, Alice Springs 15-11-2000
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia

                                                            5
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia: Abstract

Abstract
It is generally agreed that there is a relationship between productivity and
biodiversity, and productivity can be measured by remote sensing, so it should be
possible to use remote sensing for monitoring biodiversity. This idea was elaborated
in a case study in arid central Australia, using a time-series of 9 Landsat Thematic
Mapper (TM) images, and field data on species diversity of perennial plants. The
Normalized Difference Vegetation Index (NDVI) was used as a measure for green
vegetation cover, and the mean NDVI over the 9 images and temporal and spatial
standard deviations (sd) were considered to reflect relevant aspects of productivity.
Field samples consisted of species presence/absence data and vegetation cover
estimates from transects of five 20-m radius circular plots.
Field-data and remote sensing data were compared in three sets of analysis: 1. The
field-estimated vegetation cover was compared to the NDVI measures; 2. The NDVI
data was divided into classes, and the species richness compared between classes;
and 3. The species numbers per plot, transect, vegetation type and landsystem were
compared to the NDVI data for those units.
Patterns of NDVI-classes could be well recognised in the vegetation in the field.
Estimated total cover and mean NDVI correlated best, while sd NDVI was most
positively correlated with cover of grasses and herbs. The classes of mean NDVI
differed significantly in species richness, while sd classes did not. The pattern of
NDVI vs. species numbers differs between landscape types. When using data per
plots or transect all correlations are very weak. Vegetation types show a clearer
pattern relating species numbers to mean NDVI as well as temporal and spatial sd.
These relationships are roughly hump-shaped. The measure most important for
species richness in landsystems is the spatial variation. The mean NDVI in a
landsystem did not relate to species numbers.
Mean NDVI and the temporal and spatial variation, as well as other environmental
factors, interact to influence species numbers. The relationships between productivity
and biodiversity and remote sensing data can be rather complex. Several factors can
be responsible for disguising potential patterns, e.g. the measure for biodiversity, the
spatial accuracy of the data, non-vegetation influences on the NDVI, the scale of
variation, and the measures for productivity and persistence of productivity. Some
suggestions are made for possible improvement.

                                                                                       6
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia

                                                            7
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia: Introduction

2     Introduction
This study is based on two well-established ideas: 1. There is a relationship between
productivity and temporal and spatial variability of productivity, and biodiversity (e.g.
Pianka 1966, Ricklefs & Schluter 1993, Rosenzweig 1995); and 2. Productivity and
temporal and its spatial variability are detectable by remote sensing (e.g. Huete
1988, Richardson & Everitt 1992, Clevers 1997).
Adding up 1 and 2 logically leads to the conclusion: 3. Remote sensing can help to
describe patterns of biodiversity. If this conclusion holds in practice, it would offer a
great tool for monitoring biodiversity in remote areas, such as the arid inland of
Australia. Whether productivity is actually the direct cause of the differences in
species richness, or a co-variant, can be argued, but in either case, knowing how the
two relate can be useful information for practical applications.

2.1    Biodiversity - productivity
Most ecologists agree that there is some sort of relationship between species
richness and primary productivity (e.g. Pianka 1966, Tilman 1982, Begon et al. 1990,
Rosenzweig & Abramski 1993, Tilman & Pacala 1993, Wright et al. 1993, Scheiner &
Rey-Benayas 1994, Rosenzweig 1995, Marrs et al. 1996, Jørgensen & Nøhr 1996,
Ritchie & Olff 1999, Waide et al. 1999). The shape and nature of this relationship
have not been agreed upon, however, and the ‘rules’, if there are any, may indeed
have different effects in different ecosystems, for different groups of organisms and
on different scales (Wright et al. 1993, Waide et al. 1999, Lawton 1999). The
prevailing opinion is that the general shape is a unimodal one, and other shapes of
relationships are simply a part of this universal shape, e.g. the increasing or
decreasing phase, depending on the range of productivity sampled (Rosenzweig &
Abramsky 1993, Begon et al. 1990). This is an interesting but safe theoretical
assumption. The position of the optimum productivity for species richness is different
for most types of species, ecosystems, observation scales and biogeographical
regions. Therefore most relationships found can be explained using this model, which
makes it a solid, though possibly not always realistic model. Different theories
explaining the unimodal pattern and an evaluation are summarized in Rosenzweig &
Abramski (1993).
Apart from the total or average productivity, the spatial and temporal variation in
productivity have also been linked to biodiversity, both in theoretical models (e.g.
Tilman & Pacala 1993, Wright et al. 1993, Ritchie & Olff 2000) and based on
observations (e.g. Scheiner & Rey-Benayas 1994, Gough et al. 1994, Pollock,
Naiman & Hanley 1998). In most models, an increase in both spatial and temporal
heterogeneity increases the possibilities for resource partitioning, thereby increasing
the possibilities for species to co-exist, and hence biodiversity.

2.2    Biodiversity – remote sensing
Several studies have used one or several of these factors to relate remote sensing
data to ground data on species diversity (see also appendix 2). Jørgensen & Nøhr
(1996) studied bird species diversity in the Sahel in relation to landscape diversity
and biomass production, both derived from RS images. Fjelså et al. (1997) related
ecoclimatic stability of African ecosystems, assessed by means of a 10-year RS
time-series of a vegetation index, to the occurrence of biodiversity ‘hotspots: areas
with high concentration of relict species and neo-endemics. Coops et al. (1998) used
high-resolution aerial video data to predict habitat quality from spatial variation in the
reflectance of Australian Eucalypt forests.
In all of the studies above, the spectral domain is used directly through knowledge of
the reflection characteristics of objects, e.g. the vegetation cover or canopy shadows.
In many cases information is also derived from temporal and spatial patterns. In

                                                                                         8
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia: Introduction

some cases spatial patterns are in fact the only information used directly. Mack et al.
(1997) used existing land-cover maps derived from Landsat TM images, and
compared bird species numbers in woodland patches of different sizes. They found
that the land cover map underestimated the size of the patches, and hence did not
predict the number of species accurately.

Other studies have used the spectral RS data to classify their images by one of the
traditional methods, and have compared the biodiversity between the distinguished
classes. Nagendra & Gadgil (1999a) found that classes from supervised
classification could distinguish landscape element types of the Western Ghats hills in
India, and contained different species and had different species richness, but those
from unsupervised classification did not. In a study of meadows in the Greater
Yellowstone Ecosystem (USA), unsupervised classification was used in combination
with manual merging of classes based on field knowledge of the vegetation. Here it
was found that species distribution and species diversity did differ between classes
(Debinski et al. 1999).
A disadvantage of using general classification methods is, that no logical connection
is established between the RS signal and the ground data. As Tuomisto (1998) put it:
“The colour pattern in satellite imagery enable one to identify and map areas that
differ in some way; field studies are then needed to find out whether these
differences are significant in ecological and floristic terms”. A contrasting approach is
well summarised by Stoms & Estes (1993): “… richness models could be developed
that relate remotely-sensed data or indices to the underlying biophysical factors and
then to the number of species”.

Classifying an image based on the expected meaning of spectral and/or temporal
signatures, rather than just on their separability, allows for a different type of
hypothesis testing. In this study, the hypothesis was that a relationship exists
between the number of species and the productivity and persistence of productivity in
an area. Two types of analysis were used to test this hypothesis.
1. The study area was divided into distinct classes of mean productivity (indicated by
a vegetation index, as explained below) and variability of productivity, based on
boundary values. These classes were compared mutually in terms of species
diversity.
2. The species diversity and productivity parameters were compared between
different units, defined by the sample units, the vegetation composition or the
landscape type.

2.3   Productivity - remote sensing
Productivity is the rate of conversion of biomass in a certain area in a certain amount
of time. As such a rate is difficult to measure, various substitutes have been used in
productivity-biodiversity studies, for instance biomass or aboveground biomass (e.g.
Gough et al. 1994), rainfall (in Pianka 1966), actual or potential evapotranspiration
(e.g. Hoffman et al. 1994), or ocean depth (e.g. Fraser & Currie 1996). In some
theories productivity is also substituted with resources (Ritchie & Olff 1999) or energy
(Wright et al. 1993). This different terminology could lead to some confusion.
Hoffman et al. (1994) found a negative relationship between plant species richness
and potential evapotranspiration (PET) in arid and semi-arid regions of southern
Africa. They relate the PET to available energy, a relation that is technically true.
However, the ‘energy’ as used in species-energy theory, should be more related to
resources, and in arid environments thermal and solar energy are rarely the limiting
resources. In these regions more evapotranspiration is more likely to have a negative
effect on productivity by reducing the available moisture, and using it as a measure

                                                                                        9
Productivity-biodiversity patterns - a study using multitemporal Landsat TM NDVI data for the Alice Springs region, central Australia - WAGENINGEN UR
Productivity – biodiversity patterns in central Australia: Introduction

for available energy to plants is therefore deceiving. For that reason rainfall is more
commonly used as substitute for productivity in (semi-) arid regions.

Several measures related to productivity can be deduced from multi spectral remote
sensing images. The result of productivity is the standing biomass, providing no
major disturbances or grazing have taken place. Standing biomass is closely related
to cover of green and inert vegetation, which has been successfully described in
Central Australia by the ‘Perpendicular Distance band4-band 5’ (PD54) index, using
Landsat Multispectral Scanner System (MSS) band 4 (500-600 nm) and band 5 (600-
700 nm) (Pickup et al. 1993). Channel 1 from the National Oceanographic and
Atmospheric Administration – Advanced Very High Resolution Radiometer (NOAA
AVHRR) has also been used as an indication of vegetation cover (Bastin et al. 1995).
Plant productivity is directly related to the amount of photosynthesis taking place.
Photosynthesis also causes the typical spectral properties of green vegetation, which
are captured by optical remote sensing and used to calculate vegetation indices.
Vegetation indices are broadly divided into ratio-based indices and orthogonal-based
(or linear combination or n-space) indices (Clevers 1997, Huete 1988, Elvidge &
Lyon 1985). The former are calculated by some ratio between the red (used for
photosynthesis, low reflectance from green vegetation) and near infrared (NIR) (high
reflectance from green vegetation) wavelengths, while the latter are based on the
perpendicular distance to a soil line in a two or more band space.
The index used in this study is a ratio-based index called the Normalized Difference
Vegetation Index (NDVI). The NDVI has been, and still is extensively used for
monitoring of green vegetation cover (e.g. Tucker et al. 1986, Hielkema et al. 1986,
Nicholson et al. 1990, Chen & Brutsaert 1998, Teillet et al. 2000) all over the world,
including central Australia (DPIF NT 2000, Environment Australia 2000, Foran &
Pearce 1990). The combination of NDVI from multi temporal images has been used
to calculate a measure of gross primary production known as Integrated NDVI
(INDVI), or NDVI-days (e.g. Jørgensen & Nøhr 1996, Diallo et al. 1991, Prince 1991,
Foran & Pearce 1990).

The aim of this study was to relate ‘productivity’ to biodiversity. The measures of
productivity we were interested in were not the total production integrated over time,
but the rate of production at different times, as well as the temporal variability in this
rate. We chose to use the mean NDVI of 9 dates as a measure for the average
production rate at a location, and the standard deviation from this mean as a
measure of the variation.

2.4   Expected patterns
At the regional scale at which this research was conducted, the expectation was to
find a unimodal relationship between the productivity and richness in perennial plant
species, even though the whole range of productivity is relatively low (Graham Griffin,
pers.comm.). Where there is very little productivity, the circumstances are apparently
too harsh for much plant growth, and only few species will be able to survive. These
areas are dominated by ephemeral herbs and grasses, which only appear after good
rains, when circumstances are more favourable. In the most productive areas, which
in this environment are the rivers and drainage lines, competition becomes an
important factor, and only a few arid-zone species are adapted to such
circumstances. Intermediately productive areas will allow many species to survive
physiologically, while not allowing heavy dominance of any species.

Spatial heterogeneity was expected to increase species number, but temporal
variability was expected to act in a different way on the studied species: only long-
term perennial plants were included, which have as a common trait their persistence

                                                                                        10
Productivity – biodiversity patterns in central Australia: Introduction

through hard times. They do not partition their environment in time, but they are
dependent on a certain degree of persistence of a minimum level of resources.
Temporal stability should therefore have a positive effect on perennial plants.

Talk about spatial scale can be confusing, as the terminology for map scale is the
opposite for that of operational scale. In this paper a larger scale or higher scale level
means a larger area and less detail, while a smaller scale or lower scale level means
the reverse.

                                                                                       11
Productivity – biodiversity patterns in central Australia: Methods

     3     Methods

     3.1    Study area
     The study area is located in arid central Australia, and includes the town of Alice
     Springs (fig. 1). It was chosen so that the whole area was within reasonable travel
     distance from Alice Springs, but included a
     wide range of landscape types, including red
     sand dunes, hills and mountains of various
     geologies, rivers, low-relief calcareous areas
     and scrubland plains. It transverses the
     MacDonnell ranges from north to south, and
     includes a small part of the plains to the north
     of the ranges and a considerable area to the
     south, including the Waterhouse Ranges. The
     landscape types in the region have been
     classified into landsystems by Perry et al
     in1955 (Perry et al. 1961). These
     landsystems are based on a combination of
     geology, topography and vegetation, and
                                                          Figure 1 Location of the study area (small
     have been widely used in land-use and                box in center).
     environmental studies in the region. The
     landsystems found in the study-area are listed in table 1, and their distribution can be
     seen on plate 1.
     The climate of region is regarded as hot arid, with hot dry summers and cool dry
     winters. Rainfall is low and highly variable (Gentilli 1972). In the year 2000 unusually
     high rainfall has resulted in high cover of ephemeral herbs and grasses and extra
     growth and vigour in perennial plants. Images of similarly green years are included in
     the dataset.
     Table 1 Description of landsystem, after Perry et al. 1961.
Code       Name             Description
NM         ‘Northern        Combination of Alcoota, Boen and Bushy Park lndsystems. All plains N of
           Mulga’           MacDonnell ranges; weathered granite, gneiss or schist, or alluvial deposits;
                            red earths; mulga in groves over short grass or woollybutt.
Hm         Hamilton         Active alluvial fans. Plains flanking crystalline mountains, north of the
                            MacDonnell Ranges; texture-contrast soils with short grasses or Scleroleana
                            spp. or Maireana spp. (Chenopodiaceae), some red earths and red clay soils
                            with mulga and short grass.
Ha         Harts            Mountains of gneiss and granite. Uplands, steep-sided mountains, and hills,
                            relief about 300 m; pockets of shallow gritty and stony soils; sparse shrubs and
                            grasses.
Bs         Bond Springs     Hills and plains of granite, gneiss or schist. Ridges up to 180 m high and
                            rugged terrain with up to 30 m relief; some shallow gritty and stony soils;
                            sparse shrubs and grasses. Narrow plains; various soils; sparse low trees over
                            short grass.
Gi         Gillen           Dissected ranges of folded sedimentary rocks, with summit planation. Quartzite
                            and sandstone ridges up to 300 m high; little soil; spinifex. Vales with alluvial
                            plains and gravel terraces; stony soils (texture-contrast, red earth), red clayey
                            sands, and coarse soils; sparse shrubs and low trees, mulga, or witchetty bush
                            over short grass.
Td         Todd             Coalescent flood-plains of the Todd river and tributaries; sandy alluvial soils,
                            some red clayey sands and silty fine, and layered alluvial soils; sparse low
                            trees over short grass.
Mu         Muller           Undulating terrain on mainly unweathered, folded sedimentary rocks. Low hilly
                            or undulating, relief up to 25 m; calcareous earths; open or witchetty bush over
                            short grass. (Nowadays more mulga than witchetty bush)

                                                                                                   12
Productivity – biodiversity patterns in central Australia: Methods

Code       Name           Description
Ew         Ewaninga       Undulating dune-covered terrain with stony conglomerate hills, relief up to 10
                          m; red dune sands; spinifex mainly under mulga.
Si         Simpson        Parallel, reticulate, and irregular san dunes with stable flanks, minor areas of
                          mobile sands; red dune sands and red clayey sands; spinifex.
Kr         Krichauff      Sandstone mountains, not sampled.

     3.2    Image processing

     3.2.1 Data used
     Nine Landsat Thematic Mapper (TM) images of the same area (path 102, 77) were
     used, their recording dates spread out from February 1988 to March 1997 (fig. 4 and
     appendix 1). The images used were those available at CSIRO CAZR, were mostly
     cloud-free, and covered a wide range of circumstances, including high vegetation
     cover after rainfall events, and low cover in dry periods. The images were subsetted
     to 3 bands (2,3, 4) and the extent of the study area.
     Geometric registration was performed using existing ground control points (GCP’s)
     created by G. Pearce. I used UTM coordinates (zone 53 south) and the Australian
     Geodetic datum of 1984, even though this datum differs from that used for the GCP’s
     and most other spatial data in the area, the Astralian Geodetic of 1966 (AG-66),
     because the latter is not available in ENVI. However, as the two datums are only
     shifted relative to each other, not rotated or distorted, using the wrong datum did not
     cause problems in this case, and the output-image could be treated as a AG-66
     referenced image and overlain with other AG-66 data. Output accuracy was generally
     less than a pixel.
     The software used for image processing was ENVI 3.1 (BSCLCC 1998). All overlay
     operations, visualisations and other GIS functions were performed in ArcView 3.2
     (Esri 1999).

     3.2.2 Effect of atmospheric correction.
     The Landsat tm images were taken at different times of the year and different times
     of the day, having different atmospheric conditions and sun angles. These factors
     can have different effects on NIR and red light, and thereby influence NDVI values
     (Huete & Tucker 1991). An atmospheric correction can minimize the differences in
     reflectance due to these different circumstances. It was investigated whether such a
     correction would be necessary for this application, looking at the effect of a correction
     on the NDVI values and differences between years.
     Because of the limited time available, only the simplest and quickest type of
     atmospheric correction was used, which is the dark pixel subtraction method. This
     method basically presumes that real data should be starting at 0 reflection, which is
     approximately the reflection from deep water or heavy shade. The histograms of the
     original pixel values (fig 2a) can be used to determine for each band the level at
     which real data starts, showing a marked increase from near-zero frequency. All
     values below this level are considered noise. By subtracting this value from all data in
     the corresponding band, the data are made to start at zero (fig 2b). The values
     subtracted from each band can be found in appendix 1, table 2.

                                                                                                  13
Productivity – biodiversity patterns in central Australia: Methods

 a.

   b.

Figure 2 Histograms before (a.) and after (b.) dark pixel subtraction.

Another method for dark-pixel correction, subtraction of the minimum pixel-value,
which is a default option in Envi software, was also tried out. This method was
considered less suitable, because the value of one single pixel determines the value
to be subtracted, making the method very sensitive to outliers and errors.

The NDVI values calculated from the images before and after atmospheric correction
were compared for three images, to see if atmospheric correction would be likely to
have an important effect on the outcome of the analysis. The histograms of the NDVI
at different dates show that the NDVI values change considerably (fig 3). The range
of NDVI values becomes larger for all dates. In June 1988 the green vegetation cover
was high, which is reflected in the high NDVI values. After atmospheric correction
these values increase even more. The NDVI values for February 1988 (before the
rain, having low vegetation cover) are low and become even slightly lower after
atmospheric correction. The low values of February 1990 increase slightly after the
dark pixel subtraction. The irregular behaviour around NDVI zero, as seen in the
histogram before correction, also seems to largely disappear afterwards.
It is concluded that the changes in NDVI values are too large to be ignored, and that
atmospheric correction is necessary for all images to be able to derive meaningful
measures of the variation of NDVI over time.

                                                                                  14
Productivity – biodiversity patterns in central Australia: Methods

Figure 3 Histograms of the NDVI values in 3 of the images, before and after dark pixel
subtraction.

3.2.3 Calculation of NDVI
The Normalized Difference Vegetation Index (NDVI) is a widely used remote sensing
indicator for green vegetation cover. It is based on the typical spectral reflection of
green vegetation, which is very low in the red wavelengths because of energy-
absorption by chlorophyll, and which is high in the near infrared (NIR). The NDVI and
other vegetation indices can be influenced by several non-vegetation factors, such as
soil-background, dead biomass and atmospheric conditions (but the latter have
moslty been corrected for). An alternative method that could be successful for
distinguishing green vegetation cover in this arid environment is spectral unmixing
(Steven de Jong, Vanessa Chewings, pers.comm.). However, in order to use this
method, the spectral signatures of endmembers need to be known, which is not the
case for this area.
Other vegetation indices are available, some of which correct for some of the non-
vegetation influences (e.g. Huete 1988) (see Discussion; Influences on NDVI).
However, the NDVI is the most wellknown vegetation index, and also used in the
project this study is related to, which uses ready-made NOAA NDVI composites.
Foran & Pearce (1990) found a good correlation between NDVI and total green cover
in central Australia. As the observed patterns of NDVI in the images also did not
indicate a problem, the NDVI was considered a suitable index.
The NDVI is calculated by the following formula: (NIR-Red)/(NIR+Red), and was
calculated from digital numbers (DN values). For Landsat TM this NIR corresponds
with band 4, and Red with band 3. The dark pixel subtraction caused some pixels to
have value 0 in TM band 3 or 4, resulting in NDVI values –1 and 1. These values are
outside the normal NDVI range of the data (see fig 2) and were not realistic. Although
only a very small part of the pixels would have this problem, the extreme values
could have an influence by stretching the range of possible outcomes in later
analysis. Therefore the standard ENVI function for calculating NDVI was replaced by
a function that substituted 0 pixel values by 1 before calculating the NDVI, which
would still yield a low NDVI value if NIR reflectance was low, and a high value if the
red reflectance was low (appendix 1, fig 3 & box 1).

                                                                                         15
Productivity – biodiversity patterns in central Australia: Methods

Figure 4 Distributions of NDVI-values. Note the high values in June 1988, green after heavy
rain.
The NDVI values calculated for the images are very low, often below 0. This means
that the DN values of the red wavelengths are often higher than those in the NIR,
indicating a very low cover of green vegetation for most dates.
According to Foran & Pearce (1990), NDVI values in Central Australia typically range
between –0.2 and 0.7. Compared to this range, the values found here are rather low.
The lowest values, in December 1994, correspond to a dry period with very low
vegetation cover, which may have caused NDVI values below the typical range.
Values of 0.7 are very rare in these data, but do occur.

In the variable arid environment of central Australia, the green vegetation cover is
usually rather low, but also very variable. The average greenness and the variability
of this greenness over time are expected to be able to distinguish between some of
the different vegetation types in the area, and to be correlated to species richness.
The mean and standard deviation of the NDVI on the different dates were calculated
to create an image of relative measures of average and variability of green
vegetation cover.

The following functions were used for calculating mean and standard deviation of
NDVI over time:
mean = (Σbi)/9,
stdev = (Σ(bi-b10)^2)/9
Where bi = band i, and i = 1, 2, …9, and b10 = band 10
Bands 1 to band 9 are the NDVI images of different dates. Band 10 is the average
(from first formula).

One of the images, that of March 1995, had a small cloud on it. The NDVI values of
the cloud should of course not be used in the calculation of mean and standard
deviation. The mean and stdev were also calculated excluding March 1995, and
these were assigned to the cloud area using a mask, while the rest of the scene has
the 9-date measures.

The ranges of values of mean and temporal standard deviation of the NDVI were
divided into 5 and 4 classes respectively. The classes were based on the histograms
of the two measures, and boundaries were chosen so that each class was
represented by a reasonable number of pixels, and represented a specific range (e.g.
‘very low’, or ‘medium’) compared to the total range (fig 5). The classes were used
throughout the study for selecting sampling sites as well as statistical analysis. A

                                                                                          16
Productivity – biodiversity patterns in central Australia: Methods

poster-size map was produced based on these classes (larger version of plate 2) and
including roads, which was used during fieldwork.

Figure 5 Histograms of mean and SD NDVI calculated from 9 images. The frequency is the
number of pixels. Lines indicate the class-boundaries.

3.2.4 Spatial variation
The spatial variation in mean NDVI was expressed as the standard deviation (sd or
stdev) in moving windows of different sizes (3x3, 7x7, 17x17 and 33x33 pixels) (plate
3 and 4). This measure was chosen mainly because it is easy to implement, and
gives a usable result. More elaborate methods, such as variograms (see discussion),
may have given more information, but are not readily available in all image-
processing software, at least not in ENVI. Time limitations did not allow for manual
programming of such functions.

An alternative measure for variability that is also easy to calculate, the coefficient of
variation, has been used to measure temporal variation in NDVI in various studies
(Chen & Brutsaert 1998, Fjeldså et al. 1996, Eidenshink & Haas 1992, Tucker et al.
1991), although the standard deviation and variance have also been used for
temporal (Peters et al. 1993, Tucker et al. 1991) and spatial (Coops et al. 1998)
variation. To test the effect of using a different measure of variation, I replaced the
stdev by the unbiased coefficient of variance (Sokal & Rohlf 1995), adding 1 to the
mean NDVI values to avoid using negative values.

3.2.5 Resampling to lower spatial resolution
Originally the intention was to compare RS imagery of different resolutions (Landsat
TM (30x30 m) and MSS (100x100 m) and NOAA AVHRR (1.1x1.1 km) data), but
logistical difficulties did not allow this. Instead the TM images were resampled to a
100x100 m resolution and compared to the original TM image. The resampled image
is similar to a MSS image, except that the spatial accuracy is higher in the TM
derived image, and the bands are a little bit different.
Resampling to NOAA AVHRR resolution was not tried out, but considering the
observed scale of vegetation patterns, this resolution could most probably not
distinguish between vegetation types and local differences in vegetation greenness,
or be related to local species numbers.
In a study of species richness in all of inland Australia, temporal NDVI signatures of
4x4 km NOAA AVHRR composites are being used. In that study compared to this
one, not only the spatial scale changes, but so will the range of values represented,
one might say the ‘data-scale’ or extent (Waide et al. 1999). It will be interesting to
see what relationships are found at this coarse scale (Graham Griffin pers. comm.).
3.2.6

                                                                                        17
Productivity – biodiversity patterns in central Australia: Methods

Using the TM data at high spatial resolution allows small objects to be distinguished,
but also has some disadvantages. When studying large areas, the large file-size of
detailed data slows down processing considerably. Also, the spatial accuracy of the
field-data is such, that they may be compared to the wrong pixel-value, especially if
the pixels are small. Merging the 30-m TM data to larger pixels reduces the file-size,
and produces a larger accuracy of field-data appearing in the right pixels. However,
some detail is also lost, so that for instance small drainage lines may be
indiscernible. The small field-plots cannot represent the much larger pixel-areas (or
‘ground resolution elements, Atkinson 1997). This means that the field-data also
need to be merged to produce comparable scales in both datasets.
For 100x100-m resampled images the transect data were used, rather than the plot-
data (see fieldwork methods), because the 5 plots, each with 20-m radius, in each
transect have a support-area (5*2*π*20 = 628m2) more similar to the pixel size
(10.000 m2). Because the transects are linear and about 200 m long, they do not fall
within 100-m pixels. They are likely to cover more variation than a square area would
have, and the species number would therefore be an over-estimation of the number
of species. However, the area is smaller than that of a pixel, which would produce an
under-estimation. Both the over- and under-estimation will be greater in the more
heterogeneous areas, but the latter will be stronger in areas with small-scale
variation (within 100x100-m areas), while the latter depends on variation at a slightly
higher level (between 100x100-m areas). Also the shape of the spatial patterns is
important – linear patterns such as longitudinal dunes or surfacing tilted geological
layers, both common in parts of the study area, could cause lower variation in long
transects than in square areas, depending on the direction of the transect. This
complex interaction of under- and over-estimations was presumed to average out,
and not further considered.

The red and the first NIR band of the TM-images (band 3 and 4) were resampled to a
100x100-m resolution, and the NDVI recalculated. The resampling was done with the
original DN values, because this average could simulate data from lower resolution
remote sensors, for instance 100x100-m Landsat MSS data. Simply averaging the
NDVI values would not have produced the same result. This contrasts with the
temporal averaging procedure, where NDVI values were used. A recalculation of the
NDVI from average DN values would have been meaningless in this case.
The resampling did not significantly change the result of any of the analysis, which
could lead to the conclusion that the MSS resolution is as suitable for relating
biodiversity at these scales to NDVI from RS as TM data is. In that case it would be
preferable to use MSS data, because of the lower cost, storage space and
computation time.

3.3   Fieldwork
Two separate surveys provided the ground data for this study. Both consisted of
transects of 20-m radius plots, within which the presence of long-term perennial plant
species was recorded. Annual and biannual species were excluded, because these
are very variable in time. Most annuals are ephemeral herbs and grasses that only
appear after good rains. Many do show preferences for certain habitats, but they
were not expected to respond strongly to productivity patterns, because they are
vegetatively absent during low-productivity times (but see discussion). Some species
are short-term perennials under favourable circumstances, but these were also
excluded, for the same reasons. Species were either identified in the field, or
collected for later identification (Urban 1990, nomenclature from Albrecht et al. 1997)
The first survey was conducted in 1995, and concentrated on the mountain ranges. It
was aimed at collecting data for constructing species occurrence models based on
environmental variables such as substrate and hydrology. Transects were 1 km long,

                                                                                     18
Productivity – biodiversity patterns in central Australia: Methods

with 50 m between the centres of the individual plots. They were positioned within
geological strata (rock types), but laid out to cover a large range of elevations within
the strata (Griffin 1997a).
The second survey was carried out in August and September of 2000, and included
some landscape types not surveyed in 1995, such as sand dune areas and low relief
scrubland plains, as well as some more samples in the mountainous areas. This
survey was conducted specifically for this study, and the location of transects was
based on a map of mean and stdev NDVI classes. Transects were chosen to fall
within a combination of mean and stdev NDVI classes, and were located within
reasonable walking distance from roads and tracks. Transects consisted of 5 plots,
with their centres 60 m apart. Apart from species presence, an estimation of the
ground cover was recorded. Grasses, herbs, low and high shrubs, and trees were
recorded separately, in 6 cover-classes (
Productivity – biodiversity patterns in central Australia: Methods

   33% cover of grass (cover class 2) and 33% herbs (2), and 34% bare ground (2),
   then ground cover is 5*4/6 = 3.3. If herbs and grasses are 22% (both 2), bare
   ground is 56% (3), and ground cover is 5*4/7 = 2.9.
2. Ground cover = 5 * (grass cover + herb cover + spinifex cover) / (grass cover +
   herb cover + spinifex cover + bare ground + rocks).
3. Shrubs and trees cover = low shrub cover + high shrub cover + tree cover
4. Cover of perennials = shrubs and trees cover + spinifex cover
5. Total cover = ground cover + shrubs and trees cover
6. Total unvegetated = bare soil + rocks.
7. Spinifex cover was also tested separately, but as a vegetation type rather than a
   cover class.
3.4.1.2   Species richness vs. mean NDVI and temporal variation.
An analysis of variance (ANOVA) was carried out to compare the classes of mean
and stdev NDVI and their interaction. Landsystems were also included in an ANOVA,
to see whether NDVI classes behaved differently in different landscapes. Bar graphs
of the species richness per class of combination of classes were used to see the
nature of any differences.
Correlations were calculated between species numbers per plot, per transect and per
landsystem, and mean NDVI, stdev NDVI, and spatial variation at different scales.
3.4.1.3   Spatial variation
Spatial heterogeneity has often been found to correlate positively with species
richness (Pollock et al 1998, Wright et al 1993). The relationship between spatial
heterogeneity and species richness was investigated at several scale-levels.
Spatial heterogeneity was expressed as the standard deviation of the NDVI (the
temporal mean) in environments of different sizes: 90x90-m, 210x210-m, 510x510-m
and 990x990-m and landsystems. The total variation between pixels per landsystem
however, was not independent from the size of the landsystem, so that we used the
average of the spatial variation in the different sized windows rather than the total
variation.

Species numbers per plot and per transect were compared to the spatial
heterogeneity in their surroundings. It was also expected, that higher spatial
variability between pixels would increase the number of species per transect
compared to the numbers per plot. The ratio of species number per plot and species
number per transect was tested for correlation with the spatial variation.
At a higher scale level, the number of species in a landsystem was compared to the
mean and temporal and spatial variation of NDVI values in that landsystem.
3.4.1.4   Vegetation types
The species presence in the plots was combined to derive the frequency of species
occurrence per transect (5 plots per transect). These frequency data were used to
cluster transects into groups representing vegetation types. The clustering was done
in the pattern analysis package PATN, and based on the Bray & Curtis association
measure (Belbin1995). A dissimilarity matrix is calculated, and based on that the
samples are grouped. The level at which the groups are returned, and hence the
number of groups, can be adjusted.
The 400 transects were classified into 20 classes, some of which were manually
merged. Merged groups were those consisting of few transects and all of those on
grass/herb areas – these all had very few perennial species, separating them out too
easily.
Based on vegetation structure and the main constituting species, a vegetation class
was also assigned to each transect based on direct field observation during the
second survey period. These classes were compared with the calculated groups,
which, together with an inspection of the species encountered in each group, allowed
for interpretation and manual merging of the groups. Groups not represented in the
second survey were not merged.

                                                                                   20
Productivity – biodiversity patterns in central Australia: Methods

An ANOVA was performed to assess the significance of the differences in NDVI
signals and species numbers between vegetation groups. Bar graphs ordered by
different measures were used for examining the shapes of their relationship

                                                                               21
Productivity – biodiversity patterns in central Australia: Examples

  4       Examples of vegetation types and their NDVI signal

  4.1      Alice Springs
  The town of Alice Springs, the only town in the study area, offers a good example to
  explain the legend of the NDVI-classes map (see below). The town shopping and
  business centre shows up as being a dark red. This built-up area is never very
  productive in terms of vegetation growth.
  Around the centre there are various residential areas. Gardens are persistently
  watered, resulting in dark shades of green.
  The pink and yellow pattern around the town corresponds to gneiss hills with sparse
  shrubs. The linear feature south of town is a quartzite ridge, which is part of the
  Heavytree range. The Todd River is another feature that readily recognized. It cuts
  through the Heavytree ridge at Heavytree gap, and continues south from there with a
  wide bare riverbed and green banks.
  Some other human influences can be seen in the location of the rubbish dump,
  showing up bare with little variation, and the area of sewage ponds, very variable but
  on average quite productive thanks to extra inputs.

                                                                      Alice Springs

                                                                      Town centre

                                                                      Residential areas

                                                                      Heavytree Gap
                                                                      Heavytree range

                                                                      Todd River

                                                                      Rubbish dump

                                                                      Sewage ponds

              5 km
                     Mean NDVI 
             12       22    32         42       52
SD NDVI

             14       24    34         44       54
   

             16       26    36         46       56
             18       28    38         48       58

                                                                                          22
Productivity – biodiversity patterns in central Australia: Examples

    Rivers
    This image shows the Hugh River running through
    low relief country (top, landsystem Mu) and cutting
    through a range of quartzite and sandstone ridges
    (photograph, landsystem Gi). The river shows up
    clearly as being very green. The dark green
    indicates that temporal variability in greenness is
    limited. This can be attributed to the presence of
    river red gum (Eucalyptus camaldulensis), which is
    a large tree, resistant to the forces of heavy floods,
    that can get water from underground sources even
    in dry times. The river has several plaiting
    channels in most places, causing an alternation of
    green banks with trees and the actual channels,
    which generally consist of bare whitish sand and
    pebbles, and floodplains with grass-cover, both of
    which show up as having lower greenness values.

    Photograph: the Hugh River entering the
    Waterhouse Ranges. Trees are river red
    gums.

                                                  5 km

                Mean NDVI 
          12     22    32          42       52
SD NDVI

          14     24    34          44       54
   

          16     26    36          46       56
          18     28    38          48       58

                                                                                     23
Productivity – biodiversity patterns in central Australia: Examples

          4.2   Acacia scrubland
          Acacia scrubland is a very common vegetation type in the study area. It nearly always
          shows up as mean-NDVI class 4, being relatively green. The temporal variability is
          relatively low.
          The photograph shows a mulga (Acacia aneura)-dominated scrubland. This is by far
          the most common Acacia scrubland encountered in the area, followed at some
          distance by witchetty bush (Acacia kempeana). The soil was covered with herbs and
          grasses when this photo was taken. This cover is much lower in dry periods (Melinda
          Hillery, pers. comm.).

            1 km

                   Mean NDVI 
SD NDVI

           12       22    32        42      52
           14       24    34        44      54
   

           16       26    36        46      56
           18       28    38        48      58

                                                                                            24
Productivity – biodiversity patterns in central Australia: Examples

      4.3    Sand dunes
      The south of the study area is dominated by red sand dunes. These are mostly old
      inactive dunes covered in spinifex (Triodia basedowii and Triodia pungens). NDVI
      levels are intermediate here, and have relatively low spatial and temporal variability.
      The Central Australia Railway is clearly visible running North-South.

  1 km

                 Mean NDVI 
SD NDVI

            12    22    32        42      52
            14    24    34        44      54
   

            16    26    36        46      56
            18    28    38        48      58

                                                                                            25
Productivity – biodiversity patterns in central Australia: Examples

4.4    Mountain ranges
 Many mountains in the area have a clear structure of tilted geological layers,
resulting in linear shapes. The Waterhouse range (bottom image) consists of two
ranges of the same resistant quartzite layers, surrounding a valley cut in more
erodable rock types. The alternation of layers of different rock-types is reflected in a
similar pattern in the vegetation, which makes the fold-structure clearly visible on the
image. The high parts of the Waterhouse Range are quite well vegetated with Acacia
scrubland (see photograph), while the valley, has mostly open grassland and
spinifex (Triodia brizioides) on dolomite, both giving low NDVI values. Drainage
systems go west and east, and have a species-rich high scrubland vegetation.
The northern part of the MacDonnell Ranges
(top image) consists mainly of hills and
mountains of high-grade metamorphic rocks
including gneiss and granite (Griffin & Tier
1997b). The structures in this area are less
obvious. The vegetation cover is mostly open
mixed scrubland and is usually low due to
the rocky substrate and drought. Variability is
high on a small scale.

           5 km

              Mean NDVI 
         12     22   32         42     52
NDVI

         14     24   34                54
 SD

                                44
         16     26   36         46     56
         18     28   38         48     58

                                                                                     26
Productivity – biodiversity patterns in central Australia: Examples

4.5   Open grassland
Open grassland areas have low NDVI values, although not the lowest possible. In
good years like 2000, they can have full vegetation cover (top photograph), but at
other times these areas may be completely bare (small photograph). When bare, the
red soils of most of these areas may produce a false greenness-signal (Huete &
Jackson 1987 + see discussion). This may cause an increase in the mean NDVI and
a decrease in the variability.
In some grassland areas there are also some dispersed trees or shrubs. These seem
to have little influence on the NDVI values, but do increase the number of perennial
species recorded.

                                                                   Mean NDVI 
                                                   SD NDVI

                                                             12     22    32      42   52
                                                             14     24    34      44   54
                                                      

                                                             16     26    36      46   56
                                                             18     28    38      48   58

                                                                                 27
Productivity – biodiversity patterns in central Australia: Examples

   4.6    Spinifex
   Several types of spinifex-vegetation are found in the study area, some having quite
   distinct NDVI responses. Spinifex on sand dunes typically has intermediate NDVI
   values (see example above). A spinifex species encountered in most parts of the
   study area, is Triodia brizioides, which grows on dolomite or limestone rock outcrops
   and hills. This vegetation type has the lowest NDVI values of all. At the time of the
   second survey, the spinifex was all flowering, resulting in view not unlike a field of
   cereal, and certainly not very bare looking. The actual ground cover of the tussocks
   was around 50% in most areas. Three factors may explain the low NDVI values. The
   typical substrate of this vegetation has much less colour than most other rocks and
   soils, which would give it less false greenness as compared to the spinifex on red
   sand and to the grassland areas (Huete & Jackson 1987). An adaptation of spinifex
   to the arid environment is having a very small leaf-surface, to reduce water-loss.
   Also, the tussocks are long-lived and contain rather a large amount of dead biomass.
   This fraction will probably increase during dry periods. Both these characteristics can
   reduce the NDVI (see discussion).
   These vegetations contain a variable, but relatively high number of other perennial
   species. The photograph shows a Triodia brizioides area in the foreground, Acacia
   scrubland in the distance, and the northern edge of the MacDonnell ranges.

                Mean NDVI 
SD NDVI

          12     22    32          42      52
          14     24    34          44      54
   

          16     26    36          46      56
          18     28    38          48      58

                                                                                       28
Productivity – biodiversity patterns in central Australia: Results

5     Results

5.1    Species.
In total, 94 perennial plant species were found and included in the surveys. The main
plant families represented were the Mimosaceae (Acacia species), Caesalpinaceae
(Senna species), Proteaceae (Grevillea and Hakea), Myoporaceae (Eremophila
species), Myrtaceae (Eucalypts and tea tree), and Gramineae (Triodia species) (see
appendix 3 for complete species list). Vegetation types encountered ranged from
open areas with annual herbs and grasses to floodplain forests, with different types
of scrubland in between, as well as at several distinct spinifex communities.

5.2    Class distinction in the field
In the field the patterns on the NDVI-classes map could be recognised very well in
the vegetation, especially outside the MacDonnell ranges, because in the mountains
patterns are finer. Outside the mountains, the lowest NDVI values corresponded with
dolomite outcrops areas with spinifex (Triodia brizioides) occurring in many
landsystems. The next class up were areas with only grass and herb cover, a lot of
which is annual growth, and sometimes some widely dispersed trees or shrubs.
Medium NDVI values mostly corresponded with spinifex vegetations on red sandy
soils, and with Senna and some mixed low scrublands. The fourth mean-NDVI class
was typical for Acacia-scrublands dominated by mulga (Acacia aneura) and witchetty
bush (A. kempeana), sometimes combined with other high shrubs. The highest NDVI
class was found in rivers, the outflow of the Todd River and local drainage areas, and
usually indicated either river red gums (Eucalyptus camaldulensis) or a mix of high
trees and lush undergrowth of (introduced) grasses (but these grasses may look very
different in less favourable years).

5.3    NDVI values
The NDVI values used in the analysis were not the actual NDVI values as computed
in ENVI (e.g. compare fig 5 and fig 6). Something went wrong either during stretching
of the NDVI values to integer values between 0 and 255 in ENVI (done to be able to
transform the file to an ArcView *.bil file), or with the conversion from ENVI to
ArcView. Although the stretch owas a linear one changing the data from –1 to 1, into
0 to 255, the data obtained from ArcView did not return the same values when
‘destretched’ using the opposite formula (x/127.5 – 1). Relative to each other, the
values are still valid however, so they can be used for relative comparison without
problems.

5.4    Groundcover
                                                                 12
The relationships between
                                        Total vegetation cover

vegetation cover and NDVI                                        10
values in plots are shown in
                                                                 8
table 2. Mean NDVI has the
best correlation with total                                      6
vegetation cover (fig 6), while                                  4
NDVI stdev is best correlated
to the cover of grass and                                        2

herbs, the cover type notably                                    0
most variable. Perennial cover                                    -0.4   -0.3   -0.2   -0.1   0   0.1   0.2   0.3   0.4
is not correlated to NDVI                                                                Mean NDVI
variation, positively or
negatively. Unvegetated                 Figure 6 Relation between the total vegetation cover,
ground is negatively correlated         estimated from field observations, and the mean
                                        NDVI from 9 Landsat TM images (r=0.570).

                                                                                                                          29
You can also read