Grassland patterning and habitat suitability for granivorous
birds
Leader: Dr Stephen Garnett, Queensland Environmental Protection
Agency, Cairns
Project 2.2.2
Summary | Progress |
Two-dimensional modelling | One-dimensional
spatial analysis | Future directions | Methods for assessing
landscape health | One-dimensional spatial
analysis of grassland patterning | Description of trees in savanna
landscapes | Identifying water holes and areas
of dampness | Fire scar mapping | Quantifying greenness in the
landscape | Project team |
caption text
Geographic
 |
|
The fundamental aim of this project was to determine how the
spatial patterning of graminoids (granivorous birds) and grass
seeds varies across savanna landscapes. Research to date has aimed
at determining the relationship between environmental patterns at
different scales and the abundance of granivorous birds during the
wet season. The data will also be used to determine relationships
between abundance and vegetation structure for a variety of
non-granivorous birds, some of which are declining in parts of
their range.
See figure at right which maps the geographic index of decline
of granivorous birds across the tropical savannas.
A substantial dataset was gathered from six transects across the
Yinberrie Hills (north-west of Katherine in the Northern Territory)
landscape and efforts have concentrated on trying to extend the
one-dimension linear transect data to two dimensions by analysis of
multispectral imagery. This was done in order to be able to map and
measure the distribution of grassland patches across this
landscape.
Issues surrounding determining scale of patterning in landscapes
are very complex. This research contributes towards the wider body
of knowledge in this field. As far as the researchers are aware,
the use of IKONOS imagery for determining grassland patterning had
not been explored prior to this study. One-dimensional analysis of
pattern can be applied to other similar areas such as the Victoria
River District (Northern Territory) for determining the scale of
patterning and patch size. Methods that look at spatial dependence
and the level of variability in the landscape may also have
potential for modelling the structure of savanna grassland
landscapes. Launched in September 1999, the IKONOS sensor produces
1-metre resolution Panchromatic, 1-metre resolution Pan-sharpened
Multispectral and 4-metre resolution Multispectral (colour)
products.
At the beginning of July 2000, IKONOS multispectral remotely
sensed imagery was captured for a site covering five of the six
transects. This data was used in association with the transect
field data to try to examine how the spatial patterning of
graminoids varied across the landscape. The resolution of this
imagery was 1 metre for the panchromatic and 4 metres for the
four-layer multi-spectral data. Analysis of these two data sources
consisted of firstly trying to classify the multispectral IKONOS
imagery using the presence/absence graminoid data as ground
truthing. This should have determined the presence of, and area
covered by, the grass species recorded along each transect. The aim
being that once a classification had determined location and extent
of patches of each of the graminoid species in the area, patch
statistics could be applied to measure characteristics that may be
relevant to granivorous birds and the health of the landscape.
However, results showed that there were areas of contiguous
pixels within the classified image but the majority of these seemed
to correspond to bare soil or fire scar areas. In fact, large areas
of the ground cover appeared to have been burnt prior to data
capture or even were burning at the time of capture.
What tended to exist was a mosaic of classes depicting the
non-bare soil areas between trees. Where groups of potential
'grass' pixels did occur, they did not fall along the transect
lines and therefore could not be identified or evaluated. Using the
panchromatic imagery the 100-class image was evaluated and many
areas that were thought to be bare soil were in fact revealed to be
shadow from the canopies of the trees. Shadow from trees is a
significant factor within the image. An attempt was made to remove
shadow classes but there was spectral overlap with other classes so
this was impossible. It is suspected that spectral interference
from tree canopies or the shadow from the canopy, change in soil
colour, and burnt areas, or a combination of these factors all
contributed to the lack of consistency in the classes.
Because of the large number of possible combinations of
vegetative species and the 11-bit nature of the image, a 300-class
classification was used to try to separate one species from
another. Once again visual evaluations were carried out and the
result was that there was no correlation between the contiguous
class pixels in the transect area and the survey data. The next
step was to create ratio layers to try to emphasise the vegetation
and the differences within the vegetation in the image. Ratioed
layers are used to eliminate the effects of topography and to
emphasise subtle differences of characteristics within the image.
The near infrared to red ratio is called Ratio vegetation index or
RVI and emphasises healthy vegetation. This uses bands 4 / 3 or NIR
/ Red stacked with a NIR / Green layer. An unsupervised
classification of 100 classes was then carried out. The result of
this was that large areas of bare soil and shadow in the image
showed up well and wet areas were quite readily discernible from
the ratio-based images.
However, again this analysis failed to reveal anything useful in
terms of defining patches of grassland species and healthy areas of
vegetation tended to correspond with tree canopies. It was
concluded that the 4m-resolution imagery was probably not of a
sufficiently fine scale to correlate with 1m presence/absence data.
As a result, the 4 m multispectral imagery was merged with the 1 m
panchromatic to produce a multispectral dataset of 1 m resolution.
A 90 class unsupervised classification was carried out and then a
Bayesian modelling routine implemented in ArcView GIS (Aspinall,
1992) was used to try to produce probability surfaces showing the
possible locations of each of the species using presence/absence
data.
However, again this analysis failed to produce any significant
results even when the area of analysis was reduced to cover an area
immediately surrounding a particular transect. So even the 1 m
resolution data does not seem to be capable of producing a
two-dimensional representation of grassland patches when used in
association with the linear field data. Problems encountered could
be due to the resolution of the imagery still not being fine enough
to detect grassland patterning. Another limiting factor in the
results of this analysis is the fact that the field and imagery
where captured at different times of year.
The imagery was captured in the dry season when the grasses had
'browned off' and there had been substantial burning in the area,
whereas the field data was captured in the wet season. Also there
are problems associated with the inaccuracies of georeferencing the
data. Since the field data are points, if the location of these is
out by a matter of 1m in terms of the GPS reading this can have
major implications in terms of trying to identify that point on the
imagery. Also there will be a certain level of inaccuracy
associated with the georeferencing of the IKONOS imagery.
Tree and shrub species data were collected only along one
transect. It was attempted to use this data to try to classify a
local area around that transect but again the results of the study
were unsuccessful. Problems arose because exact locations of trees
were not measured, only distances off the transect line were
recorded which made pin pointing individual trees difficult. Due to
the inability to classify the imagery successfully it was not
possible to map the distribution of grassland patches and therefore
the next stage of analysis which was to run standard patch
statistics on the two dimensional data, which might give some
indication of health and suitability of the habitat for
granivorous, had to be abandoned at this stage. Other approaches to
quantifying pattern were investigated.
These included measures of spatial autocorrelation to look at
the amount of spatial dependence or clustering present in the
imagery. A GIS program (Pearson, 1998 and 2000) which produces
surfaces of local spatial autocorrelation based on Geary's C was
applied to the raw IKONOS imagery. The output identified areas
where clustering occurs. Areas of positive autocorrelation could
correspond to significant patches of grassland but field data
collected along transects was insufficient to ground truth these.
In order to verify these patches more fieldwork is needed that
examines what each patch relates to in the field.
Results of this analysis also show that 2/3 of the area is
negatively spatially autocorrelated which highlights the variable
nature of savanna landscapes and re-emphasises the problems of
trying to treat these landscapes as areas where homogeneous patches
can be identified. One-dimensional spatial analysis Due to the
problems associated with trying to expand the data to two
dimensions, one-dimensional spatial analysis of the linear
presence/absence field data was carried out to see if this could
reveal any information on the scale of pattern and scale of patches
for grassland species found in the area.
The main statistics applied were local quadrat variance and new
local variance (Dale, 1999 and Ludwig, 1979). These statistics were
implemented in a GIS environment as an ArcView script. Since each
of the transects is about 5 km in length, the ArcView script took
several days to run per species using a block size of only 500. Due
to time restrictions it was only run on some of the key grass
species that were identified in last year and for limited block
sizes. These species were Alloterposis semialata, Chrysopogon
fallax and Heteropogon triticeus.
The results of this analysis produced variance plots which
showed that the scale of pattern for Alloterposis appeared to be at
about 38 m (another peak occurred at 182 m which could indicate
pattern at another scale) with an average patch size of about 1m.
The scale of pattern for Chrysopogon appeared to be at 70 m (again
another peak occurred at 430 which could indicate pattern at
another scale) with an average patch size of 2m. The scale of
pattern for Heteropogon was 118 m (another peak was observed at 253
m which could be another scale of pattern or just the pattern at
118 m being replicated) with a average patch size of about 11m.
Analysis of pattern in perennials was also conducted by looking at
density per 50 m quadrat by distance to produce a series of
graphs.
The results of analysis to date seem to indicate that available
imagery and field data will not provide a two-dimensional
representation of grassland patterning in the area. A new image
captured at the end of the wet season was sought to see if this
improved the result—however this imagery could not be
captured at the time required. Linear point data also seems to have
major limitations for scaling up to a two-dimensional landscape
representation of pattern, with the nature of the data making it
difficult to visualise pattern.
Some indication of the scale of pattern can be obtained using
one-dimensional spatial analysis in the form of local quadrat
variance but no single analysis of variance gives all
characteristics of pattern present and scale of patches. Pattern
detection using these linear statistics is made more difficult by
fact that pattern may be present at more than one scale.
In order to be able to classify high-resolution imagery such as
IKONOS data to produce a two-dimensional surface representative of
pattern in the Yinberrie hills landscape it is necessary to make
sure that the date of imagery capture and field data capture
correspond as much as possible.
Additionally, to identify individual species of grass from an
image such as this requires coordinates of ground truth areas that
are mono-specific of each grass type being identified. Each area
would need to be uninfluenced by the presence of trees or shadows
from trees and, would need to be of a constant and reasonable
density so that the pixels representing the area could be of a
consistent type. Several areas of each grass species would be
required. If this data were available it would then be possible to
identify similar areas and it may be possible to find areas that
are predominantly, but not exclusively, of one grass species.
It is generally accepted that for such training areas to be
reliable 30 areas of at least 30 pixels each would be required.
This type of field data seem to be necessary to be able to classify
the landscape into patches of individual grass species adequately
and therefore produce a grassland-patterning surface upon which
patch statistics could be run.
Spatial analysis of the one-dimensional linear presence/absence
transect data is possible by applying several methods based on
variance. Three of these methods recommended by Dale (1999) have
been implemented in an ArcView script and can be run on individual
species within a GIS. Due to the length of each transect the
program does take considerable time to run but it would be possible
to evaluate the scale of pattern and average patch size for some of
the grassland species in the area that are seen to be most
important for granivorous birds.
The 4m resolution multispectral IKONOS image is sufficiently
fine to be able to identify individual trees spaced through the
area. It may be possible to identify individual species if several
test trees of each species had precise GPS positioning. Dense
canopies along riverbanks are also easily identified. Again the
nature of the way that the field data for trees was captured and
the fact that it was restricted to only one transect limit its use
at this stage but useful information could be obtained if more
appropriate field data was collected.
The IKONOS imagery seemed to be able to identify water bodies
and areas of dampness very well. Further studies in this area could
be useful for identifying permanent water bodies that are
inaccessible to humans but are important to granivorous birds in
the dry season.
The depiction of bare soil and fire scars is presented well in
the IKONONS imagery. In areas that are affected by fire scars, the
canopies of trees can still be discerned. It should be possible to
judge the intensity of fires affecting each area by the level of
remnant canopy in the fire-affected areas of this image.
Ratio images can be used to depict greenness through the area.
By removing all green tree canopy areas from the image it may be
possible to estimate the green coverage of grasses. To achieve this
it would be necessary to acquire an image of the area taken whilst
the grass was still thriving and green. The image would need to be
acquired before the area began to support fires and would also need
to be acquired on a day when there was little, or preferably, no
cloud cover.
Stephen Garnett QPWS
Diane Pearson NTU
Milton Lewis PWCNT
David Hooper PWCNT
Jan Riley NTU
Zhang Yue NTU
Articles
Fiona Fraser
Australian National University Canberra: Completed The ecology of the partridge pigeon and habitat impacts due to fire and grazing Summary | Habitat preferences | Variation in home range size | Reliance on specific grasses… [
read more...]
Fire, grazing and partridge pigeons
Partridge pigeons are one of a large number of tropical seed-eating birds whose abundance and distribution have declined this century Fiona Fraser one of the TS-CRC’s PhD students has been studying the needs and habits of the… [
read more...]
Grazing lands may be key to survival for rare native finch
Article on one of Australia’s most vulnerable birds, the southern black-throated finch found around Townsville. From Savanna Links, Issue 31, Jan - June 2005 [
read more...]
Figure One: Geographic index mapping the decline of granivorous birds in the tropical savannas The red areas are those where the declines have been most marked Click here to go back to: Grassland… [
read more...]
Paradise falters for seed-eating birds
This article by Don Franklin of CDU outlines concerns about the decline in seed eating bird populations [
read more...]
Project to save the Gouldian Finch
The Gouldian Finch Erythrura is a small grass-seed eating bird found only in the savanna grasslands of northern Australia It is now far less abundant than previously and has vanished completely from… [
read more...]
