Content uploaded by Kristen Jennifer Williams
Author content
All content in this area was uploaded by Kristen Jennifer Williams on Dec 28, 2020
Content may be subject to copyright.
Multispectral image analysis of woody
vegetation extent and disturbance on the
Popondetta Plain of Papua New Guinea:
Identifying areas of potential habitat for the
Queen Alexandra Birdwing Butterfly
Ralf-Dieter Schroers and Kristen J Williams
28th February 2008
A report for Conservation International Melanesia
CSIRO Sustainable Ecosystems, Atherton Australia
Enquiries should be addressed to:
Mr David K Mitchell
Melanesian Centre for Biodiversity Conservation
Conservation International PNG Programme
PO Box 943 Alotau
Milne Bay Province
Papua New Guinea
dmitchell@conservation.org
Distribution list
CI PNG Alotau
5 hard copies plus digital report
CI Melanesia CBC Atherton
5 hard copies plus digital report
CI CELB Arlington
5 hard copies plus digital report
Important Notice
© Conservation International Melanesia and CSIRO Sustainable Ecosystems, Atherton
Australia, December 2007
While every care is taken to ensure the accuracy of this report, Conservation International
and CSIRO make no representations or warranties about its accuracy, reliability,
completeness or suitability for any particular purpose and disclaims all responsibility and
all liability (including without limitation, liability in negligence) for all expenses, losses,
damages (including indirect or consequential damage) and costs which you might incur as
a result of the product being inaccurate or incomplete in any way and for any reason.
Use of this Report
This report is intended for internal use by Conservation International and partners to inform
development of a conservation plan for the Queen Alexandra Birdwing Butterfly (Ornithoptera
alexandrae) on the Popondetta Plain.
Citation
Schroers, R-D and Williams, K.J. 2008. Multispectral image analysis of woody vegetation
extent and disturbance on the Popondetta Plain of Papua New Guinea: Identifying areas of
potential habitat for the Queen Alexandra Birdwing Butterfly. Conservation International
Melanesia and CSIRO Sustainable Ecosystems, Atherton.
Multispectral image analysis of woody vegetation 3
Acknowledgements
The work presented here was funded through a grant initiative from Cargill following
discussion with the Centre for Environmental Leadership in Business (CELB) division of
Conservation International. It was managed by John Buchanan, Senior Director Business
Practices and Brian Gurr, Manager Agriculture of CELB in consultation with David Mitchell,
Conservation Strategy Specialist of the CI field program in Papua New Guinea. Additional
funding for this work was provided through the Conservation International PNG Program, the
Melanesia CBC Technical Team and CSIRO Sustainable Ecosystems.
We are grateful to Mr R.J.L. Storey (Conservation International Melanesia) for providing the
data on visually delineated land use and remnant vegetation on Popondetta Plain that was
used as a point of comparison with the remote sensing classifications presented here.
The remote sensing analyses presented here benefited from discussion and feedback with
Mr David Mitchell Conservation International PNG) and Eddie Malaisa (Oro Provincial
Government Environment Officer PNG). David Mitchell supplied valuable information about
trends in land use change of the Popondetta study area.
Phil Sherman (Department of Geography University of Papua New Guinea, Port Moresby)
kindly provided the SPOT4 imagery scene (September 2002) and the SPOT2 pan chromatic
data set (July 2002) for the Popondetta Plain.
Robert Denham (Department of Natural Resource and Water, Brisbane, Queensland)
provided advice on ground validation procedures.
Higaturu Oil Palm kindly provided GIS data on the location of Estate and smallholder oil palm
blocks that were used as a point of comparison with the remote sensing classifications of oil
palm.
Multispectral image analysis of woody vegetation 4
Table of Contents
Acknowledgements ............................................................................................................. 3
Executive Summary ............................................................................................................. 5
Objective 1 Woody vegetation mapping to support the identification of habitat for the
Queen Alexandra Bird Wing Butterfly on Popondetta Plain: An application of
multispectral imagery classification .................................................................................. 6
Introduction .......................................................................................................................................... 6
Image data for classifications .............................................................................................................. 6
Imagery used for ground verifications ................................................................................................. 7
Visual classification using Landsat imagery ........................................................................................ 9
Supervised classification ................................................................................................................... 10
Landsat Imagery ............................................................................................................................ 10
SPOT4 Imagery ............................................................................................................................. 13
Ground validation ................................................................................ Error! Bookmark not defined.
Landsat Imagery ............................................................................................................................ 17
SPOT4 Imagery ............................................................................................................................. 19
Unsupervised classification ............................................................................................................... 20
Landsat Imagery ............................................................................................................................ 20
SPOT4 Imagery ............................................................................................................................. 21
Woody remnant vegetation on the Popondetta Plain ........................................................................ 23
Conclusions ....................................................................................................................................... 26
Objective 2: Remote Detection of Disturbance among Mapped Remnants of Woody
Vegetation on the Popondetta Plain ................................................................................. 28
Introduction ........................................................................................................................................ 28
Exploring visually identified woody remnant vegetation .................................................................... 28
Unsupervised Classification of Remnant Vegetation ........................................................................ 41
Woody Remnant vegetation change: Change detection employing Landsat NIR bands (1989 -
2002) ................................................................................................................................................. 42
Spectral mixture modelling ................................................................................................................ 42
Discussion and Conclusions ............................................................................................ 47
Issues associated with landcover mapping ....................................................................................... 47
Recommendation and future work .................................................................................................... 48
Source and availability of high resolution imagery ............................................................................ 49
References ......................................................................................................................... 51
Multispectral image analysis of woody vegetation 5
Executive Summary
The Queen Alexandra Birdwing Butterfly (Ornithoptera alexandrae), the world’s largest
butterfly, is globally Endangered (IUCN Red List) and therefore a priority for conservation
within the New Guinea Wilderness. Information about current land use and land cover is
needed to support the identification and verification of potential habitat areas.
This report presents the application of remote-sensing techniques to automate the
delineation of native vegetation cover and disturbance on the Popondetta Plain. The results
will be used to inform a conservation strategy to ensure the survival of the Queen Alexandra
Birdwing Butterfly on the Popondetta Plain.
Remote sensing imagery classification and difference detection techniques were explored as
ways to map land use, vegetation cover and habitat disturbance on the Popondetta Plain of
PNG for the years 2002 and 1989. The approach and results of this work are presented in
two objectives. The first objective discusses the use of Landsat TM7 and SPOT4 imagery in
2002 to classify land cover and particularly woody remnant vegetation extent to support
habitat mapping for the Queen Alexandra Birding Butterfly. The second objective uses
Landsat TM imagery in 2002 and 1989 and explores methods for detecting disturbance of
woody vegetation through change analysis with reference to mapped habitat remnants.
The classification results demonstrate that readily available imagery data can be used to
detect current land cover on the Popondetta Plains, the accuracy of which can be broadly
assessed with verified ground features. Pan sharpened imagery of both Landsat TM7 and
SPOT4 as well as Google Earth imagery were used to define training classes for supervised
classifications and accuracy assessments. This step could be significantly improved with
properly identified and verified ground locations based on field observations. Higher
resolution imagery would also greatly enhance the detection and classification accuracy but
is expensive and often shows cloud cover.
Preliminary work toward the classification of relative disturbance is promising, but requires
further analysis based on the approach of spectral unmixing, which requires specialised
software tools to complete the processing steps and additional field data to support
classification and accuracy assessment.
The remote sensing and visual mapping classifications revealed the potential to be
complementary. The remote sensing data provides additional information about the extent of
woody vegetation and relative disturbance among the visually-delineated remnant vegetation
patches.
The resulting integrated classification of woody remnant vegetation was developed to
support tactical-scale conservation planning for the Queen Alexandra Birdwing Butterfly on
the Popondetta Plain.
This work utilised existing image acquisitions available for 2002. The cost of imagery
precluded the acquisition of more recent, high-resolution data for this project.
Data products described in Appendix 1 are available on request.
Multispectral image analysis of woody vegetation 6
Objective 1 Woody vegetation mapping to support the identification
of habitat for the Queen Alexandra Bird Wing Butterfly on
Popondetta Plain: An application of multispectral imagery
classification
Introduction
Remote sensing imagery classification techniques were explored as ways to map current
land use and vegetation cover on the Popondetta Plain of PNG. The landscape of the
Popondetta Plains is generally of heterogeneous nature, including oil palm estate
plantations, a mixture of scattered oil palm plantings, garden agriculture, open land, forest,
grassland and hamlets. The imagery employed for land cover classification was Landsat
TM7 (January 2002) and SPOT4 (September 2002). A classification of three land use types
is developed, with an emphasis on woody remnant vegetation to support habitat mapping for
the Queen Alexandra Birding Butterfly.
Image data for classifications
The imagery data Landsat TM7 (January, 2002) with spatial resolution of 28.5m and SPOT4
(September, 2002) with spatial resolution 20m were used to create supervised and
unsupervised classifications (Figure 1). The available SPOT4 imagery (courtesy P.
Sherman, University of PNG, Port Moresby) did not entirely cover the area of interest. The
selected classifiers were solely based on reflectance of plant biophysical characteristics. All
classification analyses were conducted using the image processing software ERDAS
Imagine Version 9.1 (ERDAS 2003). Some map presentations were prepared using ESRI
ArcGIS Version 9.2 (ESRI 2007).
Figure 1: The location of Landsat TM7 (red outline) and SPOT4 (blue outline) data used in the classifications,
relative to the Popondetta Plain study area of PNG (map courtesy Google Earth, accessed 22nd February 2008).
Multispectral image analysis of woody vegetation 7
Imagery used for ground verifications
Remote sensing data provides extensive information on spectral reflectance, spatial pattern
distributions and texture of land features with potential to derive representative clusters.
Field data is used to assess the accuracy of derived image classifications, to support
decisions about which classification results best represent the ground features of interest.
Since detailed ground data of known locations was not available, a visual feature
representation of the study area was obtained in its best possible form. This was achieved
by employing pan sharpened imagery of the highest available resolution, as well as Google
Earth image data.
A resolution merge was conducted by fusing the imagery employed for classification
(SPOT4, Landsat TM7) with panchromatic SPOT2 image data (10m spatial resolution). The
resulting higher resolution pan-sharpened Landsat (Figure 2) and SPOT4 imagery (Figure 3)
improved feature representations (patterns and textures) for visual identification and for
establishing training classes used for supervised classifications. A higher resolution image
was achieved with the pan sharpening of SPOT4 than was possible with Landsat.
Additional image data over the study area of Popondetta Plain were sourced from Google
Earth (Figure 4). The study area was represented by Landsat and a higher resolution
imagery of unknown origin, possibly QuickBird. Google Earth data was employed for ground
validation of supervised classification results.
Figure 2: Pan sharpened Landsat image (fused with pan chromatic SPOT2 band, 10 m).
Present landscape, Popondetta Plain
(photo: D.K. Mitchell).
Multispectral image analysis of woody vegetation 8
Figure 3: Pan sharpened SPOT4 image (fused with pan chromatic SPOT2 band, 10m).
Figure 4: Google Earth image, possibly of QuickBird origin (accessed with Google Earth Software on 14th
December 2007).
Multispectral image analysis of woody vegetation 9
Visual classification using Landsat imagery
A previous delineation of land use classes was undertaken through visual inspection of
colour and texture variation in ground feature representations on a pan sharpened, true
colour Landsat TM7 image (dated February 2002, fused with 15m panchromatic layer) by
R. Storey (described in Williams et al. 2007).
Several classes could be distinguished, including cloud, water, large scale oil palm
plantations, mixed garden/small scale oil palm plantings, garden/fallow mosaics with
established cleared areas, recent clearing and areas of remnant vegetation (Figure 5). The
remnant vegetation class defined through this visual interpretation represented “mature
forest” or woody vegetation areas with a denser canopy (Figure 6).
This visual mapping of remnant vegetation was used to support comparison and cross-
validation with the subsequent image classifications.
Figure 5: Land use mapping based on visual inspection of Landsat pan sharpened imagery (2002, 15m) by R.
Storey (described in Williams et al. 2007).
Images from Google Earth showing variation in land use and vegetation in Popondetta Plain
Grassland/burnt Hamlet Primary forest Logged forest Oil palm Estate and buffers
Multispectral image analysis of woody vegetation 10
Figure 6: Visual delineation of remnant woody vegetation by R. Storey (black outline); the pan-sharpened Landsat
(2002) source layer is shown in the background.
Supervised classification
Supervised classification requires spectral signatures from specified, known ground features.
It involves the user’s knowledge of the study area and training classes are created
representing each known land cover category that appears fairly homogeneous on the image
(Lillesand and Kiefer, 2002). Independently derived information (ideally field acquired) is
used to define the training classes which are then used to establish classification categories
from the image. Based on the statistical parameters of these training samples each image
pixel is evaluated and assigned to one of the established clusters through statistical
procedures such as Maximum Likelihood classification. The Maximum Likelihood
classification method employs a probability density function representing the probability that
the category’s spectral pattern falls into a given spectral region (Lillesand and Kiefer 2004).
Training data are used to estimate means and variances of the classes which are then used
to estimate the probabilities. An unknown pixel is assigned to the class with the greatest
probability density at the spectral region of the unknown pixel. If the highest probability is
smaller than the user-defined threshold value, the pixel remains unclassified.
The probability thresholds (weights) for the signatures can be defined by the user. The
software commonly defaults to equal probability weights of 100%. Pixels with probabilities
lower than this value will not be classified. In the following applications, the maximum
likelihood probability threshold of 95% was typically used.
Landsat Imagery
A supervised classification (probability threshold 95%) was conducted for the whole image
scene using the maximum likelihood cluster algorithm. The selection of Landsat image
training sets for the supervised classification was based on the spatial resolution of 28.5 m.
The identification of ground features was undertaken through visual inspection of Landsat
pan sharpened imagery (employing a Landsat pan chromatic band of 15 m from January,
2002, as well as SPOT2 pan chromatic band, 10m, from July 2002). Ten training classes
Multispectral image analysis of woody vegetation 11
were created based on major land cover categories (bare soil, open land, open forest,
grassland, mature forest, oil palm plantations and oil palm smallholdings, garden/mixture,
cloud, shadow, water). The location of these training classes is shown in Figure 7.
Figure 7: Establishment of training classes (red squares or rectangles) for supervised classification (Landsat
2002).
Three oil palm classes were established based on training areas defining three different
growth stages from young to mature plantations. The classification resulted in further
classes such as woody vegetation (mainly mature forest), grassland and open space
(Figure 8).
Figure 8: A supervised classification generating 10 classes (right); Landsat Bands 4,5,3 as RGB (left), The two
stages of oil palm growth are identified (red, green), mixed garden culture (blue) and woody vegetation (right in
yellow) is separated from other open land features (pink).
After merging two classes representing water bodies and further colour adjustment, a cluster
became visible, highlighting the majority of larger-scale mature oil palm plantations
(Figure 9). The “mature forest” class was compared with the visual delineation of remnant
vegetation by R. Storey (Figure 6) as both were based on the same Landsat data (January
2002). The supervised classification was judged to reasonably match the visual delineation
(Figure 10 and 6). However, a considerable amount of the “mature forest” class was also
found to account for the visually identified “garden matrix” category (Figure 10). The final
classification of “oil palm”, “mature forest” and “other” is presented in Figure 11.
Multispectral image analysis of woody vegetation 12
Figure 9: The initial supervised classification of Landsat TM7 (c. 2002) established 10 classes (2 water classes
that were merged, blue). The majority of larger scale oil palm plantations could be highlighted (light green). Cloud
and cloud-shadow is apparent (white/shadow) and other land features represent cleared/grassland/gardens and
village/urban centres (beige, bright pink, pale pink, etc).
Figure 10: The established “mature forest” class based on supervised classification of Landsat TM7 January 2002
(green), and an overlay showing the visual interpretation of remnant woody vegetation (black outline, courtesy R.
Storey).
Multispectral image analysis of woody vegetation 13
Figure 11: Supervised classification of Landsat TM7 image (2002) for the three land types of “mature forest”
(green), “oil palm” (red) and all “other” features including cleared/grassland/garden mosaics, lakes, cloud and
cloud shadow (grey).
SPOT4 Imagery
SPOT4 imagery from September 2002 was used for a further supervised classification. Using
ground features identified on the pan-sharpened SPOT4 image (fused with SPOT2
panchromatic image data) seven training classes were established. The training classes
included water bodies, cloud, cloud shadow, large scale oil palm plantations, mature forest,
bare soil and cleared/grassland/gardens. The supervised classification used the maximum
likelihood cluster algorithm with a probability threshold of 95%.
The established classes of “mature forest” and “cleared/grassland/gardens” were extracted
from the image classification and overlayed on Google Earth imagery. Figure 12 shows this
overlay for “mature forest”. The “mature forest” class reasonably delineates the established
or older growth forest fragments on the Popondetta Plain. The “cleared/grassland/gardens”
class reasonably isolated larger patches of open land (Figure 13).
The “oil palm” class covered a large proportion of the study area than the known extent of
Higaturu Oil Palm estate and small holder oil palm plantations (data not shown, courtesy
HOP, D. Mitchell pers. comm., May 2008). Errors of commission and omission were likely.
An overlay of the “oil palm” class on Google Earth failed due to the volume of data when
attempting to convert the classified image into ‘kml’ format. However, the results were
judged to be acceptable by visual comparison with the pattern of oil palm plantation on the
SPOT4 pan-sharpened image (Figure 14). Figure 15 shows the “oil palm” class derived from
the SPOT4 supervised classification overlaying the SPOT4 pan-sharpened image. The
resulting SPOT4 classified image distinguished three elements of land cover as “mature
forest”, “oil palm” and “other” features including cleared/grassland/garden mosaics,
lakes/water, cloud and cloud shadows (Figure 16).
The “mature forest” class derived from the supervised classification of SPOT4 imagery
(September 2002) was compared with the visual delineation of remnant vegetation (R.
Storey, Figure 6) based on Landsat TM7 imagery (January 2002). The total extent of
“mature forest” class was visually estimated to be larger than the extent of remnant
vegetation in the corresponding visually-delineated clusters (Figure 17).
Multispectral image analysis of woody vegetation 14
Figure 12: Extracted “mature forest” class (SPOT4, 2002, red outlines, supervised classification) overlayed on
Google Earth imagery; established clusters follow that generally found for mature forest distribution patterns.
Figure 13: Open land (pattern in background of Google Earth imagery) overlayed with a cleared/grassland class
(SPOT4, 2002, red outlines derived from supervised classification).
Multispectral image analysis of woody vegetation 15
Figure 14: The pan sharpened image of SPOT4 image (fused with pan chromatic SPOT2 band, 10m) used for
visual comparison with the SPOT4 image classifications to support decisions about merging clusters.
Figure 15: Oil palm classification (supervised, beige) overlayed on pan sharpened SPOT4 image (2002) (fused
with pan chromatic SPOT2 band, 10m). A large proportion of the study area was classified as oil palm; errors of
commission are likely.
Multispectral image analysis of woody vegetation 16
Figure 16: Supervised classification of SPOT4 image (2002) for the three land types of mature forest (green), oil
palm (red) and all other features including cleared/grassland/garden mosaics, lakes, cloud and cloud shadow
(grey).
Figure 17: Supervised classification of “mature forest” derived from SPOT4 (September 2002) (green) overlayed
with the visual interpretation of remnant woody vegetation (black outline, courtesy R. Storey).
Multispectral image analysis of woody vegetation 17
Accuracy Assessment
Ground validation sites were collected to assess the accuracy of supervised classifications.
In each case (Landsat TM7, SPOT4) the classes of “oil palm”, “mature forest” and “other”
were assessed. The class “other” resulted from a class merge of all spectral classes not
representing “oil palm” or “mature forest”. Google Earth imagery was used as a source of
ground validation data. A random stratified sampling of Google Earth imagery was
conducted to generate 30 point locations within each “other”, “oil palm” and “mature forest”
class. Only map classification areas (polygons) larger than one hectare were considered in
the random selection process. Within each selected area one location was selected
randomly representing a sample point. Established sample points (reference locations) were
overlayed on Google Earth image data and assigned a ground feature class through visual
inspection. User’s, producer’s and an overall accuracy were calculated for each image
classification. The user’s accuracy represents the reliability of an output map derived from a
classification. It informs the user of the map what percentage of a class corresponds to the
ground-truth class. The producer’s accuracy indicates the percentage of a ground feature
class that was classified correctly.
Landsat Imagery
An accuracy assessment was undertaken employing the Landsat TM7 (January 2002)
supervised classification results shown in Figure 11. A random stratified sampling generated
30 point locations within each of the three classes “oil palm”, “matured forest” and “other”.
Figure 18 shows the distribution of randomly stratified sample points within the study area,
overlaid on Google Earth imagery. The accuracy matrix is presented in Table 1 for the
classes of “oil palm” and “mature forest” and “other”.
The overall assessment of accuracy was 67% (41/61), neglecting 32% of the sample point
locations that visually were not identifiable through the Google Earth imagery: fourteen oil
palm, seven mature forest and eight other point observations could not be assigned to a
ground features class (see for example, Figure 19). The sample point data located on lower
resolution Google Earth imagery (probably Landsat TM) were often not clearly identifiable;
especially locations in the spectral class “oil palm”.
Seven oil palm ground observations were incorrectly included (error of commission) in the
spectral class “other”, and six mature forest observations were incorrectly excluded (error of
omission) (Table 1).
Table 1: Accuracy matrix for the Landsat TM7 (January 2002) supervised classification, the non oil palm and non-
mature forest features are accumulated in the class “other”. The overall accuracy was 67%, but 32% of sample
points could not be identified on Google Earth imagery and are therefore not part of this confusion matrix.
Observation
Classification
Oil Palm
Matured
Forest
Other
Row total
(% user’s
accuracy)
Oil Palm
7
2
7
16 (43%)
Matured Forest
1
21
1
23 (91%)
Other
3
6
13
22 (59%)
Column total
(% producer’s
accuracy)
11
(63%)
29
(72%)
21
(61%)
61
The overall assessment of accuracy of the Landsat TM7 supervised classification is
reasonable at 67% (based on 68% of the ground truth sample data). The accuracy
assessment may be improved with properly field verified, ground truth sample data. This is
because errors of assignment to ground observation classes based on interpretation of
Google Earth imagery are possible and may have reduced the reported accuracy of the
Multispectral image analysis of woody vegetation 18
supervised classification. However, as earlier demonstrated, a broad conformity of spatial
patterns was highlighted through the overlay of visually delineated remnant forest and the
“mature forest” class based on the supervised classification of Landsat TM imagery
(Figure 10).
Figure 18: Location of sample points generated for assessing the accuracy of the supervised classification based
on Landsat imagery (2002): 90 sample points were overlaid on Google Earth image data. A random stratified
sampling method was applied, generating 30 points in each spectral class defined as “oil palm”, “mature forest”
and “other” (see Figure 11). Each point location was to the respective ground feature classes by visual
inspection: oil palm (1), matured forest (4), and remaining features as “other” (0). At least a third of point locations
could not be identified due to low resolution imagery provided by Google Earth, represented by the image mosaic.
Figure 19: Example of
ground observation detection
problems for point locations
overlaid on Google Earth
imagery. Observation data
were generated through
visual interpretation, and are
denoted “ok” for correct
classification and “?” for
locations that were not
identifiable and could not be
assigned a ground class.
SPOT4 Imagery
An accuracy assessment was undertaken employing the SPOT4 supervised classification
results shown in Figure 16. The same method was applied as for the assessment employing
Landsat TM7 data. Random stratified sampling generated 30 point locations within each of
the three classes “oil palm”, “matured forest” and “other”, similar but not identical to that
shown in Figure 18 overlaid on Google Earth imagery. Figure 18 and 19 show the
distribution of sample points for the Landsat TM7 supervised classification evaluation.
The accuracy matrix is presented in Table 2 for the classes of “oil palm” and “mature forest”
and “other”. The overall assessment of accuracy was 53% (26/49); neglecting 44% of the
sample points that were not visually identifiable as to what ground feature class they
belonged to (see for example, Figure 20). Seventeen oil palm and 13 mature forest locations
and 10 points of the merged class “other” were not identifiable based on the validation
source Google Earth imagery.
Table 2: Accuracy matrix for the SPOT4 (September 2002) supervised classification, the non-oil palm and non-
mature forest features are accumulated in the class “other”. The overall accuracy was 53%, but 44% of sample
points could not be identified on Google Earth imagery and are therefore not part of this confusion matrix.
Observation
Classification
Oil Palm
Mature
Forest
Other
Row Total
(% user’s
accuracy)
Oil Palm
4
6
3
13 (31%)
Mature Forest
0
10
7
17 (59%)
Other
4
3
12
19 (63%)
Column Total
(% producer’s
accuracy)
8 (50%)
19 (53%)
22 (55%)
49
Figure 20: Ground validation of a point classified as oil palm based on the SPOT4 supervised classification.
Google Earth imagery did not allow confident ground verification especially in regions of low spatial resolution (left
side, Landsat TM).
Multispectral image analysis of woody vegetation 20
The accuracy assessment identified six mature forest locations from Google Earth images
that were classified as “oil palm”. This represents an error of commission, where mature
forest ground locations were incorrectly classified as “oil palm”. The “mature forest” class
was incorrectly included in seven features of the Google Earth ground representation class
“other”, but none of the oil palm ground features.
An additional difficulty for validating image classifications was cloud occurrence in the
Google Earth data used to ground truth and the time lag between image data acquisitions.
Classification data were acquired in 2002 whereas Google Earth imagery was undated but
probably 2 or 3 years older then the date of access (21st December 2007). The change of
land use in the region is rapid and oil palm farming on broad and small scales is expanding,
generating a stable income for the farming community (D. Mitchell pers. comm. 2007). This
accuracy assessment was limited because it was based on a comparison of two land cover
snapshots (SPOT4 2002, Google Earth undated) that differ in time considerably.
The overall accuracy of the SPOT4 supervised classification was low, which may in part be
attributed to the use of unknown Google Earth image data as a source of ground truth data.
However, as earlier demonstrated, a broad conformity of spatial patterns was highlighted
through the overlay of mature forest delineations in polygon form on Google Earth imagery
(Figure 12). This visual inspection showed that the broad spatial patterns of mature forest
areas on Google Earth were reasonably well described in comparison with the derived
supervised classification outlines of the “mature forest” class.
Unsupervised classification
The unsupervised classification method only employs the statistical properties of image data
to define classification clusters and does not employ the user’s input of establishing training
samples of known identity. A clustering algorithm divides the pixels into a number of
homogenous groups based on their reflectance characteristics (Jensen 2005). The
maximum number of clusters and their possible variability is user-defined and often
determined by iteration and comparison with ground data.
The unsupervised classifications presented here were based on the Iterative Self-Organizing
Data Analysis (ISODATA) algorithm, an iterative process for spectral clustering until either no
significant change in cluster statistics or a given threshold is reached (Lillesand and Kiefer,
2002). The threshold of unchanged cluster establishment was set to 95% and the maximum
number of iterations to six. The number of desired classes was user-defined by iteration.
Landsat Imagery
An unsupervised classification using the Landsat TM7 Bands 1, 2, 3, 4, 5 and 7 was
conducted. Several image classification attempts were iteratively conducted producing 10,
15, 20 and 25 spectral classes (e.g., Figure 21). Each result was visually compared with true
colour Landsat TM7 2002 and SPOT4 2002 imagery pan sharpened with a SPOT2
panchromatic band. The visual comparisons of the unsupervised classifications were not
encouraging.
An accuracy assessment would help validate the unsupervised classifications. This requires
detailed ground information, which was not available except by estimation using Google
Earth imagery. However, as the supervised classifications were considered of superior
accuracy, the unsupervised Landsat TM7 classifications were not developed further;
accuracy assessment was not attempted.
Aerial view of typical land use mosaic
on Popondetta Plain: oil palm / gardens
(left, photo D.K. Mitchell), and remnant
forest / gardens (right, image from
Google Earth)
Multispectral image analysis of woody vegetation 21
Figure 21: Unsupervised classification of Landsat TM7 (January 2002) establishing 10 spectral classes (three
vegetation clusters, four cloud/haze classes, one shadow class and two water classes).
SPOT4 Imagery
Satisfactory results were achieved with the unsupervised classification method ISODATA.
The unsupervised classification employed SPOT4 data of 3 multispectral bands (Bands 1, 2
and 3) as well as the short-wave infrared band (Band 4). All bands had a spatial resolution of
20m. Similar to classifications derived from Landsat data, the classification process based
on SPOT4 imagery identified oil palm plantations, forests, grassland, garden culture and
remnant vegetation at the point in time of image acquisition (September 2002).
The higher spatial resolution and different time of SPOT4 image acquisition resulted in
different classes to those derived from Landsat TM7 imagery (January 2002). The SPOT4
classification represented oil palm cultivations according to their various growth stages.
Acceptable results were achieved by the unsupervised classification producing 25 classes
(maximum likelihood, probability threshold 95%), shown in Figure 22.
Visual comparison with the SPOT4 pan-sharpened image (Figure 14) resulted in three
classes of the unsupervised classification being identified as oil palm plantation in different
growth stages. These three clusters of the SPOT4 unsupervised classification were merged
into one “oil palm” class. The merged class was overlain on the pan sharpened image and
visually judged to acceptably represent oil palm (Figure 23). As for the unsupervised
classification of Landsat imagery, the supervised classification of SPOT4 imagery was
considered superior to the unsupervised classification. Therefore the unsupervised
classification of SPOT4 imagery was not developed further and accuracy assessment was
not attempted.
Multispectral image analysis of woody vegetation 22
Figure 22: Unsupervised classification of SPOT4 (c. 2002) establishing 25 classes. High spectral variability
amongst land cover features is obvious. The highly complex spatial pattern is based on a mixture of plantations,
mixed garden cultures and grassland/clearings.
Figure 23: Oil palm classification (unsupervised, brown) overlayed on pan sharpened SPOT4 image (2002) (fused
with pan chromatic SPOT2 band, 10m). Oil Palm plantations in various growth stages were identified and 3
classes were merged into one based on visual comparison with the SPOT4 pansharpened image (see Figure 14).
Multispectral image analysis of woody vegetation 23
Woody remnant vegetation on the Popondetta Plain
A key objective of this work was to highlight areas of relatively intact and remnant vegetation
on the Popondetta Plain that is potential habitat for the Queen Alexandra Birdwing Butterfly.
The “mature forest” results of the remote sensing classifications were reviewed and a
method for integrating the best available outputs was developed.
The visual comparisons with Google Earth imagery and pan sharpened SPOT4 and Landsat
TM7 imagery determined that ground features were a better match with the results of
supervised classifications than were the case for unsupervised classifications. The
unsupervised classifications (Figures 21 and 22) did not satisfy ground feature
representations that were visually identified on pan sharpened SPOT4 and Landsat TM7
imagery as well as Google Earth. The generated classes did not distinguish ground feature
classes consistently and often showed errors of commission. The supervised classifications
of “mature forest” (Figures 10 and 17) were therefore adopted for further development.
The supervised classifications of “mature forest” were initially compared with the visually
delineated remnant vegetation class and then integrated to determine areas of geographic
concordance.
The “mature forest” clusters based on the Landsat TM supervised classification captured
additional spatial variability than could be mapped using visually depicted classes (Figure
24). The automated classification provided far more detail on the distribution of “mature
forest” canopy. Larger patches of open land within the visually depicted forest fragments
could be interpreted as disturbance activities related to land use such as logging, small scale
oil palm plantations or subsistence agriculture (Figure 24). Logging activities could be further
interpreted by comparison with active forest concession areas. The “mature forest” clusters
based on the SPOT4 supervised classification were similarly interpreted (Figure 25).
Figure 24: Comparison of supervised classification results of “mature forest” class (green pixels) based on
Landsat TM7 2002 and visual classifications of remnant vegetation cover using the same image data pan-
sharpened (black polygon outlines, courtesy R. Storey). The visual mapping coarsely delineated aggregate areas
of woody vegetation applicable at a scale of 1:100,000. Additional spatial variability of spectral feature
representations of woody vegetation is captured by the supervised classification, highlighting degrees of
disturbance within the forest fragments. Patches of open land (white) within the visual clusters (black outlines)
could be interpreted as disturbance attributed to human land use activities.
Multispectral image analysis of woody vegetation 24
Figure 25: Supervised classification results of “mature forest” based on SPOT4 (yellow pixels) compared with the
visual delineation of remnant vegetation (black polygon outlines, courtesy R. Storey). The visual mapping
coarsely delineated aggregate areas of woody vegetation applicable at a scale of 1:100,000. Additional spatial
variability of spectral feature representations of woody vegetation is captured by the supervised classification,
highlighting degrees of disturbance within the forest fragments. Patches of open land (beige) within the visual
clusters (black outlines) could be interpreted as disturbance attributed to human land use activities.
A final dataset to support habitat identification for the Queen Alexandra Birdwing Butterfly
was created by integrating the two supervised classifications of “mature forest” derived from
Landsat TM7 (Figure 10) and SPOT4 (Figure 17) imagery (Appendix 1). The integrated
dataset was derived as the union of both “mature forest” results and smoothed with a
convolution filter (kernel size of 3x3 cells equal to 85.5 x 85.5 m). The final raster data has a
resolution of 28.5m.
The combined “mature forest” classes from either Landsat TM7 or SPOT4 data resulted in
areas where they both mapped the forest class and areas where one or other mapped the
forest class. This occurred for various reasons such as differences in the image extents
(Figure 1) and variation in cloud cover. These three intersection categories of “mature forest”
were retained in the final dataset and are shown as separate features in Figure 26. Woody
remnant vegetation is confirmed by areas identified as “mature forest” class in both the
Landsat and SPOT4 classifications.
Cloud occurrence in remotely sensed imagery often affects land cover classification results.
Cloud free information of either image could therefore be used to substitute missing
reflectance data of ground features in the alternate image. The three categories of “mature
forest” class derived from the union of the two supervised classifications represent a potential
maximum extent of woody vegetation on the Popondetta Plain, with some errors of omission
and commission likely (viz. accuracy assessment Tables 1 and 2).
Relatively undisturbed primary and secondary mature forest are considered breeding habitat
for the Queen Alexandra Birdwing Butterfly (e.g. Parsons 1992). These areas of habitat may
be inferred from the integrated image classification (Figure 26) to be the larger, more
continuous patches of forest vegetation (orange) that also occur within the visually delineated
patches of remnant vegetation. When combined with information about the habitat
preference of the larval food plant (Pararastolochia vine), for example, this information might
be useful when designing surveys to detect the presence of the butterfly.
Multispectral image analysis of woody vegetation 25
Figure 26: Integrated classification results for the “mature forest” class derived from Landsat TM7 and SPOT4
data. Woody remnant vegetation is confirmed by areas identified through both Landsat and SPOT4 data (orange),
and probable for areas based on Landsat data only (blue), areas based on SPOT4 data only (green). Visual
remnant vegetation cover delineation (courtesy R. Storey) is shown as black polygon outlines. Note that some
areas of SPOT4 data only occur where cloud is obscured on the Landsat image.
Figure 27: Woody remnant vegetation results derived from the UNION of Landsat and SPOT4 supervised
classifications overlayed with the areas of vegetation defined by the FIMs dataset (c. 1975; Hammermaster and
Saunders 1995) (brown as undisturbed, purple as slight disturbance, orange as heavily disturbed); A larger
proportion of the woody vegetation class in 2002 is assigned to areas dominated by land use in 1975 comprising
mixed garden/small scale oil palm and fallow mosaics and the slopes of Mt Lamington (lower left) that were
affected by a volcanic eruption in 1951 (white background).
Multispectral image analysis of woody vegetation 26
The integrated classification results were overlayed on the PNG forest vegetation dataset
which delineates areas of forest in 1975 (Figure 27). Most of the classified woody remnant
vegetation area in 2002 was previously assigned undisturbed in 1975. A large extent of the
woody vegetation class also falls within the land use classes of the PNG forest vegetation
dataset (generally indicative of mixed garden/small scale oil palm and garden/fallow
vegetation mosaics). This region was affected by a volcanic eruption in 1951 and may have
regrown further since 1975.
A vector product depicting the “mature forest” class in 2002 was generated. Polygons equal
to the size of one pixel in the integrated data (28.5 x 28.5 m) were eliminated.. This post-
processing substantially reduced the size and complexity of the dataset. The vector dataset
shows clusters of “mature forest” and “other” non-forest areas (merging the land use classes
of oil palm and mosaic cleared/grassland/gardens).
Conclusions
Remote sensing imagery classification techniques were explored as ways to map current
land use and vegetation cover on the Popondetta Plain. Supervised and unsupervised
classifications were applied to Landsat TM7 and SPOT4 imagery. We found the supervised
classifications to be superior in representing ground feature classes of forest cover and land
use (oil palm, other clearings), mainly validated with Google Earth imagery.
Delineations of “mature forest” class derived from SPOT4 supervised classification followed
reasonably well the identifiable pattern of mature forest areas on Google Earth (Figure 11).
Broad mature forest patterns outlined from the integrated supervised classifications could
therefore indicate potential habitat areas for the Queen Alexandra Birdwing Butterfly (Figure
23). Also the “oil palm” class was a reasonable match with pan sharpened features identified
to be oil palm from SPOT4 data (Figure 14), providing additional information about the
boundaries between “oil palm” and “mature forest” on the Popondetta Plain. The ground
validation accuracy assessment of the Landsat TM7 supervised image classification was
acceptable at 67% (excluding 32% of the sample universe, confusion matrix, Table 1);
however the accuracy assessment of the SPOT4 supervised image classification was
relatively low overall at 53% (excluding 44% of the sample universe, Table 2).
One major issue for the accuracy assessment was the lack of known reference points in the
field (ground truth) and the form of their representations (oil palm, mature forest or “other”).
No detailed ground data were available and validations of the automated classification relied
on rather coarse visual interpretations of Google Earth imagery to deliver reference points for
comparison. The classification outputs however showed a fair conformity with the visually
identified ground features on Google Earth.
Classifications which produced a high number of classes were improved through post-
classification (merging classes describing similar features, e.g. cloud, cloud shadow and
haze; sea and fresh water). However, some clusters could not differentiate between land
cover types that were considered to be distinguished (e.g. young oil palm growth/young
natural forest regrowth). Landscape fragments and mosaics in most cases could be visually
detected based on their appearance as objects (spatial patterns), but delineations were
rather coarse. In conclusion, both approaches, visual image inspection and automated
image classification, complement one another.
A rapid-assessment of land cover classes by visual delineation was undertaken using true
colour Landsat imagery to produce a course-scale representation of remnant forest
boundaries to support habitat delineations for Key Biodiversity Areas definition (R. Storey
pers. comm. November 2007, Figure 6). The spatial pattern and texture of spectral
differences among various land uses was generally helpful when visually delineating land
cover classes. Oil palm plantations could be delineated with their various growth stages;
however a differentiation within finer heterogeneous land cover patterns (e.g. garden culture)
was rather difficult.
Both supervised and unsupervised classifications appeared to achieve a better separation of
these land cover classes, because the classification was applied at the scale of data
Multispectral image analysis of woody vegetation 27
resolution. However, differentiation between forest and other remnant vegetation was not
easy to establish, even with training classes. In contrast to visual classification, the
delineation of spatial patterns derived from image classifications did not support image object
identification because the spectral classification methods lack object oriented segregation
approaches. Object-oriented classification is a relatively new development that has been
applied to multi resolution imagery (e.g. Hubert-Moy et al. undated; Desclée et al. 2006,
Mallinis et al. 2006). For example, Wahid et al. (2005) used a combination of visual
inspection, clustering according to spatial patterns and cluster distributions, and supervised
classification which allowed the isolation of an oil palm class. In Wahid’s example the object
definitions were based on visual interpretation of ground feature representations, and could
have been complemented with the conventional classification results. The use of automated
spatial pattern recognition in conjunction with spectral classifiers results has been adopted in
recent years and proved to support a discrimination of spectrally less separable classes.
Ornithoptera
alexandrae ♂,
Voivoro Butterfly
Reserve, May
2007
(Photo courtesy
D.K. Mitchell)
Multispectral image analysis of woody vegetation 28
Objective 2: Remote Detection of Disturbance among Mapped
Remnants of Woody Vegetation on the Popondetta Plain based on
Landsat imagery
Introduction
In Objective 1, image classification approaches were used to distinguish land cover and land
use classes on the Popondetta Plain in 2002 such as mature forest, oil palm, and other
cleared/grassland/garden mosaics. The supervised classifications of SPOT4 and Landsat
TM7 each derived a “mature forest” class that were combined into a single layer representing
‘remnant woody vegetation’ (Figure 26). This layer was comparable with the visual
delineation of remnant vegetation at a coarser scale by R. Storey (Figure 6) derived from
pan-sharpened Landsat TM (January 2002). The combination of visual delineation and
image classification methods provides supporting information about the possible distribution
of relatively intact woody vegetation, but provides no additional information about the relative
condition or disturbance of that vegetation. Relatively undisturbed primary and secondary
mature forest are considered breeding habitat for the Queen Alexandra Birdwing Butterfly
(e.g. Parsons 1992). Other areas of remnant vegetation of smaller patch sizes within a
matrix of other land uses are likely to be disturbed, and may no longer be suitable breeding
habitat for the butterfly. They may however be important elements of landscape connectivity
providing refuge and nectar for dispersing butterflies. Further information about the relative
condition of delineated “mature forest” areas was therefore needed to support the field
identification of potential habitat for the Queen Alexandra Birdwing Butterfly.
Remote sensing image classification methods were therefore investigated further to
determine whether additional information about relative disturbance in the larger continuous
areas of remnant vegetation could be extracted within the visually delineated “mature forest”
class. Primary and secondary forest is represented by woody vegetation that is mature
forest or generally older growth. These areas were considered to occur mainly within the
broadly delineated areas of woody remnant vegetation cover (Figure 26).
Objective 2 reports on this investigation of remote sensing imagery and analysis techniques
for detecting land cover change potentially attributed to human disturbance. The land cover
change detection employed Landsat TM7 data over the time span 1989 and 2002. Image
manipulation (NDVI calculation) and differencing were conducted to test for detectable land
use change (deforestation, oil palm plantation). Spectral mixture analysis was also explored.
Image data for classifications and change detection
The following investigations were applied to Landsat TM7 2002 and Landsat TM5 1989
image data (spatial resolution of 28.5m). The true colour Landsat Images for 1989 and 2002
in Figures 28 and 29 provide a visual representation of the data upon which the analyses
were based.
In the first instance, the visual classification of remnant vegetation (Figure 6, courtesy R.
Storey) was used as a mask over the Landsat TM7 (2002) image to focus attention on the
spectral reflectance characteristics of the delineated areas of “remnant vegetation” only
(Figure 30). The remnant vegetation areas were then further assessed with regard to plant
vigour and forest disturbance. An unsupervised classification of the spectral reflectance
characteristics was used to test the potential to distinguish areas of primary and secondary
forest and other degraded areas. The Normalised Difference Vegetation Index (NDVI,
derived from Bands 3 and 4) was used to assess plant vigour in relation to areas of forest
disturbance/degradation. Finally, change detection using near infra red (NIR) Band 4
between 1989 and 2002 was explored as a way of identifying remnant areas of relatively
intact primary and secondary vegetation.
Subsequently, supervised (Landsat TM7 2002) and unsupervised (Landsat TM5 1989)
classifications based on all available bands were conducted to provide benchmark
classifications of “mature forest” as a further approach of change detection.
Multispectral image analysis of woody vegetation 29
Figure 28: The Landsat TM5 true colour image (July 1989) for the Popondetta Plain.
Figure 29: The Landsat TM7 true colour image (January 2002) for the Popondetta Plain.
Multispectral image analysis of woody vegetation 30
Figure 30: The Landsat TM7 image (January 2002, Bands 5, 4, and 1) masked to emphasise the spectral
reflectance characteristics of the delineated areas of remnant vegetation only. This masked data was used in the
unsupervised classification to test the potential to distinguish areas of primary and secondary forest and other
degraded areas.
Unsupervised Classification of visually delineated woody remnant vegetation
An unsupervised classification was used to analyse the spectral reflectance characteristics of
the Landsat TM7 imagery (January 2002) within the delineated areas of woody remnant
vegetation (as shown in Figure 30). Although supervised classification is preferred, an
unsupervised classification was undertaken in lieu of detailed field information or user
knowledge to discriminate areas of probable primary and secondary forest and other
degraded vegetation on Google Earth imagery.
The unsupervised classification was conducted using the IsoData algorithm with a
convergence threshold of 95% and a specified outcome of 7 spectral classes. Five classes
could be identified, three of which were cloud, cloud shade and haze. The remaining two
classes were related to vegetation features. By reference to the true colour Landsat image, it
could be visually estimated that one class was describing woody vegetation and the other
class was mainly related to grassland or drier, less active vegetative features. However,
casual ground-truth comparisons with Google Earth’s more current imagery (undated,
accessed 21st Dec. 2007) were not encouraging. The accuracy of this classification could
not be confirmed in all examined sample areas (Figures 31 and 32). Therefore it was
concluded that a simple unsupervised classification of the masked imagery into primary and
secondary forest and other degraded vegetation could not be made.
Local knowledge of primary, secondary and degraded forest in the study area is required in
order to establish a supervised classification of the spectral characteristics that could be
related to potential habitat for the Queen Alexandra Birdwing butterfly. Rather than a
masked, the full image would be used for a supervised classification.
Multispectral image analysis of woody vegetation 31
Figure 31: Classification results could not be confirmed with Google Earth imagery (Example 1): Woody
vegetation discriminated into two classes through unsupervised classification (right), compared to Landsat Bands
4,5,3 as RGB (middle), and Google Earth Image (left).
Figure 32: Classification results could not be confirmed with Google Earth image data (Example 2): Woody
vegetation discriminated into two classes through unsupervised classification (right), compared to Landsat Bands
4,5,3 as RGB (middle), and Google Earth Image (left).
Assessing plant biophysical characteristics through vegetative indices (NDVI) and NIR
within visually delineated woody remnant Vegetation
NDVI characteristics of the Landsat TM7 imagery (January 2002) were investigated as a
possible way to discriminate areas of probable primary and secondary forest and other
degraded vegetation within the delineated areas of woody remnant vegetation (as shown in
Figure 30).
NDVI
1
represents a measure of photosynthetic capacity and plant vigour for above ground
green biomass (Sellers 1985, Myneni et al. 1995). The index derives from Landsat Bands 3
(Red) and 4 (Near Infra red, NIR).
The reflectance properties of the masked imagery (Figure 30) showed NDVI values between
-0.79 and 0.64. Values larger or equal to 0 are related to the spectral response of
photosynthetically active vegetation (Lillesand and Kiefer 2004). Since the NDVI is strongly
correlated to green biomass productivity and leaf chlorophyll content, the remnant vegetation
class could be further distinguished into areas of lower, medium and higher productivity
1
NDVI, Normalized Difference Vegetation Index. NDVI represents a measure of density of chlorophyll
contained in vegetative cover. It is calculated from the visible and near-infrared light reflected by
vegetation. Healthy vegetation absorbs most of the visible light, and reflects a large portion of the
near-infrared light (higher to high NDVI values). Unhealthy or sparse vegetation reflects more visible
light and less near-infrared light (lower NDVI values). See:
http://earthobservatory.nasa.gov/Library/MeasuringVegetation/measuring_vegetation_2.html
Multispectral image analysis of woody vegetation 32
(Jensen 2005). It could be observed that homogenous mature forest areas with closed
canopies showed NDVI values below 0.5 (Figure 33), often ranging between 0.3 and 0.42
(NDVI class mean ~0.38).
Forming part of the visually delineated remnant vegetation class (Figure 6), younger growth
and regions under anthropogenic influences such as oil palm plantings and other forms of
mixed garden culture were in general characterized by higher NDVI values between 0.5 and
0.6. Also, cleared “other” areas with young ground vegetation or cultivation within the forest
fragments contributed to high NIR reflectance. These various stages of vegetation growth
and spectrally inhomogeneous soil background (wetlands, soil moisture) accounted for an
increased NDVI variability (Figure 34). The forest patch shown in Figure 34 was categorized
as heavily disturbed in 1975 according to the PNG forest vegetation dataset (Hammermaster
and Saunders 1995).
In order to test the hypothesis that NDVI might be distinguishing open forest (disturbed
woody vegetation) and closed forest (primary/secondary forest) classes, zonal statistics were
evaluated in a few areas of relatively homogenous vegetation. The choice of vegetation
patches for this comparison was based on visual inspection of the visually delineated map of
remnant vegetation (Figure 6) overlayed on the true colour Landsat TM7 image and Google
Earth.
The resulting zonal statistics for both open and closed forest canopy fragments showed that
standard deviation (STD) and ranges of NDVI values for forest patches were both smaller
than for open forest representing disturbed areas (Figure 28 and 29). However, the
differences in NDVI statistics for both forest examples were small (dense forest: STD 0.04,
range: 0.7; disturbed forest: STD 0.09, range: 1.43). Therefore it was concluded that NDVI
may not be distinctive enough for distinguishing between open and closed canopy,
potentially representing primary / secondary and more disturbed forest types.
Figure 33: NDVI values (left) were ranging between 0.3 - 0.42 for a mature forest fragment with dense canopy
(middle: Landsat Bands 4, 5, 3 as RGB, right: Google Earth representation).
Figure 34: Higher NDVI variability for an open/disturbed forest fragment: values range up to 0.57 (left), in
comparison Landsat Bands 4,5,3 as RGB (middle), and Google Earth Image (right).
Distinct reflectance patterns between dense and disturbed forest cover were possible with
Landsat TM7 Band 4 (NIR). Subsequent exploratory analysis revealed that as a measure for
Multispectral image analysis of woody vegetation 33
forest degradation, the Landsat NIR band potentially showed more distinctive character
amongst woody remnant vegetation mappings than the values of NDVI.
The NIR reflectance values of several forest fragments were compared. The forest
fragments for the exploratory comparison were selected by reference to the true colour
Landsat TM7 image and Google Earth imagery. This examination revealed that point
locations of mature/closed forest areas generally showed NIR values below 100 and those of
rather disturbed forest areas were characterized by NIR values above 100 (Figure 35).
Further investigation found that mature forest areas were better characterized by NIR values
between 79 and 100 for the year 2002 (Landsat TM7) and NIR values between 53 and 79 for
the year 1989 (Landsat TM5). The NIR reflectance value of 100 was therefore considered a
threshold value for distinguishing closed canopies from the more open/disturbed forest edge
areas induced by human impact such as agricultural activities (garden matrix / cleared) and
oil palm growth. The spectral response of undisturbed forests is characterized by canopy
shade and – in contrast to regrowth and crops – by moderate photosynthetic activity and
therefore contributed to lower NIR reflectance measures. Figures 36 and 37 show the NIR
values between between 53 and 79 applied to the Landsat 1989 NIR band as well as NIR
values between 79 and 100 for the Landsat 2002 NIR band.
Figure 35: Spectral profile of point locations (red crosshairs on Landsat imagery from 1989 (below left) and 2002
(below right); Band 4 (NIR) reflectance of matured woody vegetation showed values considerably below 100
(graph left); same locations were logged and replanted with oil palm years later (2002; visually identified through
Google Earth imagery) and revealed higher NIR values above 100 (graph right).
Based on the Landsat TM7 (January 2002) NIR band, a vector dataset was generated
describing “relatively undisturbed woody vegetation” (NIR values between 79 and 100),
“disturbed vegetation” (NIR values above 100). NIR values below 100 (often water) together
with an applied cloud/water mask dervived the class “other” (Figure 38). The classes “other”
and “disturbed vegetation” showed errors of commission. The “other” class includes water
and shade, including hill shade, which comprise forested areas. The “disturbed vegetation”
class included areas affected by cloud/haze cover.
This analysis provided information about relative disturbance within the visually delineated
remnant vegetation class which is additional to that derived from the integrated mapping of
“mature forest” shown in Figure 26. However, further confirmatory evidence was needed, for
example through comparison with change detection to test for areas of geographic
concordance.
Multispectral image analysis of woody vegetation 34
Figure 36: NIR values between 79 and 100 (turquoise) derived from the Landsat TM7 Band 4 (January 2002)
estimated to represent closed canopies of relatively undisturbed forest, possibly primary or secondary forest on
the Popondetta Plain, overlaying the false-colour Landsat image (Bands 5, 4, 1) and showing the visually
delineated boundaries of remnant vegetation (black outlines, courtesy R. Storey).
Figure 37: NIR values between 53 and 79 (turquoise) derived from the Landsat TM5 Band 4 (July 1989)
estimated to represent closed canopies of relatively undisturbed forest, possibly primary or secondary forest on
the Popondetta Plain, overlaying the false-colour Landsat image (Bands 5, 4, 1) and showing the visually
delineated boundaries of remnant vegetation (black outlines, courtesy R. Storey).
Multispectral image analysis of woody vegetation 35
Figure 38: Vector data based on Landsat TM7 January 2002 showing ‘relatively undisturbed woody vegetation’
(NIR values between 79 and 100, green), ‘disturbed woody vegetation’ (NIR values above 100, orange), and
‘other’ (coud and water mask as well as NIR values below 79, red). Classifications based on the single NIR band
(Band 4) were sensitive to hill shade (included in class “other”). The classification results are shown within the
visually delineated boundaries of remnant vegetation (black outlines, courtesy R. Storey).
Change detection of “mature forest” based on single bands (Landsat NIR bands,1989
– 2002)
Change detection using NIR band 4 between 1989 and 2002 was explored as a way of
identifying remnant areas of relatively intact primary and secondary vegetation. On the
Popondetta Plain deforested land often triggers rapid regrowth and closed vegetative cover
is established within a growing season (D. Mitchell, pers. comm. May 2007). A cursory
visual inspection of the imagery time series revealed that cleared areas are often under
different cultivations at different times (e.g. oil palm regrowth stage, other subsistence
farming activities). The variety and dynamics of land use on the Popondetta Plain and the
limitations of Google Earth imagery for verification meant that it would not be easy to
describe ground cover changes by spectral response of plant biophysical characteristics.
Variation in NIR between1989 and 2002 were used to detect patches of “mature forest” that
have remained relatively undisturbed over time (Figures 39 and 40). NIR serves as a
measure for spectral response of active vegetation (Jensen, 2005). In the previous analysis,
it was observed that NIR reflectance could depict varying ground conditions (Figure 36).
Undisturbed forests showed NIR reflectance values between 74 and 100 in 2002 (Landsat
TM7 2002) and NIR values between 53 and 79 in 1989 (Landsat TM5). However, regrowth,
oil palm and other agricultural cultivation were detected with NIR values above 100 in 2002
and values above 79 in 1989. These factors of change influencing the NIR reflectance were
highlighted and possibly allowed a delineation of spatial variation of forest ground cover
change.
Multispectral image analysis of woody vegetation 36
Figure 39. The Landsat TM5 NIR Band 4 (July 1989) for the Popondetta Plain.
Figure 40. The Landsat TM7 NIR Band 4 (January 2002) for the Popondetta Plain.
Multispectral image analysis of woody vegetation 37
A change dataset was produced through image differencing based on the Landsat 1989 and
2002 NIR data. The values of the change dataset varied between –255 and +246. The
change dataset was compared with both 1989 and 2002 true colour Landsat images. . The
purpose of this comparison was to identify a threshold value representing the change from
forest to other land use, such as oil palm and garden fallow regrowth, in order to distinguish
the forest class for change detection.
A threshold value of -42 of the difference image was derived from various sampling points
representing a breakpoint for a change from forest to oil palm/other land use of similar higher
NIR reflective features. Values equal or below -42 of the difference image were identified as
potential areas of deforestation. A binary output dataset (level slice) was derived from the
division of the difference image into two classes, one above (1) and the other below (0) or
equal to -42, outlining formerly forested areas (Figures 41, 42).
A cloud, shade and water mask was derived from a composite of the two input datasets
(1989, 2002). This dataset was used to mask out potential errors in the difference image.
The level-slice was subsequently derived from the difference image, highlighting areas of
deforestation. Areas identified as potential deforestation were selected where they occurred
within the 1989 NIR forest mask (i.e. NIR values between 53 and 79 as shown in Figure 37)
but not in the 2002 forest NIR layer (i.e. NIR values between 79 and 100 as shown in Figure
36). Areas of forest regrowth and intact forest areas were derived using same raster overlay
of the NIR band classifications of 1989 and 2002. After applying a majority filter (3x3, i.e.
28.5mx28.5m) and eliminating areas less then two pixels (1624.5 m2), a vector layer was
derived showing areas of potential deforestation, reforestation and intact forest between
1989 and 2002 (Figure 43).
Figure 41: Landsat 1989 (above left) and 2002 (above right) showing a change of vegetative cover over time. Oil
palm growth followed after deforestation, further spread of farming activities were found on Google Earth (below
Multispectral image analysis of woody vegetation 38
left); Highlighted deforestation with oil palm growth replacement in black based on differences of NIR reflectance
(below right).
Figure 42: The resulting difference image derived from 1989 and 2002 Landsat TM band 4 (NIR) for Popondetta
Plain.
Figure 43: Change detection based on Landsat NIR bands 1989 and 2002, employing the difference image, NIR
based forest classes of 1989 and 2002, and a cloud/water mask: Potential deforestation areas between 1989 and
2002 derived from Landsat data (red). Areas of intact forest with closed canopy were identified in as matured
forested areas in 2002 and 1989 (green). Potential forest regrowth areas in 2002 are shown in yellow. All other
areas are defined as “other” (beige).
The potential deforestation areas were overlaid on Google Earth as a source of ground-truth
data. Figure 44 and 45 present examples of areas identified as “mature forest” in 1989
Multispectral image analysis of woody vegetation 39
(using Landsat data NIR values between 53 and 79) that were logged and appear as cleared
and regrowth or cultivated areas in 2002. The results were encouraging and supported the
general premise of change, delineating areas of “mature forest” that have been deforested
since 1989 contrasting with areas which were still intact by 2002.
Figure 44: Formerly forested areas in red outlines (1989, Landsat TM5 derived) are overlaid on Google Earth
imagery (undated, possibly 2-3 years prior to 2007).
Figure 45: Formerly forested areas in red outlines (1989, Landsat TM5 derived) are overlaid on Google Earth
imagery (undated, possibly 2-3 years prior to 2007).
Multispectral image analysis of woody vegetation 40
Change detection of woody vegetation employing Landsat data (1989 – 2002)
A further change detection based on classification of Landsat image data for the years 1989
and 2002 was explored as a way of confirming areas of relatively intact primary and
secondary vegetation. A supervised classification of 2002 Landsat TM7 image data (Bands
1, 2, 3, 4, 5 and 7) was previously conducted (Figure 10). An unsupervised classification
using the 1989 Landsat TM5 image data (Bands 1, 2, 3, 4, 5 and 7) was conducted for land
cover classification. Both data sets were then used detecting change in forest cover to 2002.
The unsupervised classification was conducted using the IsoData algorithm with a
convergence threshold of 95%. The maximum number of iterations was set to six and the
amount of spectral classes to twelve. On the base of visual inspection using true colour
Landsat TM5 (1989), pan-sharpened Landsat TM5 (1989) and Google Earth data, four
spectral classes were merged to a forest class, approximating matured forest areas on the
ground as of 1989 (Figure 46).
Ground validation employed the 30 sample points derived for the Landsat 2002 supervised
classification (Table 1), 23 of which were identified as forest on Google Earth imagery.
Seven points were not identifiable on Google Earth. Fourteen of the 23 ground truth points,
identified as forest on Google Earth and classed as “mature forest” in 2002, were accurately
classed by the unsupervised classification of Landsat 1989 data. Nine forest point locations
from Google Earth imagery and the “mature forest” class in 2002 were inaccurately classed
as non-forest in 1989, resulting in an overall accuracy of 61%.
Employing the supervised classification outcomes of Landsat TM7 data 2002 (Figure 10) a
comparison of forested and deforested areas between 1989 and 2002 was conducted
through image differencing (Figure 47). The results highlight forest areas that were intact
over a 13 year period from 1989 to 2002 and other areas which have been disturbed. The
image data were applied through a cloud, shade and water mask and a post-classification
merge of ground feature clusters apart from mature forest into one class (“other”).
Multispectral image analysis of woody vegetation 41
Figure 46: Unsupervised classification of Landsat data TM5 (July 1989) establishing 12 spectral classes (four
spectral classes were merged into one forest class in light green colour).
Figure 47: Image differencing result of Landsat data 1989 and 2002; forested areas intact in both 1989 and 2002
(green), areas deforested between 1989 and 2002 (red), areas reforested between 1989 and 2002 (yellow), areas
representing all “other” ground features in 1989 and 2002 (brown), areas showing cloud, shadow and water from
1989 and 2002 (grey and dark grey).
A vector product depicting forest change between 1989 and 2002 as shown in Figure 47 was
generated. Polygons equal to the size of one pixel in the Landsat TM7 data (28.5 x 28.5 m)
were eliminated by merging with their neighbour sharing the longest boundary. This post-
processing substantially reduced the size and complexity of the dataset. The vector dataset
is a derivative of the raster calculation of the unsupervised (1989) and supervised
classification (2002) of Landsat data showing clusters of deforested, reforested and
untouched (intact) forest areas, as well as areas that were not forested in 1989 or 2002
(Figure 48). The results included some error due to misregistration of the two Landsat
datasets (e.g. Figure 49).
Misregistration error
The results of change detection were affected by a misregistration error derived from the two
Landsat image scenes. The root mean squared error (RMSE) calculated for the geometric
offset between the Landsat images 1989 and 2002 was 139.7m. The RMSE was derived
from ten sample sites where deviation measurements were taken of the same identified point
locations on the Landsat true colour 1989 and 2002 images (Figure 49).
Multispectral image analysis of woody vegetation 42
Figure 48: Vector data set based on supervised (2002) and unsupervised (1989) classification of Landsat data:
Mature forest areas intact in 1989 and 2002 (dark green), areas being deforested (red), regrowth to matured
forest in 2002 since 1989 (light green) and non-forest areas (beige); a cloud mask was applied (white) depicting
composite areas of cloud, shadow and water in 1989 and 2002.
Figure 49: Misregistration error of Landsat imagery: true colour Landsat data (2002, left; 1989, right); the offset in
this example was 149 m, the RMSE for the two Landsat images was 139.7 m based on 10 sample site deviation
measurements.
Spectral mixture modelling deriving forest degradation
A spectral mixture analysis operating on reflectance data was explored as a method for
investigating forest degradation in the Popondetta Plain study area. The inspiration for
investigating the approach came from the outcomes of recent research by Souza et al.
(2003) in deriving a forest degradation map for Eastern Amazon using SPOT4 image data.
Unmixing algorithms were employed to unmix the spectral signatures of forest features.
Spectral mixture analysis mostly uses a linear mixing model, where the mixed spectrum of a
pixel is assumed to be a linear combination of the original spectra present in the pixel,
Multispectral image analysis of woody vegetation 43
weighted by the quantities of their occurrences. According to Souza et al. (2003) the
assumption was made that forested features are composed of a mixture of green vegetation,
non-photosynthetic vegetation (logs, wood), soil (often associated with unpaved dirt roads)
and shade (canopy shade). Spectral signatures of pure pixel representations of these
ground feature classes represented so called “end members”.
Souza et al. (2003) showed that through the application of linear mixing, fraction images
were derived and proportions of these end members could be estimated within the image.
This allowed the definition of the proportion of end members in each pixel, associating a
measure of degradation expressed through the fraction of green vegetation, non-
photosynthetic vegetation and soil. Souza et al. (2003 and 2005) suggested this approach
for tropical forest areas in Brazil using SPOT4 imagery. They were able to establish a forest
degradation classification through further image manipulation and the use of field and
external image data: 1-m high resolution imagery (IKONOS) was used as image fusion data
in conjunction with field data for validating image classifications.
On the Popondetta Plain, SPOT4 imagery data were used to derive a supervised
classification including a “mature forest” class. This “mature forest” class, in conjunction with
pan sharpened and Google Earth imagery as ground data, offered the opportunity to follow
the Souza et al. method of degradation mapping.
This method looks promising, but due to the lack of appropriate software tools (such as
ENVI, http://www.ittvis.com/ENVI/) a linear unmixing approach could not be conducted.
Conclusions
Several approaches to image classification were investigated to identify areas of forest
disturbance that would distinguish potential areas of intact primary or secondary forest.
In the first instance we attempted an unsupervised classification of Landsat TM7 2002 data
within the visually delineated areas of remnant vegetation. Our logic in this application being
to mask areas of land use and focus the unsupervised classification on distinguishing a wider
array of spectral features within the forest mosaic. However, we lacked suitable ground-truth
data to distinguish between forest degradation classes and primary/secondary forest, and
therefore could not apply a supervised classification. We concluded from this exploration
that a simple unsupervised classification of the masked imagery into primary and secondary
forest and other degraded vegetation could not be made. Local knowledge of primary,
secondary and degraded forest in the study area is required in order to establish a landuse
classification that could be related to potential habitat for the Queen Alexandra Birdwing
butterfly. This could be attempted using Google Earth where higher resolution imagery and
local knowledge can reasonably depict the different forest classes. We further conclude that
a supervised classification is best applied to the full image extent, rather than masked to
remnant forest delineations.
We also explored the use of NDVI as a measure of plant vigour and a possible way to
discriminate areas of probable primary and secondary forest and other degraded vegetation,
again applied to the Landsat TM7 2002 image masked by the remnant forest delineations.
Zonal statistics for sample patches of remnant vegetation were used to test the hypothesis
that NDVI might distinguish open forest (disturbed woody vegetation) and closed forest
(primary/secondary forest) classes. The results were inconclusive and we decided not to
progress further with NDVI.
We further explored variation in reflectance patterns between dense and disturbed forest
cover using the Landsat TM7 NIR band (Band 4). Initial examinations suggested the NIR
band showed more distinctive character within woody remnant vegetation than NDVI.
Mature forest areas were found to be reasonably characterised by NIR values between 79
and 100 in 2002 (Landsat TM7) and between 53 and 79 in 1989 (Landsat TM5). The NIR
reflectance value of 100 was therefore considered a threshold value for the boundary
between closed canopies and open/disturbed forest edge areas. The results looked
promising and a vector dataset was generated depicting areas of woody vegetation classified
as relatively undisturbed, disturbed, and other. The classification was applied within the
Multispectral image analysis of woody vegetation 44
visually delineated areas of remnant vegetation (Figure 37). This analysis provided useful
information about relative disturbance additional to that derived from the integrated mapping
of “mature forest” shown in Figure 26. However, further confirmatory information was
needed which we developed through image-differencing and change detection.
Two methods of change detection were explored for the 13 year period between 1989 and
2002 of Landsat imagery. In the first instance, we applied NIR thresholds to distinguish
forest and non-forest areas in 1989 and 2002. Regrowth vegetation, oil palm and other
clearings or agricultural cultivations appeared to correlate with NIR values above 100 in 2002
and above 79 in 1989. A change dataset was produced through image differencing based
on the Landsat 1989 and 2002 NIR data and a threshold value of -42 below which potentially
identified areas of deforestation. The results were encouraging and supported the general
premise of change, delineating areas of “mature forecast” that have been deforested since
1989 contrasting with areas which were still intact in 2002 (Figure 43).
A further change detection was needed to confirm areas of relatively intact primary and
secondary vegetation. The first approach utilised a single spectral band in Landsat – NIR
Band 4 – with image differencing (1989-2002), but 7 spectral bands are potentially available
for classification. We therefore developed a change detection based on comparable
classifications of the 6 spectral bands to define a “mature forest” class in 1989 and 2002.
Although we had generated a reasonable supervised classification of “mature forest” in 2002
(overall accuracy estimated 67% based on 68% of the ground truth sample data, Table 1),
we lacked ground-truth data for a similar classification in 1989. We therefore generated an
unsupervised classification of “mature forest” for the 1989 image and tested accuracy using
ground-truth data identified as mature forest on Google Earth and also classified as such in
2002 (overall accuracy estimated 61%, based on 77% of the ground truth sample data). A
comparison of forested and deforested areas between 1989 and 2002 was conducted
through image differencing (Figure 47). The results highlight forest areas that were intact
over a 13 year period from 1989 to 2002 and other forest areas which have been disturbed
or deforested (Figure 48).
The two maps of potential deforestation areas (Figures 43 and 48) are comparable;
emphasising the same parts of the landscape where forest edges are gradually being
cleared and replaced by other land uses. There are differences however. The image
differencing, based on supervised and unsupervised classification of “mature forest” in 2002
and 1989, appears to include errors of commission around the margins of cloud, shade and
water (Figure 48), evidenced by comparison with the integrated classification of “mature
forest” as at 2002 presented in Figure 26. There appears to be less emphasis on these
errors of commission for the image differencing based on the NIR level slice which
distinguishes forest and non-forest areas in 1989 (Figure 43). The two approaches provide
conclusive evidence of considerable changes in forest cover and incremental disturbance
over a 13 year period from 1989 to 2002 on the Popondetta Plain. However, we were unable
to determine which result provides the most accurate representation of deforestation over
this period.
Our objective was to identify areas of intact primary or secondary forest and to distinguish
areas of forest disturbance. Principally, we aimed to ‘add value’ to the previous delineations
of remnant vegetation (Figure 6) and “mature forest” (Figure 26). Figure 48 is an integrated
output combining the results presented in Figures 11 (supervised classification Landsat
2002) and 46 (unsupervised classification Landsat 1989). Figure 50 shows the pattern of
forest cover and disturbance on the Popondetta Plain compared with the manual delineation
of remnant vegetation cover. Figure 51 shows the same patterns compared with the
integrated classification of “mature forest” based on Spot4 and Landsat TM7 data (shown in
Figure 26). These two maps demonstrate how the change detection has contributed
additional information about potential areas of forest disturbance and clearings within the
delineated boundaries of remnant forest cover.
Not withstanding the context of classification methods, accuracy assessments and the
significant misregistration between the 2002 and 1989 Landsat images (+/- 150 m), the
results presented here provide useful information and guidance for field-based studies to
Multispectral image analysis of woody vegetation 45
identify suitable remnant habitat for the Queen Alexandra Birdwing Butterfly on the
Popondetta Plain.
Figure 50: Forest cover and disturbance on the Popondetta Plain in 2002 compared with the manual delineation
of remnant vegetation (black outline, courtesy R. Storey, shown in Figure 6). Within delineated areas of remnant
vegetation, intact forest (green) potentially represents primary or secondary forest, deforestation (orange) shows
areas of disturbance since 1989, regrowth since 1989 (yellow) is indicative of past and potential continuing
disturbance, and non-forest areas (purple) represent other land uses of varying intensity including oil palm,
grassland, agricultural cultivations and related clearings.
Multispectral image analysis of woody vegetation 46
Figure 51: Forest cover and disturbance on the Popondetta Plain in 2002 compared with the integrated
classification of “mature forest” based on SPOT4 and Landsat TM7 data (based on Figure 26). Within the “mature
forest “ class, intact forest (green) potentially represents primary or secondary forest, deforestation (red) shows
areas of disturbance since 1989, regrowth since 1989 (yellow) is indicative of past and potential continuing
disturbance.
Multispectral image analysis of woody vegetation 47
General Discussion and Conclusions
Taking a “rapid-assessment” approach, we investigated several methods of image
classification and change detection in order to compare and confirm delineations of ground
features or land cover change in the absence of field data. Our principle objective and the
emphasis of this work have been to provide information about native vegetation cover and
disturbance on the Popondetta Plain to inform a conservation strategy for the Queen
Alexandra Birdwing Butterfly.
In Objective 1, we developed a classification of “mature forest” areas derived from Landsat
TM7 and SPOT4 (Figure 26) and compared this with the visual delineation of remnant woody
vegetation by R. Storey (Figure 6). In Objective 2, we developed two classifications of forest
disturbance based on changes in forest cover between 1989 and 2002 using Landsat data
and compared these with both the visual delineation of remnant woody vegetation and the
integrated classification results.
The two objectives resulted in delineated areas of remnant vegetation, within which areas of
forest disturbance and clearings could be distinguished (Figures 50, 51). Forest disturbance
is evident within remnant forest patches that are surrounded by land use and clearings
associated with human activity, and this disturbance extends into contiguous forest areas
where timber harvesting is prevalent.
The integrated classification of woody remnant vegetation (Figure 26) provides some insight
into relative disturbance of habitat potentially suitable for breeding or foraging by the Queen
Alexandra Birdwing Butterfly. Relatively undisturbed primary and secondary mature forest
are considered breeding habitat for the butterfly. These areas of habitat may be inferred
from the integrated image classification to be the larger, more continuous patches of forest
that also occur within the visually delineated remnant vegetation. When combined with
information about the habitat preference of the larval food plant (Pararastolochia vine)
(discussed in Williams et al. 2007), the maps of forest cover and disturbance (Figure 50, 51)
might be useful, for example, when designing surveys to confirm the presence of the
butterfly.
The supervised classifications of Landsat TM7 and SPOT4 based on spectral reflectance of
ground features resulted in a reasonable map of forest cover and land use on the
Popondetta Plain. The classification outputs were mainly validated with Google Earth
imagery and showed a fair conformity with visually-identified ground features.
Major land cover classes in 2002 such as oil palm plantations, small oil palm farms, mixed
garden culture and remnant vegetation could be mapped (Figures 5, 11, 16). However, the
highly dispersed mosaic of land uses (garden matrix/cleared/grassland) could not be
accurately delineated. The remote sensing detection explored here demonstrated that
increased spatial heterogeneity could be captured by classification algorithms (supervised
and unsupervised classifications) but delineations were restricted to variations of spectral
reflectance within a land cover class. The woody land cover mosaics consisted of vegetative
features (e.g. oil palm, forest) in various growth stages.
Land use mapping based on remotely sensed data reflects a state of land cover at a
particular point in time (snapshot). Ground feature or land use classes such as mature
forest, disturbed vegetation and oil palm plantations appeared in various growth stages and
their plant physiological characteristics therefore change over time. Vigorously growing
plants appear different to older and mature growth vegetation due to changing chlorophyll
and other pigment density in their leaves. It was therefore difficult to consistently delineate
all ages of oil palm plantings from natural regrowth within the Popondetta Plain using either
visual delineation or distinct spectral signatures assigned to that ground class.
Thus, corresponding spectral reflectance varied within ground feature clusters highlighting
the need for detailed ground truth information to establish reliable training classes for the
supervised classifications. Due to the rapid change of land use on the Popondetta Plain, the
accuracy of current land use and land cover mapping will depend on the degree of
correspondence between the date of the image data aquisition and ground truth data.
Multispectral image analysis of woody vegetation 48
An accuracy assessment of supervised classifications presented in this report showed low to
reasonable overall accuracy (Tables 1 and 2). Detailed field information and local knowledge
of land use and land cover in the study area are essential for establishing image training
areas and conducting ground validation assessments (Jensen, 2005). Google Earth imagery
provided a guide for assessing classification accuracy but was not sufficient in lieu of ground
truth data. We found that ground features could not confidently be identified using Google
Earth alone because of variable and low image resolution over some parts of the landscape
and obvious time lags between image acquisitions.
Data products
Metadata records for the derived classification products resulting from this work are
presented in Appendix 1. Three vector data products are described:
o Multispectral classification of remnant woody vegetation land cover (Landsat TM7
and SPOT4 2002) on the Popondetta Plain of Oro Province, Papua New Guinea
(based on Figure 26).
o Woody remnant vegetation change (Landsat NIR bands 1989-2002) on the
Popondetta Plain of Oro Province, Papua New Guinea (based on Figure 43).
o Woody remnant vegetation change (Landsat classification 1989-2002) on the
Popondetta Plain of Oro Province, Papua New Guinea (based on Figure 48).
The metadata descriptions also refer to a number of interim data products which are the
original raster classifications or merged classes upon which the vector derivatives are based,
or ground-truth data used in the training of supervised classifications or accuracy
assessments. These raster data products are:
o Landsat TM7 image January 2002
o Landsat TM7 image July 1989
o SPOT4 image September 2002
o Visual delineation of remnant vegetation based on Landsat TM7 pan-sharpened true-
colour image January 2002 (courtesy R. Storey).
o Spatial data containing training sites and their ground feature interpretations used in
generating supervised classifications.
o Supervised classification of Landsat TM7 (January 2002) into 10 classes (based on
Figure 9).
o Three land types – mature forest, oil palm, other – derived from the supervised
classification of Landsat TM7 (January 2002) (based on Figure 11).
o Three land types – mature forest, oil palm, other – derived from the supervised
classification of SPOT4 (September 2002) (based on Figure 16).
o Difference image derived from 1989 and 2002 Landsat TM band 4 (NIR) (based on
Figure 42).
o Potential deforestation areas between 1989 and 2002 derived from Landsat TM5 NIR
and TM7 data (based on Figure 43).
o Unsupervised classification of Landsat TM5 (July 1989) establishing 12 spectral
classes (based on Figure 46).
o Difference image derived from classifications of “mature forest” in 1989 and 2002
Landsat TM (based on Figure 47).
Recommendation and future work
This report demonstrated that detecting woody vegetation changes deriving degrees of
disturbance over time could produce satisfactory results based on higher temporal and
Multispectral image analysis of woody vegetation 49
spatial resolution of image time series. Satisfactory results were achieved by NIR image
differencing and raster based calculation of derived land use classifications, both showing
changes of remnant vegetation over time.
Ground truth data, however, would enhance the remote process for mapping land use or
land cover. Ground feature information and corresponding records of field locations are
essential for generating accurate ground feature representations and validations of
classification outcomes. Training clusters for supervised classifications could then be
established to more precisely describe discrete land use classes.
We conclude that spectral unmixing techniques, although not explored here in detail, could
allow the computation of forest disturbance measures and improve the land cover
assessment. Recent literature demonstrates the utility of unmixing techniques (Souza et al.
2003, 2005). Furthermore, change detection using Principal Component Analysis (PCA)
could be explored. This is a commonly applied image analysis tool for identifying vegetation
change over time (Richards 1994). Through the reduction of multispectral data redundancy,
the PCA will build uncorrelated components representing the variance of image data. Higher
components are often associated to areas of vegetation change during the observed time
interval and could be useful to detect forest disturbance.
High resolution imagery would help refine the delineations of land cover, land use and forest
disturbance in the case of visual interpretation and training classes with low spectral
variation. Furthermore, higher spectral resolution data (e.g. hyper spectral imagery) would
enable increased differentiation among vegetative ground features due to the enlarged
specificity of spectral signatures associated with plant biophysical characteristics.
Without exception, all image data needs to be accurately georectified in order to correctly
highlight changes on the ground, thus minimizing the error of change/no-change areas. The
error in registration between the 1989 and 2002 Landsat TM data, for example, confounded
the detection of land cover change and forest disturbance.
A combination of both visual and spectral classifications is recommended to refine the land
use mapping. An important input will be expert knowledge of local environment and land
use, on the ground. Furthermore, an object-oriented segmentation approach could support a
refined identification of land cover features such as oil palm plantations (Wahid et al. 2005).
In summary, for future applications of remote sensing techniques applied to land cover and
land use mapping for local area planning on the Popondetta Plain we make the following
recommendations:
o Acquire recent, high-resolution imagery within cost constraints (see below) applicable
to the purpose and currency needs of the project
o Acquire field-derived ground-truth data defining the observation date and
latitude/longitude derived from GPS records, GPS technology description, ground
feature classification, ground feature description, ground photographs, adjacent
feature description
o Apply supervised classifications based on use of field-derived ground-truth training
data and accuracy assessments
o Evaluate potential for improvement in discriminating forest disturbance classes
through spectral mixture modelling or object-oriented approaches
Source and availability of high resolution imagery
General enquiries about cost and availability of high resolution imagery were made for the
Popondetta Plain study area. Often images contain a fair amount of cloud cover and are
relatively costly for a larger area (e.g. ~20 AUD$ per square kilometre, QuickBird
multispectral, enquired January 2008).
Imagery such as Ikonos, QuickBird, SPOT, ALOS and Aster, are available but archives of
local (i.e., Brisbane-based) retailers are often only current to 2005. High amounts of cloud
Multispectral image analysis of woody vegetation 50
cover during the time of image acquisition are always a considerable issue influencing image
quality and utility.
The main distributor for high resolution imagery in Queensland Australia is Geoimage
(Brisbane; www.geoimage.com.au). They stock image data such as Ikonos, QuickBird,
SPOT and Aster.
Some other distributors for image data are:
o IKONOS: http://www.geoeye.com/
o QUICKBIRD: http://www.digitalglobe.com/
o SPOT: http://www.spot.com/welcome/us.php
o ALOS: http://www.ga.gov.au/acres/prod_ser/ALOS/index.jsp
o ASTER: http://edcimswww.cr.usgs.gov/pub/imswelcome/
Multispectral image analysis of woody vegetation 51
References
Desclée, B., de Wasseige, C., Bogaert, P., Defourny, P. 2006. Tropical forest monitoring by
object-based change detection: towards an automated method in an operational
perspective. In The International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, Vol. 34. Proceedings of the 1st International Conference on
Object-based Image Analysis (OBIA 2006) Salzburg University, Austria, July 4-5, 2006.
http://www.commission4.isprs.org/obia06/Papers/13_Automated%20classification%20IC
%20I%20-%20Forest/OBIA2006_Desclee_et_al.pdf (last accessed 21st March 2008).
ERDAS, 2003. ERDAS Field Guide. Atlanta, Leica Geosystems.
ESRI, 2007. ArcGIS 9.2, Redlands, CA, www.esri.com
Hammermaster, E.T. and Saunders, J.C. 1995. Forest Resources and Vegetation Mapping of
Papua New Guinea. Australian Agency for International Development, PNG Resource
Information System (Publication 4), Canberra.
Hubert-Moy, Michel, K., Corpetti, T., Clément, B. undated. Object-oriented mapping and analysis
of wetlands using SPOT 5 data. Online Manucsript Université Rennes,
http://ieeexplore.ieee.org/iel5/4087812/4087813/04088697.pdf (last accessed 21st March
2008).
Jensen, J. R. 2005. Introductory Digital Image Processing: A Remote Sensing Perspective,
Second Edition. Prentice Hall, Upper Saddle River, USA.
Lillesand, T.M. and Kiefer, R.W. 2004. Remote Sensing and Image Interpretation (Fifth Edition).
John Wiley & Sons, 724 pgs.
Mallinis et al. 2006. G. Mallinis, G., Karamanolis, D., Karteris, M. and Gitas, I. 2006. An object
oriented approach for the discrimination of forest areas under the criteria of forest
legislation in Greece using very high resolution data. In The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. 34.
Proceedings of the 1st International Conference on Object-based Image Analysis (OBIA
2006) Salzburg University, Austria, July 4-5, 2006.
http://www.commission4.isprs.org/obia06/Papers/13_Automated%20classification%20IC
%20I%20-%20Forest/OBIA2006_Mallinis_et_al.pdf (last accessed 21st March 2008).
Parsons, M.J. 1992. The world’s largest butterfly endangered: the ecology, status and
conservation of Ornithoptera alexandrae (Lepidoptera: Papilionidae). Tropical
Lepidoptera, 3 (Suppl): 33-60.
Richards, J.A., Remote Sensing Digital Image Analysis, 1994, Springer Press, Berlin
Souza, C. Jr., Firestone, L., Silva, L.M. and Roberts, D. 2003. Mapping forest degradation in the
eastern Amazon from SPOT4 through spectral mixture models. Remote Sensing of
Environment 87: 494–506.
Souza, CM, Roberts, DA and Cochrane, MA. 2005. Combining spectral and spatial information to
map canopy damage from selective logging and forest fires. Remote Sensing of
Environment, 98: 329–343.
Wahid, B O; Nordiana, A A and Tarmizi, A M, 2005. Satellite Mapping of Oil Palm Land Use,
Malaysian Palm Oil Board Information Series, No. 255, Kuala Lumpur.
Williams, K.J., Storey, R.J.L., Mitchell, D.K. 2007. Preventing the Extinction of the Queen
Alexandra Birdwing Butterfly: Land Use and Forest Loss on the Popondetta Plain, Oro
Province, Papua New Guinea. Conservation International Melanesia and CSIRO
Sustainable Ecosystems, Atherton.
Multispectral image analysis of woody vegetation 52
Appendix 1: Metadata
Multispectral classification of remnant woody vegetation land cover (Landsat
TM7 and SPOT4 2002) on the Popondetta Plain of Oro Province, Papua New
Guinea
The described data is available on request to the project manager:
Mr David Mitchell
Melanesian Centre for Biodiversity Conservation
Conservation International PNG Programme
PO Box 943 Alotau
Milne Bay Province
Papua New Guinea
dmitchell@conservation.org
This metadata template is used by CSIRO Sustainable Ecosystems
Dataset Title:
Multispectral classification of remnant woody vegetation land cover (Landsat
TM7 and SPOT4 2002) on the Popondetta Plain of Oro Province, Papua New
Guinea
Abstract:
Supervised classification of Landsat TM7 (2002) and Spot4 (2002) was
conducted. Established training classes were based on Google Earth and pan-
sharpened imagery (Spot2, 2002) and pan chromatic Landsat layer (2002).
Established training classes were bare soil, open land, open forest, grassland,
mature forest, oil palm plantations and oil palm smallholdings,
garden/mixture, cloud, shadow, water. The matured forest class was extracted
from both classification results. The two datsets were unioned, and the final
raster shows three classes: Forest areas identified through both Landsat TM7
and Spot4 classifications, and two additional classes showing matured forest
where one or other classification mapped the forest class. The unioned data
was applied a 3x3 kernel majority filter deriving the final raster dataset with a
spatial resolution of 28.5m
Overall Discipline:
Conservation Planning
Originator Organisation:
CSIRO Sustainable Ecosystems
Dataset Creation Date:
25th February 2008
Metadata Creation Details Date:
25th February 2008
Custodian
CSIRO Sustainable Ecosystems
Metadata Contact Person Name:
Mr Ralf-D Schroers
Metadata Contact Email:
Ralf-Dieter.Schroers@csiro.au
Data Details
ANZLIC Identifier:
ANZQL----
RFCD Subject Classifier:
ANZLIC Theme:
Vegetation
Jurisdiction:
Papua New Guinea
Beginning Date:
2002
Ending Date:
2002
Progress:
Version 23rd Feb. 2008, completed
Maintenance and Update
Frequency:
Data is complete for the purpose
Multispectral image analysis of woody vegetation 53
Access Constraint:
For non-commercial use within Conservation International, CSIRO and
partners
Lineage:
Based on supervised classification oof Landsat TM7 and Spot4 image data; 3
matured forest : 1 - only Spot4 derived; 4 – only Landsat TM7 derived; 5 –
derived from both Spot4 and Landsat TWM7;
Spatail resolution: 28.5 m
8 bit, signed integer
Completeness:
Complete for purpose
Alternate Title:
Author Assigned Keywords:
Vegetation, Papua New Guinea, Classification result (Landsat TM7, SPOT4)
GCMD Parameters:
Online Link Name:
Data
Online Link URL:
Online Link Type:
Dataset Last Revision Date:
23rd Feb. 2008
Stored Data Format:
Stored Data Location:
Stored Data Identification:
Stored Data Volume:
Distribution Medium:
Metadata Last Update Details
Date:
Metadata Future Review Details:
n/a
Parent Metadata Record:
Additional Metadata:
SPOT4 (2002) image data; Landsat TM7 (2002)
CSIRO Specific Information
CSIRO Division Site:
CSIRO Sustainable Ecosystems – Atherton
CSIRO Project Name:
Understanding and Mitigating Risks of Existing Plantations – Higaturu Oil Palm:
QABB Conservation Strategy and KBA Delineation Coastal New Guinea
Wilderness
Platform Source Model Name:
ERDAS Imagine image processing and ESRI ARCGIS spatial analysis
Survey Experiment or Model
Name:
QABB Conservation Strategy: Multispectral image analysis of woody vegetation
extent and disturbance on the Popondetta Plain of Papua New Guinea
CSIRO Project PSS Identifier:
PNO LE96A
CSIRO Project Leader Name:
Kristen J Williams
Acknowledgements:
As given in the report
Bibliography:
Schroers, R-D and Williams, K.J. 2008. Multispectral image analysis of woody
vegetation extent and disturbance on the Popondetta Plain of Papua New
Guinea: Identifying areas of potential habitat for the Queen Alexandra
Birdwing Butterfly. Conservation International Melanesia and CSIRO
Sustainable Ecosystems, Atherton.
Spatial Specific Information
Attribute Accuracy:
attribute accuracy depends on georegistration accuracy of source image data;
validation assessments provide an indication;
Logical Consistency:
No inconsistency
Spatial Representation Method:
ArcGIS, vector data, shapefile
Multispectral image analysis of woody vegetation 54
North Bounding Latitude:
9072010m
South Bounding Latitude:
9004009m
West Bounding Longitude:
590128m
East Bounding Longitude:
662774.5m
Location Keyword:
Popandetta Plain, Oro Province, Papaua New Guinea
Geographic Extent Polygon:
Minimum Vertical Extent:
Maximum Vertical Extent:
Vertical Datum:
Positional Accuracy:
Source data dependent
Scale Denominator:
Spatial Resolution Size:
Spatial Resolution Units:
meter
Horizontal Reference System -
Projection:
AGD 1984 Transverse Mercator
Horizontal Reference System -
Datum:
GCS Australian 1984
Horizontal Reference System -
Additional:
Geographic Name Extent:
GEN Category:
GEN Custodial Jurisdiction:
GEN Name:
Originator Organisation's Contact Information
Data Contact Point Name:
Ralf-D. Schroers
Data Contact Point Email:
Ralf-Dieter.Schroers@csiro.au
Telephone:
+61 7 4091 8822
Contact Organisation:
CSIRO
Contact Organisation Mail
Address:
PO Box 780, Atherton, 4883, QLD, Australia
Contact Organisation Locality:
Atherton, Queensland, Australia
Contact Organisation State:
Queensland
Contact Organisation Postcode:
4883
Contact Organisation Country:
Australia
Dataset Contributing Persons:
Kristen Williams
Multispectral image analysis of woody vegetation 55
Woody remnant vegetation change (Landsat NIR bands 1989-2002) on the Popondetta Plain
of Oro Province, Papua New Guinea
This metadata template is used by CSIRO Sustainable Ecosystems
Dataset Title:
Woody remnant vegetation change (Landsat NIR bands 1989-2002) on the
Popondetta Plain of Oro Province, Papua New Guinea
Abstract:
Change detection data set, derived from image differecing of Landsat NIR
bands (Landsat TM5 1989 and TM7 2002). The NIR band raster calculation
generated a difference image, where deforested areas were corresponding to
values below -42. Two classes were established: areas equal or below -42 and
areas showing values above -42. The raster data was cloud masked, and a
majority filter (3x3 cells) applied, and converted into a vector dataset.
Polygons greater than two cells (1624.5 sqm) were eliminated.
Overall Discipline:
Conservation Planning
Originator Organisation:
CSIRO Sustainable Ecosystems
Dataset Creation Date:
25th February 2008
Metadata Creation Details Date:
25th February 2008
Custodian
CSIRO Sustainable Ecosystems
Metadata Contact Person Name:
Mr Ralf-D Schroers
Metadata Contact Email:
Ralf-Dieter.Schroers@csiro.au
Data Details
ANZLIC Identifier:
ANZQL----
RFCD Subject Classifier:
ANZLIC Theme:
Vegetation
Jurisdiction:
Papua New Guinea
Beginning Date:
1989
Ending Date:
2002
Progress:
Version 10th March 2008, completed
Maintenance and Update
Frequency:
Data is complete for the purpose
Access Constraint:
For non-commercial use within Conservation International, CSIRO and
partners
Lineage:
Change detection data set, derived from image differecing of Landsat NIR
bands (Landsat Tm5 1989 and TM7 2002).
Completeness:
Complete for purpose
Alternate Title:
Author Assigned Keywords:
Vegetation, Papua New Guinea, Classification result (Landsat TM7 2002,
Landsat TM5 1989)
GCMD Parameters:
Online Link Name:
Data
Online Link URL:
Online Link Type:
Dataset Last Revision Date:
10th March 2008
Stored Data Format:
Stored Data Location:
Stored Data Identification:
Stored Data Volume:
Multispectral image analysis of woody vegetation 56
Distribution Medium:
Metadata Last Update Details
Date:
Metadata Future Review Details:
n/a
Parent Metadata Record:
Additional Metadata:
Landsat TM7 2002; Landsat TM5 1989
CSIRO Specific Information
CSIRO Division Site:
CSIRO Sustainable Ecosystems – Atherton
CSIRO Project Name:
Understanding and Mitigating Risks of Existing Plantations – Higaturu Oil Palm:
QABB Conservation Strategy and KBA Delineation Coastal New Guinea
Wilderness
Platform Source Model Name:
ERDAS Imagine image processing and ESRI ARCGIS spatial analysis
Survey Experiment or Model
Name:
QABB Conservation Strategy: Multispectral image analysis of woody vegetation
extent and disturbance on the Popondetta Plain of Papua New Guinea
CSIRO Project PSS Identifier:
PNO LE96A
CSIRO Project Leader Name:
Kristen J Williams
Acknowledgements:
As given in the report
Bibliography:
Schroers, R-D and Williams, K.J. 2008. Multispectral image analysis of woody
vegetation extent and disturbance on the Popondetta Plain of Papua New
Guinea: Identifying areas of potential habitat for the Queen Alexandra
Birdwing Butterfly. Conservation International Melanesia and CSIRO
Sustainable Ecosystems, Atherton.
Spatial Specific Information
Attribute Accuracy:
attribute accuracy depends on georegistration accuracy of source image data
(Landsat);
Logical Consistency:
No inconsistency
Spatial Representation Method:
ArcGIS, vector data, shapefile
North Bounding Latitude:
-950113m
South Bounding Latitude:
-992236m
West Bounding Longitude:
598770m
East Bounding Longitude:
656853m
Location Keyword:
Popandetta Plain, Oro Province, Papaua New Guinea
Geographic Extent Polygon:
Minimum Vertical Extent:
Maximum Vertical Extent:
Vertical Datum:
Positional Accuracy:
Scale Denominator:
Spatial Resolution Size:
Spatial Resolution Units:
Horizontal Reference System -
Projection:
WGS 1984 UTM ZONE 55N, Transverse Mercator
Horizontal Reference System -
Datum:
WGS 84
Horizontal Reference System -
Additional:
Geographic Name Extent:
Multispectral image analysis of woody vegetation 57
GEN Category:
GEN Custodial Jurisdiction:
GEN Name:
Originator Organisation's Contact Information
Data Contact Point Name:
Ralf-D. Schroers
Data Contact Point Email:
Ralf-Dieter.Schroers@csiro.au
Telephone:
+61 7 4091 8822
Contact Organisation:
CSIRO
Contact Organisation Mail
Address:
PO Box 780, Atherton, 4883, QLD, Australia
Contact Organisation Locality:
Atherton, Queensland, Australia
Contact Organisation State:
Queensland
Contact Organisation Postcode:
4883
Contact Organisation Country:
Australia
Dataset Contributing Persons:
Kristen Williams
Dataset Title:
Woody remnant vegetation change (Landsat classification 1989-2002) on the
Popondetta Plain of Oro Province, Papua New Guinea
Abstract:
Supervised classification of Landsat TM7 (2002) and unsupervised
classification of Landsat TM5 (1989) was conducted. Two datasets were
classed into forest, and non-forest class, and additional cloud mask was
applied. Change detection derived areas that were deforested in 2002
(forested in 1989), areas that were reforested in 2002 (formerly deforeswted
in 1989), intact forested areas (forested 1989 and 2002), and non-forest areas
(not forested in 1989 and 2002). Areas smaller than one cell unit
(28.5mx28.5m) were eliminated.
Overall Discipline:
Conservation Planning
Originator Organisation:
CSIRO Sustainable Ecosystems
Dataset Creation Date:
25th February 2008
Metadata Creation Details Date:
25th February 2008
Custodian
CSIRO Sustainable Ecosystems
Metadata Contact Person Name:
Mr Ralf-D Schroers
Metadata Contact Email:
Ralf-Dieter.Schroers@csiro.au
Data Details
ANZLIC Identifier:
ANZQL----
RFCD Subject Classifier:
ANZLIC Theme:
Vegetation
Jurisdiction:
Papua New Guinea
Beginning Date:
1989
Ending Date:
2002
Progress:
Version 20th March 2008, completed
Maintenance and Update
Frequency:
Data is complete for the purpose
Access Constraint:
For non-commercial use within Conservation International, CSIRO and
partners
Multispectral image analysis of woody vegetation 58
Lineage:
Change detection data set, derived from image differecing of supervised
classification of Landsat TM7 (2002) and unsupervised classification of Landsat
TM5 (1989)
Completeness:
Complete for purpose
Alternate Title:
Author Assigned Keywords:
GCMD Parameters:
Online Link Name:
Data
Online Link URL:
Online Link Type:
Dataset Last Revision Date:
15th March 2008
Stored Data Format:
Stored Data Location:
Stored Data Identification:
Stored Data Volume:
Distribution Medium:
Metadata Last Update Details
Date:
Metadata Future Review Details:
n/a
Parent Metadata Record:
Additional Metadata:
Landsat TM7 2002; Landsat TM5 1989
CSIRO Specific Information
CSIRO Division Site:
CSIRO Sustainable Ecosystems – Atherton
CSIRO Project Name:
Understanding and Mitigating Risks of Existing Plantations – Higaturu Oil Palm:
QABB Conservation Strategy and KBA Delineation Coastal New Guinea
Wilderness
Platform Source Model Name:
ERDAS Imagine image processing and ESRI ARCGIS spatial analysis
Survey Experiment or Model
Name:
QABB Conservation Strategy: Multispectral image analysis of woody vegetation
extent and disturbance on the Popondetta Plain of Papua New Guinea
CSIRO Project PSS Identifier:
PNO LE96A
CSIRO Project Leader Name:
Kristen J Williams
Acknowledgements:
As given in the report
Bibliography:
Schroers, R-D and Williams, K.J. 2008. Multispectral image analysis of woody
vegetation extent and disturbance on the Popondetta Plain of Papua New
Guinea: Identifying areas of potential habitat for the Queen Alexandra
Birdwing Butterfly. Conservation International Melanesia and CSIRO
Sustainable Ecosystems, Atherton.
Spatial Specific Information
Attribute Accuracy:
attribute accuracy depends on georegistration accuracy of source image data
(Landsat);
Logical Consistency:
No inconsistency
Spatial Representation Method:
ArcGIS, vector data, shapefile
North Bounding Latitude:
-851537m
South Bounding Latitude:
-1068707m
West Bounding Longitude:
521501m
East Bounding Longitude:
768169m
Multispectral image analysis of woody vegetation 59
Location Keyword:
Popandetta Plain, Oro Province, Papaua New Guinea
Geographic Extent Polygon:
Minimum Vertical Extent:
Maximum Vertical Extent:
Vertical Datum:
Positional Accuracy:
Source data dependent
Scale Denominator:
Spatial Resolution Size:
Spatial Resolution Units:
Horizontal Reference System -
Projection:
WGS 1984 UTM Zone 55S
Horizontal Reference System -
Datum:
WGS84
Horizontal Reference System -
Additional:
Geographic Name Extent:
GEN Category:
GEN Custodial Jurisdiction:
GEN Name:
Originator Organisation's Contact Information
Data Contact Point Name:
Ralf-D. Schroers
Data Contact Point Email:
Ralf-Dieter.Schroers@csiro.au
Telephone:
+61 7 4091 8822
Contact Organisation:
CSIRO
Contact Organisation Mail
Address:
PO Box 780, Atherton, 4883, QLD, Australia
Contact Organisation Locality:
Atherton, Queensland, Australia
Contact Organisation State:
Queensland
Contact Organisation Postcode:
4883
Contact Organisation Country:
Australia
Dataset Contributing Persons:
Kristen Williams
