Você está na página 1de 10

SEGEMENTATION OF MULTI TEMPORAL RADAR FOR

WETLAND INVENTORY
Geoffrey Horn
School of Biology, Earth and Environmental Science, UNSW
209 Cobra St, Dubbo, NSW, 2830
02 6841 7515
ghdielectric@gmail.com

Abstract
The Wetlands Policy of the Australian Government states the need to identify
and complete national wetland inventories for all wetlands in Australia. An
important part of inventory is the mapping of wetlands. However, the wet/dry
tropical climate in the north of the continent has yet to be mapped to a baseline
level with any reasonable accuracy. The highly complex seasonal cycling in the
region means that any attempt at mapping the location of a particular wetland
component must account for the potential change throughout the seasonal
cycle.
Optical remote sensing systems are non optimal for the wet/dry tropical
wetlands of Northern Australia because the area is almost continually covered
by either clouds or smoke. Radar imagery is insensitive to both smoke and
cloud. A multi-temporal approach to the situation is well suited to the highly
dynamic nature of the wet/dry tropics of northern Australian and Kakadu
National Park in particular.
This paper presents results regarding how multiple date radar data can be
processed to gather information about the extent of wetland inundation, the
period of wetland inundation and the pattern of drainage. The speckle problems
associated with radar imagery were avoided by using a customised
segmentation program written specifically for this research. Analysis of the
segments to extract information for wetland inventory was achieved by using
robust thresholds for each distinct wetland component and inventory
requirement. Once defined, these thresholds were applied and the results
assessed to determine the accuracy of the results.
The research is significant because it creates new ecological information for the
region. The application of a new segmentation technique and wetland inventory
information based on segment analysis are the key results of this research.

Introduction
Wetlands have been variously defined by different organisations, from different
perspectives or for different purposes, which has led to some confusion. The
definition of wetlands used in this research is from the Ramsar Convention,
which states wetlands are “Areas of marsh, fen, peatland or water, whether
natural or artificial, permanent or temporary, with water that is static or flowing,
fresh, brackish or salt, including areas of marine water the depth of which at low
tide does not exceed 6 metres” (Ramsar Convention Secretariat, 2006).

1
Globally, little information has been compiled for wetland areas, particularly
from an inventory perspective (Finlayson et al. 2001; Phinn et al. 1999). The
fundamental components of a wetland inventory are wetland area, boundary,
location and geomorphic setting, all of which can be captured from remotely
sensed imagery (Phinn et al., 1999; Finlayson & von Oertzen, 1993).
One of the major goals of the Wetlands Policy of the Australian Government is
to conserve, repair and manage wetlands wisely (Commonwealth Government
of Australia 1997). As an implicit part of this policy, accurate and up-to-date
information about existing wetlands in Australia is required. One of the major
strategies of the policy is to “Develop and maintain a comprehensive inventory
of wetlands on Commonwealth lands and waters, areas leased by the
Commonwealth agencies and areas under co management arrangements
involving the Commonwealth.” (Commonwealth Government of Australia 1997).
Developing wetland inventories in tropical northern Australia is difficult for two
main reasons. Firstly the area is blanketed by either cloud or smoke cover most
of the year, making cloud free optical images rare. Secondly, the region
undergoes an annual cycle of wet to dry conditions, with the timing and extent
of the cycle varying from year to year. The wetlands are part of this dynamic
system, changing rapidly with the onset of the wet season, and so a multi-
temporal approach is required in their mapping. These problems can be
overcome by the use of radar imagery, which is not affected by clouds or
smoke, and is available throughout the year.
However, Radar data is intrinsically noisy and is difficult to use with many
standard classification techniques. Radar noise, termed speckle, is
encountered in all radar imagery as a result of backscatter from several
structural elements interacting in a multiplicative fashion, leading to additive and
destructive interference in the returned wave of each imaging cell (Ulaby et al.
1986). Applying standard classification techniques to radar imagery produces
noise affected results (Crawford & Ricard 1998). Preprocessing the radar
imagery, through segmenting areas of similar pixel values and textures reduces
noise and improves classification results (Haralick & Shapiro 1992; Caves &
Quegan 1998).
Segmentation of radar imagery is relatively complex with the mathematical logic
for segment boundary placement difficult to reproduce consistently (Nussbaum
& Menz 2008). The use of a robust and time-independent technique is required
to allow the valid comparison of time series data. Others have developed
techniques that require varying segmentation parameters by image acquisition
date, although this introduces a subjective component giving a less reliable
result (Nussbaum & Menz 2008). This paper presents a method that avoids
subjective segmentation parameters. When applied to multi temporal radar
imagery this method allows the extraction of wetland inventory information from
the resulting segments. As Radarsat-1 data has a 24-day repeat cycle and
consistent viewing geometry change detection can be easily performed.
Method
Segmentation can be used to reduce the impact of speckle on radar imagery,
and can be used to identify sub-populations of backscatter within a previously

2
continuous distribution. Segmentation takes multiplicative noise into account, by
using the statistics of a small region to determine its association with its
surrounding neighbours. If the segment is statistically the same as its
neighbours it is joined with them to produce a bigger segment. However, if the
statistics of two neighbouring regions are dissimilar, a boundary is drawn and
the two regions are kept separate. For each segment it is the statistics of the
that are recorded in the final output, namely that of the segment mean and its
standard deviation, which is a measure of texture. A custom segmentation
software routine was written in C++ (Borland C++ Builder) to take an input radar
image and divide it into a number of adjoining segments where each segment
member pixel is statistically similar to other member pixels. A full description of
the method is detailed in Horn (2009).
Once image noise is reduced through segmentation the images can be
classified using standard thresholding. Thresholds can be based on an
understanding of the way radar physically interacts with different wetland
components. Both open water and flooded forest are readily identifiable in radar
imagery due to two distinctly different interactions with incident radiation.
Open water is relatively smooth with respect to incident radar beams and so
scatters most of the energy away from the receiving antenna, giving a very low
returned value in the subsequent imagery.
Flooded forests (e.g. Melaleuca or paperbark swamps) are areas inundated
with freshwater for long periods and have extensive areas of emergent trees.
These areas typically occur at the upland margin of the floodplain surfaces. The
incident radar beam reflects from the water to the trunk and back to the receive
antenna at a very high fraction of the original energy, resulting in an extremely
high backscatter response (Richards et al, 1999). This response is typically
maintained while the water level remains high, and has a uniform texture (low
standard deviation) in areas of almost permanent inundation. Backscatter
values of above -8 dB are typical of these areas after segmentation.
In order to determine wetland extent, thresholds can be applied to the mean of
the time series. Inundation period could be determined by counting the number
of instances at which a location was within the wetland threshold.
Because the defined thresholds have a physical justification (such as specular
reflection for open water or a double bounce for flooded paperbark swamps)
they should be consistent with time after calibration of the imagery. For a given
season, the same threshold should apply to a similar season image from a
different year.

Application
Study Site
Kakadu National Park covers an area of approximately 20 000 km2 in the
Northern Territory of Australia. The Park was established in 1979. The park
contains a diverse range of habitats and ecological communities, from upland
Eucalypt woodlands totalling 47 000 ha to vast floodplains of approximately 147
000 ha (Finlayson & Moser 1991). Areas of paperbark and mangrove swamps

3
are also found in the park. (Finlayson et al. 2001; Finlayson & von Oertzen
1996)
An image subset from the Munmalari area was used as a type example of the
paperbark swamp landscape element (Figure 1). The subset is dominated by a
large ENE-WSW trending paperbark swamp, surrounded to the north and south
by eucalypt woodland (Figure 2). To the southwest of the image subset, is the
right bank of the South Alligator River and the associated floodplain complex. In
the NE corner is a perched window lake called Alligator Hole with associated
outwash channel running to the north.

Figure 1. Image subset location within Kakadu National Park

4
Figure 2. Paperbark image subset, Landsat 3,2,1 as R,G,B

A series of 11 Radarsat-1 (C-HH) scenes were collected from January to


October 1998 in a 24 day repeat cycle. Each image was collected in the same
standard 4 beam mode with an incidence of 34-40o. The image data covered a
100 km by 100 km footprint with a pixel size of 12.5 m. Each image was
rectified and converted to backscatter coefficient (σ0). After the images had
been rectified and calibrated, they were segmented using the custom
segmentation software described earlier resulting in a two band image
(backscatter mean and standard deviation). The segmentation results were
then stacked into a single image and the mean and range for the entire series
was calculated.
Primary thresholds for the main landscape elements were determined from on-
ground investigations (Table 1). The thresholds were then applied to generate
layers for each landscape element, from which the wetland extent and
inundation maps were derived.

5
Landscape element Mean (dB) Range (dB)
Open Water < -15 any
Floodplain ≤ -8 > 7.5
Paperbark > -8 ≤ 7.5
Woodland > -12 and < -8 any

Table 1. landscape element thresholds.

Accuracy assessment
An accuracy assessment was undertaken using a stratified random distribution
of 10 000 points. Wetland extent was compared with an existing structural
vegetation classification (Environmental Data Directory, DEHA 2008). For the
purposes of the analysis each of the map categories was grouped into a
wetland or non wetland category. Stratification was performed proportional to
the extent of each reference class.

Results
The subset image shows a good relationship between the identified wetland
class and the reference dataset (Figure 3). Some commission errors are
evident in the open forest and woodland classes. The floodplain and paperbark
swamp classes are well matched.

class Users’ Accuracy Producers’ Accuracy


wetland 0.526 0.739
non wetland 0.915 0.808

Table 2. Users’ and producers’ accuracy for wetland extent layer.

The Kappa statistic was 0.479 which indicates that the wetland extent layer has
a moderate agreement with the existing vegetation layer. Users’ accuracy is
relatively low (0.526) for the wetland class due to the confusion with the open
forest-woodland and pandanus savannah classes (Table 2). Producers’
accuracy for the wetland class is higher (0.739) as the measure is not sensitive
to omission errors. Using simple thresholds to determine wetland extent worked
well for some extreme landscape elements such as paperbark swamp and
open water and less well for more similar elements such as woodlands and mid
floodplain classes.

6
Figure 3. Wetland extent for the Paperbark Subset

7
Figure 4. Paperbark subset inundation period.

8
Inundation period was ascertained by the number of times a pixel was classified
as open water or paperbark swamp. By counting the number of times an area
was classed as inundated, the period of inundation in 24-day increments was
mapped. Inundation ranged from 0 to 11 occurrences across the 11 image
dates. These occurrences were multiplied by the number of days between
images (24) to create an inundation period in days.
The subset image shows a strong relationship between inundation period and
paperbark vegetation (Figure 4). The paperbark swamp areas are dominated by
extended water residence from 145 to 240 days. This is a defining
characteristic of the class and indicates the ease with which double bounce
characterises flooded forest. The floodplain area in the Southwest of the image
has a good agreement with inundation period expected from the floodplain
topography, with a general increase in water residence with increasing distance
from river channel.

Conclusion
Tropical northern Australia presents unique difficulties for establishing wetland
inventories. Smoke and cloud coverage and the dynamic landscape require a
combination of longer wavelength imaging and multi temporal sampling. The
speckle problems inherent with radar data have been overcome through a
segmentation routine, and information has been extracted for the resulting
segments using robust thresholds.
This paper presents the initial results for wetland extent and inundation period
from a subset of the full dataset. Further research is required to assess the
accuracy of the inundation period results and fine tune the segmentation
routine. The extraction of wetland extent and inundation period maps from
multiple radar images shows a great deal of promise for wetland inventory.

References
Commonwealth Government of Australia, 1997, The Wetlands Policy of the
Commonwealth Government of Australia, Australian Government Publishing
Service, Canberra.
Coppin, P., Jonckheere, I., Nackaerts, K., Muys, B., Lambin, E., 2004 ‘Digital
change detection methods in ecosystem monitoring: a review’, International
Journal of Remote Sensing, vol. 25, no. 9, pp. 1565 – 1596.
Drury, S.A., 1993, Image Interpretation in Geology, 2nd ed, Chapman & Hall,
London.
Environmental Data Directory, DEHA. 2008. http://www.environment.gov.au/
metadataexplorer/explorer.jsp.
Finlayson, M.J., and Moser , M., 1991, Australasian and Oceania in: Finlayson,
C.M. and Moser, M. (eds.) Wetlands. International Waterfowl and Wetlands
Research Bureau, United Kingdom.
Finlayson, C.M., and von Oertzen, I., 1993, ‘Wetlands of Australia Northern
(tropical) Australia’, in D. F. Whigham, D. Dykyjova & S. Hejny (eds), Wetlands

9
of the world I: inventory, ecology and management, Kluwer Academic
Publishers, Dordrecht, pp. 195-243
Finlayson, C.M., and von Oertzen, I., (eds)., 1996, ‘Landscape and vegetation
ecology of the Kakadu region, Northern Australia’, Kulwer Academic Publishers,
London
Finlayson, C.M., Davidson, N.C., and Stevenson, N.J., (eds)., 2001. Wetland
inventory, assessment and monitoring: Practical techniques and identification
nd
of major issues. Proceedings of Workshop 4, 2 International Conference on
Wetlands and Development, Dakar, Senegal, 8-14 November 1998,
Supervising Scientist Report 161, Supervising Scientist, Darwin
Hess, L., Melack, J., and Simonett, D., 1990, Radar detection of flooding
beneath the forest canopy: a review, International Journal of Remote Sensing,
vol 11, no 7, pp1313-1325.
Horn, G.D., 2009, Segmentation of multi-temporal radar images for wetland
inventory. Unpublished Ph.D thesis, School of Biology Earth and Environmental
Sciences, UNSW, Sydney Australia.
Kasischke, E.S., Melack, J.M., and Dobson, M.C., 1997, The use of Imaging
Radars for Ecological Applications -A Review, Remote Sensing of Environment,
vol. 59, pp 141-156.
rd
Lillesand, T. and Kiefer, R., 1994. Remote Sensing and Image Interpretation 3
ed. John Wiley & Sons, New York, U.S.A.
Phinn, S.R., 1998, A framework for selecting appropriate remotely sensed data
dimensions for environmental monitoring and management, International
Journal of Remote Sensing, vol. 19, no. 17, pp.3457-3463.
Phinn, S.R, Hess, L., and Finalyson, C.M., 1999. An assessment of the
usefulness of remote sensing for coastal wetland inventory and monitoring in
Australia. In Finlayson C.M. & Spiers A.G. (eds). Techniques for enhanced
wetland inventory and monitoring. Supervising Scientist Environment Australia,
Canberra.
Ramsar Convention Secretariat, 2006. The Ramsar Convention Manual: a
guide to the Convention on Wetlands (Ramsar, Iran, 1971), 4th ed. Ramsar
Convention Secretariat, Gland, Switzerland.
Richards, J.A., Sun, G.Q., and Simonett, D.S., 1987, L Band radar backscatter
modelling of forest stands, IEEE Transactions on Geoscience and Remote
Sensing, vol. 25, no. 4, pp.487-498.
Story, M., and Congalton, R.G., 1986, Accuracy assessment: a users
perspective. Photogrametric Engineering and Remote Sensing, 52, 397– 399.
Ulaby, FT, Moore, RK & Fung, AK 1986, Microwave remote sensing, Active and
passive, volumes I-III, Artech House, INC. Dedham, Massachusetts.

10

Você também pode gostar