UGANDAN DAMBO WETLAND CLASSIFICATION USING MULTISPECTRAL AND
TOPOGRAPHIC REMOTE SENSING DATA
Matthew K. Hansen, Research Assistant
Philip E. Dennison, Assistant Professor
Scott A. Graves, Research Assistant
Department of Geography
University of Utah
260 S. Central Campus Dr., Rm. 270
Salt Lake City, UT 84112
matt.hansen@geog.utah.edu
dennison@geog.utah.edu
scott.graves@geog.utah.edu
David J. Brown, Assistant Professor
Department of Crop and Soil Sciences
Washington State University
PO Box 646420
Pullman, WA 99164
david_brown@wsu.edu
ABSTRACT
The seasonally saturated dambo wetlands of Central and Southeastern Africa are likely a substantial source of
methane (CH
4
), an important greenhouse gas. Dambo soil, vegetative, and hydrologic characteristics vary along a
gradual topographic gradient; from relatively dry uplands, to occasionally inundated margins, to more frequently
inundated floors, to perennially inundated bottoms. CH
4
production presumably also varies along the gradient from
uplands to bottoms, making mapping dambo topography an important first step in determining regional dambo
methane production. Multispectral remote sensing data and topographic attributes were used to map upland, margin,
floor, and bottom classes of dambo wetlands in a 2,200 square kilometer study area located in central Uganda. Two
Système Pour l’Observacion de la Terre (SPOT) 4 multispectral scenes were acquired to coincide with the beginning
and end of the January-to-March dry season in the study area. Spectral indices and spectral mixture modeling
fractions were calculated for both SPOT scenes. Phenological changes occurring within each dambo class were
quantified by comparing remote sensing metrics in each SPOT scene. Multispectral metrics were combined with
topographic data from the Shuttle Radar Topography Mission (SRTM) using a binary decision tree (BDT) to
classify the study area into upland, margin, floor, and bottom classes. Field data were used for training the classifier
and assessing the accuracy of the final classification. This classification will be used to locate methane emission
sampling sites within the study area as part of the first region-scale dambo methane emissions modeling project.
BACKGROUND
The anaerobic microbial decomposition of biomass that occurs in wetlands makes them the largest natural
source of atmospheric CH
4
. Wetlands are responsible for 20-45% of the total annual global emissions of CH
4
(Shindell et al., 2004), and CH
4
is approximately 20 times as effective a greenhouse gas as carbon dioxide (CO
2
) on
a molecular level (Rodhe, 1990; Faulkner, 2004). CH
4
emissions from wetlands in the tropics are among the least
understood (Bartlett and Harriss, 1993), adding to the significance of this research.
“Dambo” is one of several dialectal terms used to describe the seasonally saturated, grassy, narrow depressions
covering as much as 20% of the plateau regions of Central and Southern Africa (Bullock, 1992; Mäckel, 1974).
While most dambo wetlands are waterlogged for at least a portion of the year, many of them dry out at the surface
during the four- to six-month dry season typically found in semi-arid Africa (Acres et al., 1985). However, the
“sponge-like” center of most dambos stays moist even during the dry season, sustaining the higher-density
herbaceous vegetation in this zone (Mäckel, 1985). The relatively planar topography (½ – 2º slope) typical of
ASPRS 2008 Annual Conference
Portland, Oregon April 28 - May 2, 2008
dambo wetlands produces little hydraulic energy, which in turn facilitates soil saturation and inhibits channel
formation (Acres et al., 1985; Von der Heyden, 2004).
margin
margin
floor
floor
bottom
upland
upland
Figure 1. Dambo cross-section.
Dambo vegetation and soil types vary, but researchers have identified certain patterns characteristic of the
landforms. Acres et al. (1985) used vegetation, soil type, and topography to identify three general zones within a
dambo: the margin, floor, and bottom (Figure 1). The margin, at the perimeter of a dambo, is the transition zone
from higher, often wooded terrain to grasslands that gently slope toward the center of the wetland (Von der Heyden,
2004). The dambo floor is a continuation of the grassland that transitions to sedges and other herbaceous plants
near the bottom. At the center of the idealized dambo is the bottom, an area usually only suitable as habitat for
those plants adapted to saturated soils. Because inundation varies according to these dambo zones, becoming
progressively wetter from margin to bottom, distinguishing between them is essential to modeling methane
emissions. The purpose of this project was to use remotely sensed information to classify dambo wetlands within a
study area in Uganda, resulting in an accurate map for use in subsequent region-scale methane emissions
measurement and modeling.
METHODS
This landscape classification designated each 20 meter pixel within a study area as bottom, floor, or margin,
with the addition of an upland class for the surrounding terrain. The 2,200 sq. km. study area (Figure 2) was
selected from a remote region of central Uganda bounded by the Mayanja and Lugogo Rivers. The study area
demonstrated the topography, climate, and vegetation typical of African dambos. Additionally, the site offered a
range of elevation and precipitation patterns in a relatively compact area, hopefully allowing for broader application
of the methane model that will result from future research.
Field sampling of the dambo classes occurred in January and February, 2007, during one of the annual dry
seasons typically experienced in the study area. The 236 sites sampled (Figure 2) during the field campaign were
chosen in one of two ways. First, to capture dambo wetlands on multiple scales, a stratified random sample set was
selected according to stream order. Second, samples were taken opportunely as they were observed within the study
area. Two sampling techniques were also employed; one recorded the location of relatively homogeneous areas
representing each of the dambo classes, and one focused on variation by measuring the extent of each class along
transects running perpendicular to hydrologic flow.
Dambo classes were identified in the field according to a combination of soil, topography, and vegetation
characteristics (Table 1). Bottoms were readily identified by their distinctive hummocky micro-topography and
frequent inundation, margins were distinguished by sandy surface soils and slope measurements in excess of 2%,
and uplands were categorized by their reddish soils according to Munsell chroma and hue. Because floors exhibited
fewer unique soil or slope characteristics, they were identified by not belonging to any other class. In practice, flat,
grassy terrain and the absence of hummocks were generally indicative of dambo floors.
Several sources of remotely sensed topographic and multispectral image data served as inputs to the
classification (Table 2). Two four-band Système Pour l’Observacion de la Terre (SPOT) 4 satellite images, captured
December 10, 2006, and February 21, 2007 (Figure 2), provided the primary inputs and set a standard spatial
resolution for the classification of 20 meters. The SPOT images were also used to calculate two vegetation indices,
the Normalized Difference Vegetation Index (NDVI; Rouse et al., 1973) and the Normalized Difference Infrared
Index (NDII; Hunt and Rock, 1989). Endmember fractions resulting from a spectral mixture analysis (SMA; Adams
et al., 1993), specifically, the percentages of photosynthetic vegetation (PV), non-photosynthetic vegetation
(NPV)/soil, and shade for each SPOT pixel, provided additional vegetation-related inputs. The differences between
the vegetation index and SMA values for the two SPOT scenes were also included in the classification to observe
phenological changes that occurred during the dry season. To measure the spatial heterogeneity of a given variable,
the classification also included the standard deviation within nine-pixel neighborhoods for each input.
ASPRS 2008 Annual Conference
Portland, Oregon April 28 - May 2, 2008
Figure 2. Study area with sample sites overlaid on the February SPOT 4 image.
(includes material © CNES 2007, Distribution Spot Image S.A., France SICORP, USA, all rights reserved)
Table 1. Field identification criteria
Dambo Class
Identification Criteria
Upland
Surface soil Chroma ≥ 4 and Hue ≤ 7.5 YR,
or Chroma ≥ 3 and Hue ≤ 5 YR
Margin Slope
>
2%
Floor
NOT Upland, Margin, or Bottom
Bottom Hummocky
microtopography
ASPRS 2008 Annual Conference
Portland, Oregon April 28 - May 2, 2008
Digital elevation model (DEM) data obtained from the Shuttle Remote Topography Mission (SRTM) provided
the base topographic input for the classification. Prior to resampling, these 90-meter resolution DEM data were also
used to calculate percent slope and a series of ranked relative elevation layers. Relative elevation was calculated as
the rank of a given pixel within a moving window, with window sizes from 11 to 201 DEM pixels (roughly 1 km to
18 km). Inclusion of these relative elevation layers was intended to improve dambo classification by distinguishing
topographic variation at multiple spatial scales.
Finally, northing and easting values from the projected coordinate system, Universal Transverse Mercator
(UTM), were inputs to the classification. These coordinates were chosen to serve as a proxy for climatic gradients,
particularly precipitation, which vary spatially across the study area.
The field samples for each dambo class were then randomly subdivided into training and accuracy assessment
sets for the classification. As is typical for a supervised classification, the training set provided the locations for
which pixel values were extracted from corresponding areas of each input variable. These extracted values were
evaluated using the “tree” package (Ripley, 2006) developed for the statistical software R to construct a binary
decision tree (BDT) classifier (Figure 3).
BDTs consist of a series of “nodes” and “branches,” composed of classification input variables and connections,
respectively. As pixels from the input layers progress through the tree, they are sorted at the nodes by comparing
their value for a given variable with a threshold value established by a decision rule. They eventually arrive at a
terminal node and are thereby assigned to a specific class. In the case of the dambo wetland classifier, uplands were
identified exclusively by relative elevation. Slope, SPOT bands, vegetation indices, and UTM northing values were
used to classify bottoms, floors, and margins.
Table 2. Classification input variables
(*denotes decision tree node)
Multispectral Inputs
Topographic Inputs
SPOT 10 Dec 06
SRTM Elevation Data
Band 1: green
DEM
Band 2: red
Slope*
Band 3: near infrared
Relative elevation 11
Band 4: short-wave infrared
Relative elevation 21*
SPOT 21 Feb 07
Relative elevation 31
Band 1: green
Relative elevation 41
Band 2: red
Relative elevation 51
Band 3: near infrared
Relative elevation 61
Band 4: short-wave infrared*
Relative elevation 71
Vegetation Indices
Relative elevation 81
NDVI Dec
Relative elevation 91
NDVI Feb
Relative elevation 101
NDVI difference*
Relative elevation 111
NDII Dec
Relative elevation 121
NDII Feb*
Relative elevation 131
NDII difference
Relative elevation 141
Spectral Mixture Analysis
Relative elevation 151
PV Dec
Relative elevation 161
NPV Dec
Relative elevation 171
Shade Dec
Relative elevation 181*
PV Feb
Relative elevation 191
NPV Feb
Relative elevation 201
Shade Feb
UTM Coordinates
PV difference
Northing*
NPV difference
Easting
Shade difference
ASPRS 2008 Annual Conference
Portland, Oregon April 28 - May 2, 2008
Figure 3. Binary decision tree classifier.
RESULTS
The BDT classification resulted in a reasonably accurate dambo class map from a qualitative perspective
(Figure 4). In general, the mapped dambo classes corresponded with the spatial patterns described in the literature
and observed in the field. Bottoms were identified in the central, lowest portions of larger drainages, with floors,
margins, and uplands surrounding them in succession. The classification was less accurate beyond the boundaries of
the study area where field samples were not obtained. In these areas, floors and margins tended to be overclassified.
Within the wider Lugogo, Towa (center left), and Mayanja River channels, bottoms were occasionally misclassified
as upland or margin. In quantitative terms, the classification resulted in an overall accuracy of 75.5% and a Kappa
coefficient of 0.67. Subdividing the accuracy assessment set in a confusion matrix (Table 3) revealed that uplands
were the most readily identifiable class. Confusion with adjacent classes was the most common error among all of
the classes.
Table 3. Pixel-level confusion matrix for BDT classification
Ground Truth
Bottom Floor Margin Upland
Bottom
520
142 88 22
67.36%
Floor
130
1002
232 0
73.46%
Margin
56 258 1159
294 65.59%
Upland
0 0 102
1396
93.19%
Classifie
d
73.65% 71.47% 73.31% 81.54%
Producer’s
Accuracy
User’s Accuracy
ASPRS 2008 Annual Conference
Portland, Oregon April 28 - May 2, 2008
Figure 4. Dambo class map of study area.
DISCUSSION
Given the relationship between topography and hydrology, it was not surprising to learn that elevation data
would play a major role in this landscape classification. Uplands were identified only by elevation, with both fine-
and coarse-scale relative elevation being selected for the decision tree. DEM-derived slope was used to distinguish
the relatively steep margin class from the flatter bottoms and floors. Future work could benefit from the use of
higher-resolution DEM data. Were these data available, the accuracy of the classification would likely increase,
while the need for additional optical remote sensing inputs might decrease.
ASPRS 2008 Annual Conference
Portland, Oregon April 28 - May 2, 2008
ASPRS 2008 Annual Conference
Portland, Oregon April 28 - May 2, 2008
With the relatively coarse elevation data used in this study, however, vegetation-related variables served as
important inputs to the classification. Drier margins were distinguished from wetter bottoms by the phenological
change detected in the NDVI between December and February. Shortwave Infrared (SWIR) reflectance, the fourth
band provided by SPOT 4 imagery, appears to be important for discriminating floors from bottoms, with two SWIR
variables selected for the decision tree. NDII, also a measure of water content that was used to distinguish floors
from bottoms, is calculated using the Near Infrared (NIR) and SWIR bands. The SWIR band may have been
important because it experiences large changes in reflectance due to grass senescence. The UTM northing input’s
significance may be due to climatic variation within the study area. A marked latitudinal difference in green
vegetation and standing water was observed while conducting field work. This variation may be linked to
precipitation gradients and a general decrease in elevation moving from south to north in the study area.
The results of this first-of-its-kind classification of dambo wetlands were promising. However, the classifier
did not appear to perform as well outside of the study area where no sampling occurred. Further refinement of the
classifier will likely be required if it is to be applied in other areas. Due to the range of vegetation and topography
found in the areas of dambo occurrence, it may not be possible to develop a universal dambo wetland classification
technique. In spite of this fact, this classification delineated dambo extents within the study area accurately enough
to proceed with the intended methane emissions measurement and modeling. These efforts will provide an estimate
of dambo contributions to global methane emissions. In addition, this work will add to the understanding of the
interplay between dambo methane production, hydroclimate, and anthropogenic land cover change.
ACKNOWLEDGEMENT
This material is based upon work supported by the National Science Foundation under Grant No. 0620206.
REFERENCES
Acres, B.D., Rains, A.B., King, R.B., Lawton, R.M., Mitchell, A.J.B. and Rackham, L.J., 1985. African dambos:
their distribution, characteristics and use, In Dambos: small channelless valleys in the tropics (M.F.
Thomas and A. S. Goudie, Eds.), Zeitscrift für Geomorphologie, 52: 63-86.
Adams, J.B., Smith, M.O., and Gillespie, A.R., 1993. Imaging spectroscopy: Interpretations based on spectral
mixture analysis, In Remote Geochemical Analysis: Elemental and Mineralogical Composition (C.M.
Pieters and P.A. Englert, Eds.), Cambridge University Press: Cambridge, UK, pp. 145-166.
Bartlett, K.B. and Harriss, R.C., 1993. Review and assessment of methane emissions from wetlands, Chemosphere,
26: 261-320.
Bullock, A., 1992. Dambo hydrology in Southern Africa - Review and reassessment, Journal of Hydrology, 134:
373-396.
Hunt, E.R. and Rock, B.N., 1989. Detection of changes in leaf water content using near- and middle-infrared
reflectances, Remote Sensing of Environment, 30: 43-54.
Faulkner, S., 2004. Soils and sediment: understanding wetland biogeochemistry, In Wetlands (Spray, S.L. and
McGlothlin, K.L., Eds), Rowman & Littlefield: Lanham, MD, pp. 30-54.
Mäckel, R., 1974. Dambos: a study in morphodynamic activity on the plateau regions of Zambia, Catena, 1: 327-
365.
Mäckel, R., 1985. Dambos and related landforms in Africa – an example for the ecological approach to tropical
geomorphology, In Dambos: small channelless valleys in the tropics (M.F. Thomas and A. S. Goudie,
Eds.), Zeitscrift für Geomorphologie, 52: 63-86.
Ripley, B.D., 2006. The tree package (Version 1.0-24) [Computer software]. The R Project for Statistical
Computing. Retrieved September 20, 2006. Available from
http://www.r-project.org/
.
Rodhe, H., 1990. A comparison of the contribution of various gases to the greenhouse effect, Science, 248: 1217-
1219.
Rouse, J.W., Haas, R.H., Schell, J.A., and Deering, D.W., 1973. Monitoring vegetation systems in the Great Plains
with ERTS, Third ERTS Symposium, NASA SP-351, 1, 309-317.
Shindell, D.T., Walter, B.P. and Faluvegi, G., 2004. Impacts of climate change on methane emissions from
wetlands, Geophysical Research Letters, 31: L21202.
Von der Heyden, C.J., 2004. The hydrology and hydrogeology of dambos: A review, Progress in Physical
Geography, 28: 544-564.
Document Outline - Next Page
- Previous Page
- ========================
- Table of Contents
- Author Index
- Exhibitors
- Copyright
- ============================
- Print
Dostları ilə paylaş: |