Mapping North African landforms using continental scale unmixing of MODIS imagery John-Andrew C. Ballantine a, * , Gregory S. Okin b , Dylan E. Prentiss a , Dar A. Roberts a a Department of Geography, University of California Santa Barbara, EH3611, Santa Barbara, CA 93106, USA b Department of Environmental Sciences, University of Virginia, Charlottesville, VA 22904-4123, USA Received 14 December 2004; received in revised form 25 April 2005; accepted 29 April 2005 Abstract We describe the production of a landform map of North Africa utilizing moderate resolution satellite imagery and a methodology that is applicable for sub-continental to global scale landform mapping. A mosaic of Moderate Resolution Imaging Spectroradiometer (MODIS) apparent surface reflectance imagery was compiled for Africa north of 10- N. Landform image endmembers were chosen to characterize ten different types of vegetated and unvegetated desert surfaces: alluvial complexes, dunes, dry and ephemeral lakes, open water, basaltic volcanoes and flows, mountains, regs, stripped, low-angle bedrock surfaces, sand sheets, and Sahelian vegetation. Multiple Endmember Spectral Mixture Analysis (MESMA) was applied to the MODIS mosaic to estimate landform and vegetation endmember fractions. The major landform in each MODIS pixel was identified based on the majority endmember fraction in two- or three-endmember models. Accuracy assessment was conducted using two data sources: the historic Landform Map of North Africa [Raisz, E. (1952). Landform Map of North Africa. Environmental Protection Branch, Office of the Quartermaster General.] and Landsat Thematic Mapper (TM) data. Comparison with the Raisz landform map gave an overall classification accuracy of 54% with significant confusion between alluvial surfaces and regs, and between sandy and clayey surfaces and dunes. A second validation using 20 Landsat images in a stratified sampling scheme gave a classification accuracy of 70%, with confusion between dunes and sand sheets. Both accuracy assessment schemes indicated difficulty in vegetation classification at the margin of the Sahel. A comparison with minimum distance and maximum likelihood supervised classifications found that the MESMA approach produced significantly higher classification accuracies. This digital landform map is of sufficiently high quality to form the basis for geomorphic studies, including parameterization of the surface in global and regional dust models. D 2005 Elsevier Inc. All rights reserved. Keywords: Sahel; Sahara; Landform; Desert; Dust source; MESMA; Landform endmembers; Wind erosion 1. Introduction The Sahara Desert is the largest in the world and arguably the most diverse in terms of the range of landforms found within it. Tucker and Nicholson (1999) found the mean size of the Sahara Desert, between 1980 and 1997, was 9,149,000 km 2 . The size of this region makes the traditional mapping methods of aerial photography and field surveying of limited use. Moderate resolution multispectral data allows continental- to global-scale mapping of the Earth’s surface while retaining sufficient resolution for geomorphic and ecological studies. In this paper we describe the production of a landform map of North Africa (Fig. 1) using data from the moderate resolution imaging spectroradiometer (MODIS) reflectance products and a modified spectral mixture analysis (SMA) approach. The principal purpose of creating a landform map is to provide a basis for geomorphic studies and providing greater insight into the landforms in a zone that is often referred to as ‘‘barren’’ in land cover studies. SMA models each pixel in an image as a mixture between a landform endmember (spectrum), vegetation 0034-4257/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2005.04.023 * Corresponding author. Tel.: +1 207 299 3703. E-mail address: [email protected] (J.-A.C. Ballantine). Remote Sensing of Environment 97 (2005) 470 – 483 www.elsevier.com/locate/rse
14
Embed
Mapping North African landforms using continental scale ... · Mapping North African landforms using continental scale unmixing ... The major landform in each MODIS ... but did not
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
www.elsevier.com/locate/rse
Remote Sensing of Environm
Mapping North African landforms using continental scale unmixing
of MODIS imagery
John-Andrew C. Ballantine a,*, Gregory S. Okin b, Dylan E. Prentiss a, Dar A. Roberts a
aDepartment of Geography, University of California Santa Barbara, EH3611, Santa Barbara, CA 93106, USAbDepartment of Environmental Sciences, University of Virginia, Charlottesville, VA 22904-4123, USA
Received 14 December 2004; received in revised form 25 April 2005; accepted 29 April 2005
Abstract
We describe the production of a landform map of North Africa utilizing moderate resolution satellite imagery and a methodology that
is applicable for sub-continental to global scale landform mapping. A mosaic of Moderate Resolution Imaging Spectroradiometer
(MODIS) apparent surface reflectance imagery was compiled for Africa north of 10- N. Landform image endmembers were chosen to
characterize ten different types of vegetated and unvegetated desert surfaces: alluvial complexes, dunes, dry and ephemeral lakes, open
water, basaltic volcanoes and flows, mountains, regs, stripped, low-angle bedrock surfaces, sand sheets, and Sahelian vegetation. Multiple
Endmember Spectral Mixture Analysis (MESMA) was applied to the MODIS mosaic to estimate landform and vegetation endmember
fractions. The major landform in each MODIS pixel was identified based on the majority endmember fraction in two- or three-endmember
models. Accuracy assessment was conducted using two data sources: the historic Landform Map of North Africa [Raisz, E. (1952).
Landform Map of North Africa. Environmental Protection Branch, Office of the Quartermaster General.] and Landsat Thematic Mapper
(TM) data. Comparison with the Raisz landform map gave an overall classification accuracy of 54% with significant confusion between
alluvial surfaces and regs, and between sandy and clayey surfaces and dunes. A second validation using 20 Landsat images in a stratified
sampling scheme gave a classification accuracy of 70%, with confusion between dunes and sand sheets. Both accuracy assessment
schemes indicated difficulty in vegetation classification at the margin of the Sahel. A comparison with minimum distance and maximum
likelihood supervised classifications found that the MESMA approach produced significantly higher classification accuracies. This digital
landform map is of sufficiently high quality to form the basis for geomorphic studies, including parameterization of the surface in global
products and a modified spectral mixture analysis (SMA)
approach. The principal purpose of creating a landform map
is to provide a basis for geomorphic studies and providing
greater insight into the landforms in a zone that is often
referred to as ‘‘barren’’ in land cover studies.
SMA models each pixel in an image as a mixture
between a landform endmember (spectrum), vegetation
ent 97 (2005) 470 – 483
10oW
30oN
20oN
30oN
20oN
0oE 10oE 20oE 30oE 40oE
10oW 0oE 10oE 20oE 30oE 40oE
Lake
Study Area
Country
River
0 337.5 675 1,350 2,025 2,700Kilometers
Legend
Morocco
WesternSahara
Mauritania
Senegal
Mali
Algeria
Niger
Libya
Chad Sudan
Egypt
Tunisia
Niger River
Lake Chad
Lake Nasser
Nile River
N
Fig. 1. A map of the study area covering North Africa from 20- W to 40- E and 10- N to 40- N. Political boundaries and major water bodies are included as
references for locations described in the paper.
J.-A.C. Ballantine et al. / Remote Sensing of Environment 97 (2005) 470–483 471
endmember (if it is present and detectable) and shade, an
endmember included to account for shadows and variable
regs, and sandsheets) based on the classes identified in the
Raisz map. The spectra from each broad class were
statistically divided into two to three subclasses using a k-
means unsupervised cluster analysis (Funk et al., 2001). As
an example, the broad mountain class had three statistically
separate subclasses, two of which were identified as
representing mountains, and the other identified as repre-
senting basaltic volcanoes and flows.
Tompkins et al. (1997) emphasized the importance of the
selection of good spectral endmembers for any SMA.
Because of the diversity of classes in this study, we chose
the endmember average RMS (EAR) method of Dennison
and Roberts (2003) as the principal method for selecting
landform endmembers. The EAR technique selects the
endmember that is most representative of a class of
endmembers to represent that class. EAR worked well for
the broadly defined classes of this study where finding
extreme endmembers might not have represented the full
diversity of each landform class.
In order to identify the landform spectrum that was most
distinct, the spectral library of 109 landform endmembers
was unmixed by two endmember models of shade and each
of the other endmembers in the library to calculate EAR
values as described by Dennison and Roberts (2003). The
model constraints described by Dennison and Roberts were
found to work well for these data. Endmember fractions
were constrained to less than 106% with best-fit models
greater than this value being set to 106% and the RMSE
calculated from this value.
The EAR method selected an endmember representative
for a subclass by finding the endmember with the lowest
RMSE when modeling other endmembers in its subclass:
EARAi;A ¼
Xnj¼1
RMSEAi;Aj
n� 1ð4Þ
where A was the subclass, Ai was the landform endmember
in question, n was the number of spectra in subclass A, and
Aj was the spectrum being modeled by Ai. Thus, the
endmember representing basaltic flows in the Ahaggar
Mountains of Algeria was used to model the other basalt
endmembers (basalt is subclass M1 in Fig. 3a). The average
of the RMSE values from each model was the EAR value
for the Ahaggar basaltic endmember. Because the Ahaggar
basaltic endmember’s EAR value was lower than that of the
Tibesti Mountains and the other basalt endmembers, the
Ahaggar basaltic endmember was picked as the representa-
tive endmember for subclass M1. The landform spectra thus
picked are shown in Fig. 3a.
3. Modeling with MESMA
The modeling of the image with MESMA followed the
methodology of Dennison and Roberts (2003). Two-
endmember models (an endmember and shade) were run
with the constraint that non-shade fractions had to be
between �6% and 106% of the pixel. Cases where
residuals exceeded 2.5% of reflectance for more than 7
contiguous bands or RMSE exceeded 2.5% of reflectance
were left unmodeled. Similar constraints were used for
three endmember models (two endmember spectra and
shade). All possible models were considered such that
there were 24 possible two-endmember models (17 land-
form subclasses and 7 vegetation endmembers) and 276
three-endmember models. For each pixel, the lowest
RMSE two-endmember model was chosen unless the
RMSE of the three-endmember improved upon that of
Fig. 4. The shade endmember fraction image derived from the MODIS mosaic in Fig. 2. The shading bar shows % shade from �6% to 100% with values of
�100 representing bright desert surfaces with effectively no shade. Bright areas near the coast indicate deep water that has not been masked out. Note that the
shade fraction is greater on the east-side of each MODIS tile, expressing the BRDF of vegetation.
J.-A.C. Ballantine et al. / Remote Sensing of Environment 97 (2005) 470–483476
the two endmember model by more than 0.8% of
reflectance (Roberts et al., 2003).
4. Results
MESMA produced fraction images of the dominant and
secondary (in three endmember cases) endmembers, along
with the fraction of shade present. Each pixel was coded
with the classes that describe it and the fraction of each
class.
4.1. Shade fraction
In Fig. 2, it is apparent that the east side of each MODIS
tile in the mosaic is darker than the west, particularly in
vegetated regions. This occurs because the bi-directional
reflectance distribution function (BRDF) of surface materi-
als is asymmetrical with more light being back-scattered
than forward-scattered in the case of vegetation. The
variation in brightness across the tile is a function of the
amount of shadowing imaged by the sensor. An advantage
Fig. 5. The vegetation endmember fraction derived from the MODIS mosaic in Fig
of the transects shown in Fig. 6 are indicated.
of using a spectral unmixing approach is that the shade
fraction contains most of this BRDF effect. As a result, the
landform and/or vegetation endmember(s) used to determine
the landform class did not exhibit any variation due to
BRDF effects. The expression of the anisotropic BRDF of
vegetation is apparent in the shade fraction image (Fig. 4).
4.2. Vegetation fraction
The modeled vegetation endmember fraction represents
the fractional ground cover of vegetation (Fig. 5). Fig. 5
expresses the vegetation cover increase from the Sahara
south into the Sahel. Vegetation along the northern margin
of the continent is also apparent. Scattered vegetation in the
core of the Sahara is usually associated with mountain
ranges where orographic rainfall and springs make signifi-
cant vegetation viable. If multi-temporal imagery had been
used, the vegetation fraction could tracked be to show
changes in vegetation cover and its response to rainfall
variation.
The red stems vegetation endmember dominated the
scene because it best modeled Sahel vegetation (Table 1). In
. 2. The shading bar shows percent vegetation cover. Approximate locations
Table 1
Areal coverage and proportion of vegetation fractions in the Fig. 5
Class # Name Area (km2) % Vegetation
18 Dry grass 92,886.75 4.89
19 Plant litter 28,845.25 1.52
20 Red stems 1,568,183.8 82.48
21 Bursage 180.75 0.01
22 Big leaf sage 1226 0.06
23 Rabbit brush 27,119 1.43
24 Image riparian 182,950 9.62
Class # is relative to all endmembers used in this study. Class name
corresponds to those used in Fig. 3b. The ‘‘% Vegetation’’ column describes
the percentage of all vegetated pixels covered by the class in question.
J.-A.C. Ballantine et al. / Remote Sensing of Environment 97 (2005) 470–483 477
terms of areal coverage, the image riparian vegetation
endmember was the next most common endmember
because of its occurrence in a few dense areas including
the Niger Delta, Lake Chad, and along the Nile River and
into the Nile Delta. The elevated vegetation in southern
Libya (Fig. 5) classified as dry grass and was mixed with the
reg class.
Variation in vegetation cover across the Sahara is further
illustrated by the three transects shown in Fig. 6. All three
transects clearly show the steady rise in vegetation from
north to south in the Sahel (towards the right in the shaded
zone at the right of Fig. 6). At longitude 4.5- W, there is a
zone of higher vegetation cover along the northern margin
of the Sahel because of the vigorous vegetation of the Niger
Delta in southern Mali. North of that, there is very little
vegetation across the Sahara until one reaches the Atlas
Fig. 6. Transects of vegetation fraction oriented north–south across the vegetation
north edge of the image to the south, and show locations of high and low vegetat
shaded squares (16- E), and black triangles (30- E). The Sahel and Nile River and D
arrows.
Mountains in Morocco. The high vegetation fraction at
about 3000 km south in the 16- E transect indicates the
location of Lake Chad. This transect shows an elevated
vegetation fraction in the heart of the Sahara where it
crosses the western side of the Tibesti Mountain range. At
the northern end of the transect, there is a notable vegetated
zone extending from southern Libya to the coast, possibly
representing the Fezzan region which has historically been a
fertile pastoral zone (Bovill, 1968). This agrees with the
selection of dry grass as the endmember for this region.
Traversing the 30- E transect northward from the Sahel, one
finds little vegetation until reaching the Nile River and its
floodplain in the center of Fig. 6. The 30- E transect follows
the northward course of the Nile into the Nile Delta from
this point, showing elevated vegetation cover.
4.3. Classification
In many cases, a two endmember model was adequate to
describe the spectral response of the pixel, in which case the
pixel was labeled as the dominant class. In cases where
shade was the majority class, the pixel was labeled as
unclassified. These high-shade unclassified pixels occurred
in mountainous or basaltic regions, in heavily vegetated
regions of the Sahel, and over open water. Should any these
classes be important to a given user, the class could be
assigned manually.
Some very bright areas were also unclassified because
the brightness of the pixels exceeded 106% of the brightest
fraction image shown in Fig. 5. Samples are taken every 2500 m from the
ion cover. Points on each transect are symbolized by open circles (4.5- W),
elta are shaded and other zones of elevated vegetation fraction are noted by
Table 2
Distribution of landform classes in classified image
Class Cover fraction Color
Alluvial 0.14 Olive
Dunes 0.22 Yellow
Open water 0.01 Blue
Lakebed 0.01 Cyan
Basalt 0.01 Purple
Mountain 0.09 Brown
Reg 0.21 Red
Bedrock 0.03 Magenta
Sandsheet 0.15 Orange
Vegetation 0.13 Green
J.-A.C. Ballantine et al. / Remote Sensing of Environment 97 (2005) 470–483478
endmember (D1 in Fig. 3b). Some pixels were brighter
than the brightest endmember because the EAR endmem-
ber selection technique picked endmembers that were
representative of classes, as opposed to the most extreme
in the image. The most prominent case was in the Tenere
Desert of Niger where active dunes had a very high
albedo. This unclassified area was small and encompassed
by the dune class and therefore could easily be corrected
by hand.
In cases where the three endmember model was chosen,
the class with the greatest fraction was considered the
landform class but the relative fractional abundances were
retained. This was particularly useful in the case where
vegetation was the secondary class because a percentage
vegetation cover value could be determined (Elmore et al.,
2000).
4.4. Landform map
The results of the landform classification are shown in
Table 2 and Fig. 7. Dunes and regs were the dominant
landform classes (22% and 21% cover respectively). Dunes
are areas of high sediment availability, limited by transport
capacity, whereas regs are areas of potentially high sediment
storage that have low sediment availability due to the
armoring of the surface (Kocurek & Lancaster, 1999). The
Fig. 7. The landform map of North Africa produced through MESMA applie
(For interpretation of the references to colour in this figure legend, the reader is
sandsheet class is in an intermediate position on this
sediment availability continuum. Although lakebeds and
alluvial complexes are affected by wind erosion and
deposition, they are also heavily influenced by fluvial
activity. Other classes are less affected by the wind transport
system. All of the mosaic, except for very dark and very
bright regions, was classified.
All of the vegetation classes, including NPV, were
lumped into a single vegetation class for the map. The
dominant component of the vegetation class was red stems,
with approximately an order of magnitude more coverage
than any other vegetation type. The dominance of the
vegetation class by an NPV endmember was an indication
both of the dominance of woody vegetation in the Sahel,
and the fact that these images were taken during the
Sahelian dry season.
Some classes appeared in particular regions. The alluvial
class largely occurred in the north in Algeria and from the
coast to the Fezzan region of Libya. This may have been a
result of distinct lithologies in this region (e.g. limestone
bedrock), the more frequent occurrence of alluvial fans
along the flanks of the Atlas Mountains, or the possibility
that the alluvial endmember contained some fraction of
vegetation which was stronger in the north during Novem-
ber and December (the beginning of the Mediterranean wet
season). Both the alluvial and mountain classes appeared in
the transition zone between the Sahel and the Sahara. There
is no mountain range in this zone and this effect will be
discussed in the next section.
With the exception of the band of the mountain class in
the northern Sahel, most occurrences of the mountain and
basalt classes coincided with well known mountain ranges
and basaltic formations. Dry lakebeds were largely confined
to zones in the Bodele Depression of Chad, the coastal
sabkhas of Mauritania, and the ephemeral chotts of Tunisia
and northern Algeria.
Another notable regional tendency was that the stripped
bedrock surfaces occurred in the plateaus of the western
Sahara and along the coast of the Red Sea in the east. The
d to the MODIS mosaic in Fig. 2. Unclassified areas appear in black.
referred to the web version of this article.)
Table 3
Error matrix showing reference classes from the Raisz landform map on the
x-axis and modeled classes from the MODIS-derived landform map on the