Urban vegetation mapping using sub-pixel analysis and expert system rules: A critical approach S. W. MYINT* Department of Geography, Arizona State University, 600 E. Orange Street, SCOB Building, Room 330, Tempe, AZ 85287-0104 (Received 29 September 2004; in final form 1 December 2005 ) Since the traditional hard classifier can label each pixel only with one class, urban vegetation (e.g. trees) can only be recorded as either present or absent. The sub- pixel analysis that can provide the relative abundance of surface materials within a pixel may be a potential solution to effectively identifying urban vegetation distribution. This study examines the effectiveness of a sub-pixel classifier with the use of expert system rules to estimate varying distributions of different vegetation types in urban areas. The Spearman’s rank order correlation between the vegetation output and reference data for wild grass, man-made grass, riparian vegetation, tree, and agriculture were 0.791, 0.869, 0.628, 0.743, and 0.840 respectively. Results from this study demonstrated that the expert system rule using NDVI threshold procedure is reliable and the sub-pixel processor picked the signatures relatively well. This study reports a checklist of the sources of limitation in the application of sub-pixel approaches. 1. Introduction Vegetation influences urban environmental conditions and energy fluxes by selective reflection and absorption of solar radiation (Gallo et al. 1993) and by function of evapotranspiration (Owen et al. 1998). The presence and abundance of vegetation in urban areas has long been recognized as a strong influence on energy demand and development of the urban heat island (Oke 1982, Huang et al. 1987). Urban vegetation abundance may also influence air quality and human health (Wagrowski and Hites 1997) because trees make their own food from carbon dioxide in the atmosphere, sunlight, water, and a little amount of soil elements, and release oxygen in the process. They also provide surface area for sequestration of particulate matter and ozone. The loss of trees in our urban areas not only intensifies the urban heat island effect due to the loss of shade and evaporation, but we lose a principal absorber of carbon dioxide and trapper of other pollutant as well. A noticeable phenomenon that has arisen as a result of urbanization is that urban climates are warmer and more polluted than their rural environments (Lo and Quattrochi 2003). Urban development increases the amount of impervious surfaces in watersheds as farmland, forests, and meadows are converted into buildings, driveways, pavements, roads, and car parks with virtually no ability to absorb storm water. Surrounding urban environments (i.e., forests, grasslands, agriculture, water, etc.) are also very important because the decisions that need to be made regarding planning *Corresponding author. Email: [email protected]International Journal of Remote Sensing Vol. 27, Nos. 12–14, July 2006, 2645–2665 International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online # 2006 Taylor & Francis http://www.tandf.co.uk/journals DOI: 10.1080/01431160500534630
21
Embed
Urban vegetation mapping using sub-pixel analysis and ...smyint/publications/Myint-IJRS-Subpixel-July2… · Urban vegetation mapping using sub-pixel analysis and expert system rules:
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
Urban vegetation mapping using sub-pixel analysis and expert systemrules: A critical approach
S. W. MYINT*
Department of Geography, Arizona State University, 600 E. Orange Street, SCOB
Building, Room 330, Tempe, AZ 85287-0104
(Received 29 September 2004; in final form 1 December 2005 )
Since the traditional hard classifier can label each pixel only with one class, urban
vegetation (e.g. trees) can only be recorded as either present or absent. The sub-
pixel analysis that can provide the relative abundance of surface materials within
a pixel may be a potential solution to effectively identifying urban vegetation
distribution. This study examines the effectiveness of a sub-pixel classifier with
the use of expert system rules to estimate varying distributions of different
vegetation types in urban areas. The Spearman’s rank order correlation between
the vegetation output and reference data for wild grass, man-made grass, riparian
vegetation, tree, and agriculture were 0.791, 0.869, 0.628, 0.743, and 0.840
respectively. Results from this study demonstrated that the expert system rule
using NDVI threshold procedure is reliable and the sub-pixel processor picked
the signatures relatively well. This study reports a checklist of the sources of
limitation in the application of sub-pixel approaches.
1. Introduction
Vegetation influences urban environmental conditions and energy fluxes by selective
reflection and absorption of solar radiation (Gallo et al. 1993) and by function of
evapotranspiration (Owen et al. 1998). The presence and abundance of vegetation in
urban areas has long been recognized as a strong influence on energy demand and
development of the urban heat island (Oke 1982, Huang et al. 1987). Urban
vegetation abundance may also influence air quality and human health (Wagrowski
and Hites 1997) because trees make their own food from carbon dioxide in the
atmosphere, sunlight, water, and a little amount of soil elements, and release oxygen
in the process. They also provide surface area for sequestration of particulate matter
and ozone. The loss of trees in our urban areas not only intensifies the urban heat
island effect due to the loss of shade and evaporation, but we lose a principal
absorber of carbon dioxide and trapper of other pollutant as well. A noticeable
phenomenon that has arisen as a result of urbanization is that urban climates are
warmer and more polluted than their rural environments (Lo and Quattrochi 2003).
Urban development increases the amount of impervious surfaces in watersheds as
farmland, forests, and meadows are converted into buildings, driveways, pavements,
roads, and car parks with virtually no ability to absorb storm water. Surrounding
urban environments (i.e., forests, grasslands, agriculture, water, etc.) are also very
important because the decisions that need to be made regarding planning
community growth include where to locate new residential areas, transportation
infrastructure, new retailers, school catchment zones (Mesev 2003), emergency
management systems, industrial zones, public offices, commercial areas and how to
reduce and monitor air pollution, noise pollution, water pollution, soil erosion,
deforestation, land degradation, urban heat island effects, crime pattern and rate,
disaster risk, and traffic congestion. Urbanization alters the natural ways energy
flows through the atmosphere, land, and water systems. The modification of the
urban landscape influences the local (microscale), mesoscale, and even the
macroscale climate (NASA/GHCC Project Atlanta). Hence, the spatiotemporal
distribution of vegetation is generally considered a key component of the urban
environment.
Identification of urban land use and land covers from remotely sensed images has
usually been based on the hard classification of spectral response from image pixels.
The brightness value of each pixel represents either one homogeneous land cover or
the combination of a number of different land-cover classes. Since the traditional
hard classifier can label each pixel only with one class, urban vegetation (e.g. trees)
can only be recorded as either present or absent. Information on the percentage
distribution of spatially mixed spectral signatures from different ground-cover
features is not possible with the per-pixel classifiers. Hence, the traditional
classification of mixed pixels may lead to information loss (Wang, 1990),
degradation of classification accuracy, and degradation of modelling quality in
successive applications (Ji and Jensen 1996, 1999).
The sub-pixel analysis that can provide the relative abundance of surface
materials within a pixel may be a potential solution to per-pixel classifiers especially
when dealing with medium to coarse resolution satellite images (e.g. Landsat TM,
MODIS, AVHRR). There have been several approaches to sub-pixel analysis—
linear mixture models (Smith et al. 1990, Settle and Drake 1993, Van der Meer 1997,
Wu and Murray 2003, Rashed et al. 2003), Bayesian probabilities (Wang 1990a,
Wang 1990b, Foody et al. 1992, Eastman and Laney 2002, Hung and Ridd 2002),
neural network (Foody and Aurora 1996, Zhang and Foody 2001); fuzzy c-means
methods (Fisher and Pathirana 1990, Foody and Cox 1994, Foody 2000), and fuzzy
set possibilities (Eastman 1999).
This study aims to determine the effectiveness of the IMAGINE sub-pixel
classifier with the use of expert system rules to quantify varying amounts and
distributions of different vegetation types in urban and suburban areas using
Landsat TM data.
2. IMAGINE sub-pixel analysis
The sub-pixel processing tool used in this study is a sub-pixel classifier, an add-on
module to ERDAS IMAGINE geographic imaging package. The tool is intended to
quantify materials that are smaller than image resolution. IMAGINE sub-pixel
processor is based on the concept reported by Schowengerdt (1995) that the spectral
reflectance of the majority of the pixels in remotely sensed image data is assumed to
be a spatial average of spectral signatures from two or more surface categories.
Hence, the brightness value of a pixel in an urban image can be considered a
combination of spectral response from multiple materials, such as grass, trees,
shrubs, cement roads, tarmac roads, metal roofs, wooden roofs, pavements,
driveways, and car parks.
2646 S. W. Myint
The sub-pixel processor is designed to classify each pixel in an image as itsfraction of material of interest (MOI) present. For example, if the MOI is trees, each
pixel in the image will hold a number from 0 to 1.0 representing the fraction of trees
within the pixel. The procedure to obtain the proportion of trees in each pixel is
explained below.
Following Huguenin et al. (1997), it is assumed that the total spectral response of
each urban image pixel, Am, can be separated into a component of trees, C, and a
background component, Bm, of all other materials (e.g. grass, cement parking). In
figure 1, if trees were the material of interest (C) and if the fraction of C, fm, is equal
to 33%, then the fraction of the rest of the materials (Bm) would be (12fm)50.67. Itshould be noted that in figure 1, C represents a single specified material of interest
(e.g. trees) whereas Bm refers to all other materials in the pixel, representing a single
set of combined materials of the background.
Am~ fmCð Þz 1{fmð ÞBmf g ð1Þ
Assuming C and Bm are optically thick, meaning no transmittance through thematerial in at least one of the spectral bands, the radiance contributions from C and
Bm can be considered linearly additive in spectral bands, n.
where, Xm[n], Z[n], and Ym[n] are radiances from, Am, C, and Bm in pixel m and bandn, respectively. The fraction of spectral radiance contributed by Z[n] (MOI) in pixel
m and band n is gm[n]. It should be noted that the radiant fraction, gm[n], can vary
from band to band, because the spectral signatures from the MOI and the
background material can vary from band to band since different spectral bands have
different spectral sensitivity to different materials. If there is negligible difference in
spectral response between the material of interest and the background for all
multispectral bands, gm[n] will be approximately equal to fm.
The sub-pixel processor detects the MOI (e.g. trees) in each urban pixel by
iteratively subtracting fractions of candidate background spectra. The set ofcandidate background, Ym[n], is unique for each pixel in the image and is
independently selected for each pixel based on the assumption that the background
for the pixel under investigation can be represented by other pixels in the same
scene. Then, the processor identifies the background and its fraction that gives the
residual spectrum that most closely matches the spectrum for the material of
Figure 1. Material of interest C (trees) and a background component Bm (all othermaterials) in a pixel of an urban image.
Urban vegetation mapping using sub-pixel analysis and expert system rules2647
interest. The residual, Zm[n], is obtained by using the expression
(0.45–0.52 mm), green (0.52–0.60 mm), red (0.63–0.69 mm), and near infrared (0.76–
0.90 mm) acquired over Norman, Oklahoma, on 20 March 2000 with the aid of
1 : 50,000 scale aerial photographs were used. Both IKONOS and Landsat ETM +data were ortho-rectified. The aerial photographs were not very useful in identifying
accurate ground covers in comparison with the Landsat TM since they were
acquired in 1985. However, field verification was also carried out to supplement the
identification of land cover classes accurately. Several field trips were conducted to
identify uncertain features and classes.
4. Environmental correction
The environmental correction function first converts the brightness value of each pixel
into the true radiance of ground materials. This is done through environmental
correction using the pseudo-calibration materials (e.g. the darkest regions—deep clear
water or terrain shadow) indigenous to the scene. The environmental correction tool
calculates a set of factors to compensate for variations in environmental and
atmospheric conditions during satellite data acquisition. These correction parameters
are output to a file, and used during signature derivation and classification. This step
is required because in order to use equation (2) to search for the materials of interest,
the raw digital numbers for pixel m, DNm[n], need to be corrected to remove the
atmospherically scattered solar radiance component and the sensor offset factors. An
environmental correction tool in the IMAGINE sub-pixel classifier uses sampled
pixels from the scene being processed to derive a correction factor that is subtracted
from DNm[n] to provide the requisite proportionality to Xm[n].
The use of dark pixels, representing deep clear water and shadowed terrain, to
remove atmospherically scattered solar radiance from scene pixels in not at all
unusual. However, the success of the approach is questionable because water pixels
and shadowed pixels generally contain significant unwanted signatures from surface
features, such as reflected sky radiance and sun glints in water pixels and solar
illuminated terrain in shadowed pixels (Huguenin et al. 1997). The sub-pixel
procedure allows a more accurate atmospheric spectrum to be derived by blending
spectra from both reflected sky-radiated and sun-glinted water pixels and solar
illuminated terrain shadow pixels (Applied Analysis 2003). It is expected that the
unwanted glints and illuminated terrains are effectively suppressed, generating a
more accurate atmospheric spectrum.
5. Signature derivation
Even though the use of laboratory-based measurement of pure signatures would be
the optimal approach to sub-pixel classifiers (Wu and Murray 2003), a common
approach for determining pure signatures is to select representative pixels from
homogeneous land covers from satellite images (Rashed et al. 2001, Small 2001,
Eastman and Laney 2002, Hung and Ridd 2002, Wu and Murray 2003). One of the
reasons for the selection of pure signatures from images is to overcome the
substantial problems that exist in correcting atmospheric absorption and scattering
(Settle and Drake 1993). Other reasons may be due to unavailability of laboratory
based reference data, the nature of the landscape under study, and/or classification
specificity.
The signature of the material of interest (MOI) consists of a signature spectrum
and a non-parametric feature space. The signature derivation function generates the
Urban vegetation mapping using sub-pixel analysis and expert system rules2649
signature spectrum and feature space from user-defined training samples and their
parameters. The function allows us to select signatures by either a whole-pixel or
sub-pixel training set. Whole-pixel signatures are signatures derived from training
sample pixels that contain more than 90% of the MOI. It was suggested that a sub-
pixel training approach should be applied only when a whole pixel signature cannot
provide satisfactory accuracy (Applied Analysis 2003). The whole-pixel signature
derivation strategy was employed in this study to identify vegetation distribution in
Norman. The user-specified parameters include the approximate fraction of the
MOI in the training pixels (material pixel fraction) and the estimated probability
that any specified training pixels actually contains the MOI (confidence level). For
all signature samples, we used 0.90 for material fraction and 0.80 for the confidence
level. The signatures that we selected as training samples for the material of interests
include shrubs (Gr1), tall wild grass (Gr2), short wild grass (Gr3), agriculture (Agr),
man-made grass (Gol), riparian vegetation (Rip), and trees (For). By referring to
IKONOS image data and aerial photos with the help of field checks, the above seven
vegetation covers were selected from the TM image and statistics of the samples
were computed using ERDAS IMAGINE software. Figure 3 is a line chart of the
mean brightness values per band for the seven vegetation types. The mean and
standard deviation brightness values by band for the selected covers are shown in
table 1, and figure 4 shows the band 3 vs. band 4 scatter plot of the mean brightness
values of the classes.
Since we found three visually and statistically different signatures in the wild grass
category under investigation (table 1), we identified three different training samples
for wild grass. It should be noted that the grass 2 sample was selected from land
where natural vegetation is purely tall grass. The grass 1 category sample was
selected from a grass-like plant, forbs, or bushes area whereas the grass 3 class
sample was selected from a short wild grass area. We treated man-made grass
vegetation as a separate category instead of combining it with wild grass in this
study, since the spectral response from man-made grass is significantly different
from other wild grasslands. This information could be useful for some specific urban
and environmental planning of the study area.
It is important to note that the sub-pixel tool is designed to identify one material
of interest at a time. In other words, each material of interest was identified in an
Figure 3. Mean brightness values of seven vegetation types. Gr1, shrubs; Gr2, tall wildgrass; Gr3, short wild grass; Agr, agriculture; Gol, man-made grass; Rip, riparian vegetation;For, trees.
2650 S. W. Myint
image independently, i.e. in one analysis MOI was grass, and in another analysis
MOI was trees. However, in some cases, there may be two or more signatures that
represent a material of interest (e.g. signatures of grasses from wild grassland, man-
made grass, and dry rangeland). It was anticipated that classification accuracy could
be improved by using more than one signatures of the same class because different
signatures of the same class could produce more complete identification of the
material of interest. This is partly because the sub-pixel processor uses pure
signatures (.90% of MOI) and does not consider the variance of the training
samples.
6. Classification and expert system rules
The signature combiner tool in the IMAGINE sub-pixel classifier can be used to
combine signatures of the same category. This tool allows us to combine different
signatures to form a signature family (e.g. grass) as well as signatures of different
materials such that they are not in the same family (e.g. grass and trees). However, it
should be noted that signatures of the same family are treated separately during the
classification process. They do not compete with each other as in linear mixture
modelling approaches (Smith et al. 1990, Settle and Drake 1993). For example,
Table 1. Mean and standard deviation of the brightness values for the seven vegetationcovers. The last column is the mean of normalized difference vegetation index.
TrainingSamples
Band1 Band2 Band3 Band4 Band5 Band7
NDVIMean Std Mean Std Mean Std Mean Std Mean Std Mean Std
Gr1, shrubs; Gr2, tall wild grass; Gr3, short wild grass; Agr, agriculture; Gol, man-madegrass; Rip, riparian vegetation; For, trees.
Figure 4. Mean brightness values of the seven vegetation types in a band 3 vs. band4 featurespace plot. Gr1, shrubs; Gr2, tall wild grass; Gr3, short wild grass; Agr, agriculture; Gol,man-made grass; Rip, riparian vegetation; For, trees.
Urban vegetation mapping using sub-pixel analysis and expert system rules2651
when using grass signature, a given pixel is classified as containing 60% of the MOI
(i.e. Gr1), as containing 70% using signature 2 (i.e. Gr1), and as containing 70%
using signature 3 (i.e. Gr1). The total of these three signatures is well over 100%, but
the processor identifies the average fraction (66.7%). The classification output for
each pixel will consist of four layers, one for each signature and the fourth layer
containing the average material fraction of all signatures. It may be acceptable to
take the average of two or more signatures of the same family, but it may not
provide useful information or it may lead to information loss if we take the average
of different family members. For example, an average fraction value of 45% is
assigned as vegetation for a candidate pixel with signature responses from 10%
grass, 80% trees, and 10% background. Apparently the multiple signature approach
(Huguenin et al. 1997) employed in this study falls short in handling multiple family
members (multiple MOIs). This does not necessarily mean that the processor is
incapable of accurately identifying the material of interest. To overcome thislimitation, a set of expert system rules based on the signatures of selected materials
and their greenness vegetation biomass in relation to the normalized difference
vegetation index (NDVI) were developed to identify vegetation distribution in
Norman. This is basically to prepare a single output map showing all vegetation
types over the study area. In the sub-pixel classification stage, each of the signatures
was used to produce each vegetation map of eight levels.
Table 1 lists the mean brightness and standard deviation value by band for the
seven signature samples. This shows the purity of the selected training signatures.
The expert system rules developed in this study are based on the assumption that
there is a dominant vegetation cover type related to its normalized difference
vegetation index within the candidate pixel. The other vegetation type that might
occur in that candidate pixel is negligible. Hence, the output map contains one
dominant vegetation cover type at a time with its fraction value for each pixel under
investigation. The rules basically determine the dominant vegetation component
within the candidate pixel. Hung and Ridd (2002) used threshold values derived
from the ratio of band 4 over band 3 to adjust pixels with invalid percentages
representing a situation in which the training sample statistics do not handle well. In
this study, the mid-value of two successive NDVI values was set to determine the
dominant vegetation type in this study. For example, 0.5765 was used as the first
threshold value in this study to give priority to Gr1 vegetation type, since this value
is the mid-value between two successive NDVI values (0.587 and 0.566). Hence, the
threshold values for Gr1, Gol, Rip, Tree, Gr2, Gr3, and Agr classes were 0.5765,
0.5435, 0.4915, 0.4555, 0.4255, and 0.2425 respectively. The index was used to
distinguish all selected vegetation cover types. Before applying the expert systemrule, we recoded eight levels of fraction classes for all seven class outputs to avoid
confusion (e.g. 1–8 for Gr1, 9–16 for Gol, 17–24 for Rip). The rules are explained by
the following procedure.
If NDVI.5threshold value then
Gr1 is dominant (take the fraction of Gr1)
Else if NDVI.5threshold value and not Gr1 then
Gol is dominant (take the fraction of Gol)
Else if NDVI.5threshold value and not Gr1 and not Gol then
Rip is dominant (take the fraction of Rip)
Else if NDVI.5threshold value and not Gr1 and not Gol and not Rip then
For is dominant (take the fraction of For)
2652 S. W. Myint
Else if NDVI.5threshold value and not Gr1 and not Gol and not Rip and not
For then
Gr2 is dominant (take the fraction of Gr2)
Else if NDVI.5threshold value and not Gr1 and not Gol and not Rip and not
For and not Gr2 then
Gr3 is dominant (take the fraction of Gr3)
Else if NDVI.5threshold value and not Gr1 and not Gol and not Rip and not
For and not Gr2 and not Gr3 then
Agr is dominant (take the fraction of Agr)
After completion of this process, those pixels with NDVI above the first threshold
value will be filled with fraction values of dominant class (Gr1) and all other
possible classes in the output map. Experiments from this study showed that there
were only a very few pixels left to be assigned to other classes after Gr1 was given
priority. It was also found that the leftover pixels were, in most cases, assigned to
one or two sample classes next to the dominant signature (e.g. Gr1). This implied
that the NDVI threshold procedure is reliable and the sub-pixel processor picked the
signatures effectively. It could also be inferred that the signatures selected in this
study are pure and appropriate. The next step is to develop another set of
procedures to give priority to Gol using NDVI between 0.5765 and 0.4535 (mid
value between 0.566 and 0.521). The second set of rules can be described as
If NDVI,first threshold value and.second threshold value then
Gol is dominant (take the fraction of Gol)
Else if NDVI,first threshold value and.second threshold value and not Gol
then
Gr1 is dominant (take the fraction of Gr1)
Else if NDVI,first threshold value and.second threshold value and not Gol and
not Gr1 then
Rip is dominant (take the fraction of Rip)
Then follow the same procedure until the last class is obtained.
After completion of the second process, those pixels with NDVI values between
the first and the second thresholds were filled with fraction values of the second
dominant class (Gol) and all other possible classes in the output map.
Five more expert system rules following the procedures described above were
developed for the rest of the training samples (i.e. For, Rip, Gr2, Gr3, Agr) using
their respective NDVI threshold values to complete the whole study area for all
dominant training samples.
7. Results and Accuracy Assessment
Figure 6 (a) to (e) show the single category percentage images with grey level display.
Brighter areas in output maps represent a higher percentage, and darker areas
indicate a lower percentage. For comparison purposes, the normalized difference
vegetation index (NDVI) of the study area is provided in figure 5. Figure 6 (a)
illustrates the grass coverage of Norman. We combined three different grass
signature classes to show total grass distribution in Norman. It should be noted that
the grass coverage is not the average fraction of all three grass layers generated by
the three grass signatures. It was formed by integrating the values of the outputs
produced by the expert system rules described above. The rest of the classes are
Urban vegetation mapping using sub-pixel analysis and expert system rules2653
shown separately. Figure 7 shows the final map with all five vegetation distributions
found in Norman.
Some interesting observations were found in the study when examining the
number of pixels quantified into each of the seven vegetation classes in the interim
maps, as well as the final map after expert rules. Apparently, there were some
overlapping categories found in the classification. It was anticipated that there were
some overlaps among the three different grass categories and man-made grass
vegetation because samples from the same family tend to possess similar signatures.
We believe that this situation is obvious and understandable. If there were some
overlaps between two different families, it would have been a limitation in the sub-
pixel processor and consequently would have led to certain errors. However, it was
possible that there were some situations where those two different classes co-existed
in some pixels. It was difficult to trace which overlaps were acceptable and which
were not. Some categories had a reduction in the number of pixels during the
integration. This may be due to the effect of overlap among signatures from the
same family or signatures from the different families. In general, without looking at
the classification accuracy or correlation of the referenced classes and the classes
generated by the sub-pixel processor, the percentage for each category looks
reasonable and seems to conform to the vegetation distribution in the city of
Norman.
By overlaying the final vegetation distribution map and the Landsat TM image, it
was found that there was some confusion among the three types of grassland and
man-made grass vegetation. We anticipated this situation since they all were
basically from the same family. We also found that there was a little confusion
between agriculture and grass since their signatures were also similar. However, our
main concern was the signature confusion between completely different families (e.g.
grass and tree or riparian vegetation). This is because they are not only different in
terms of the amount of vegetation biomass but also the level of importance with
regards to assessment and monitoring of the urban heat island, air pollution, and
environmental degradation. On the other hand, the percentage of trees (crown
closure) or percentage of riparian vegetation in urban land-use classes (e.g.
residential, commercial) may be crucial in the planning and management of the
urban environment (McPherson 1994, Lo et al. 1997). Hence, we observed the
Figure 5. NDVI image of the study area.
2654 S. W. Myint
overlap of riparian vegetation classes and three different grass categories (figure 8 (a)
to (c)) by using the Matrix function in the IMAGINE software. It was found that
the highest overlap between riparian vegetation and Gr1, Gr2, and Gr3 were
0.095%, 0.187%, and 0.125% respectively. We believe that the percent overlaps are
low and acceptable. This implies that the classifier identified the classes effectively
and the selected samples contained pure signatures. This also indicates that our
expert system rules developed for overlapping classes of different families as well as
Figure 6. Vegetation distribution maps derived from the IMAGINE sub-pixel analysis: (a)wild grass; (b) man-made grass; (c) riparian vegetation; (d) tree; (e) agriculture. Brighter areasrepresent higher-percent category of a certain class.
Urban vegetation mapping using sub-pixel analysis and expert system rules2655
same families were justifiable and reliable. It can be observed from Figure 8 (a) to (c)
that majority of the pixels for the overlap between Rip and Gr1 and Rip and Gr3
belonged to lower-level classes, whereas Rip and Gr2 belonged to higher-level
classes. It is understandable that the overlap of higher-level classes from different
families is more important to consider than lower-level classes. It was anticipated
that there could be some confusion between Rip or For and Gr2 (table 1) since their
signatures were somewhat similar. However, the percentage of total overlapping
pixels was less than 1.5% and it was not as high as we expected. This is one of the
reasons why we developed the expert system rules to identify classes more accurately
Figure 6. (Continued.)
2656 S. W. Myint
by giving priority to different classes according to their NDVI threshold values. We
believe that the above problem was effectively taken care by the system rules.
As mentioned earlier, IKONOS 4 metre resolution multispectral image of
Norman, Oklahoma, with the aid of 1 : 50,000 scale aerial photographs, was used for
assessing the accuracy of the levels of vegetation distribution from sub-pixel analysis
Figure 6. (Continued.)
Figure 7. Final map with all five vegetation distributions derived from the IMAGINE sub-pixel analysis and expert system rules. Note: The original map contains forty levels of classes(8 levels65 vegetation classes540) and only eight levels of distributions for all classes areshown for better visualization and interpretation.
Urban vegetation mapping using sub-pixel analysis and expert system rules2657
and the expert system rules employed in this approach. Field verification was also
carried out to identify the classes accurately. Remote sensing accuracy generally
refers to thematic accuracy according to the difference between referenced and
classified data. There is some uncertainty in accurately determining the percentage
of vegetation types in each pixel under study due to the rectification accuracy and
the date of acquisition of aerial photographs. It should be noted that even though
Figure 8. Percent overlap of riparian vegetation classes and three different grass categories:(a) Rp vs. Gr1; (b) Rp vs. Gr2; (c) Rp vs. Gr3.
2658 S. W. Myint
we kept the root mean square (rms) error for the rectification of Landast TM image
less than one, there would have been some significantly high locational errors in
image geometric accuracy. This is often referred to as spatial accuracy, due to the
position shift between coarse resolution and finer resolution data pointed out by
Singh (1989); for example, an rms error of one means that the referenced pixel is
30 m away from the transformed pixel. This could have contributed more than 49
pixels (.767 pixels) off for the IKONOS image in this study. In some cases, there is
also some uncertainty in identifying some features and objects in aerial photographs
as well as in IKONOS images. This limitation could be referred to as spectral
limitation. The differences in the dates of the acquisition of Landsat TM, IKONOS,
and aerial photographs could also be considered a limitation in effectively assessing
the accuracy. This could be referred to as temporal accuracy. We believed that it was
impossible to accurately determine the percentage of different vegetation types in
each pixel. On the other hand, unlike per-pixel classification approaches, the classes
we identified were not completely spectrally different classes, but rather the discrete
levels representing a percentage of material of interest in each pixel.
A regression analysis is generally performed and presents the correlation
coefficient between two sets of ground component percentages from sample points
for most linear spectral un-mixing approaches (Smith et al. 1990, Foody and Cox
1994, Bastin 1997, Small 2001, Hung and Ridd 2002). One set of data is generated
by the classifier, and the other set is obtained from the reference data (e.g. aerial
photograph). Unfortunately, the output of the IMAGINE sub-pixel classifier in its
current form does not support linear regression analysis (Ji and Jensen 1999) since it
requires data to be in interval ratio scale. Hence, we decided to assess the accuracy
by showing the Spearman’s correlation between referenced classes (actual) and