Environmental Engineering Reference
In-Depth Information
After classification, the two red surface classes (light, dark)
and the three gray surface classes (light, medium, dark) were
aggregated into one red surface class and one gray surface class
respectively, reducing the total number of classes to seven plus
shadow. The overall performance of the classifier on an inde-
pendent stratified random sample of validation pixels for these
classes was characterized by a kappa index-of-agreement of 0.91.
The accuracy and the spatial coherence of the classification was
further improved by applying dedicated post-classification tech-
niques to remove shadow, correct remaining classification errors
and reduce structural noise typical for pixel-based classifications
(Van de Voorde, De Genst and Canters, 2007). Post-classification
improvement of the land-cover map resulted in an overall accu-
racy of 95.7% and increased the kappa index from 0.91 to
0.95 (Table 18.3). The seven remaining land-cover classes were
grouped into a single vegetation class (including shrub and trees,
grass, and crops), a single impervious class (including red, and
gray surfaces), water and bare soil.
To characterize land-cover at the medium resolution, an
MLP sub-pixel classification model was developed, estimating
the proportion of the four major classes (impervious surfaces,
vegetation, bare soil and water/shade) in each pixel of the Landsat
ETM
significantly from the trend were considered as indicative of
major changes in land-cover and were not used for training and
validation (Van de Voorde, De Roeck and Canters, 2009).
Feature selection and neural network building were accom-
plished with NeuralWorks Predict® software, starting from a
set of transformations of all multispectral Landsat ETM
+
bands
(1 - 5, 7) and all possible ratios between these bands. Application
of the model to the part of the Landsat ETM
+
image that overlaps
the IKONOS image resulted in four proportion maps (vegeta-
tion, impervious surfaces, water and bare soil). Figure 18.5 (left)
shows the proportion map obtained for impervious surfaces. A
visual comparison with the proportion map of impervious sur-
faces, derived from the IKONOS image (Fig. 18.5, right), shows
a good correspondence in the overall pattern of imperviousness,
although the sub-pixel classification result is more generalized
and includes less structural detail.
The performance of the sub-pixel classifier was assessed on
an independent validation set by calculating per-class mean error
(
ME
C
), as well as per-class mean absolute error (
MAE
C
):
i
=
1
(
P
ij
−
P
ij
)
N
N
ME
Cj
=
(18.9)
image. To train and validate the sub-pixel classifier, the
land-cover classification derived from the IKONOS image was
spatially aggregated to produce reference proportions for the four
majorland-coverclassesforeachETM
+
pixel in the overlapping
zone between the two images. Because the images are not of the
same date, precautions had to be taken not to include pixels in the
training and validation phase with different land-cover in both
images due to seasonal shifts (leaf condition, crop cycles) or due
to a change in land-use (e.g., transition from non-built to built
area). Since urban areas are dominated by impervious surfaces
and vegetation, most changes in the 8-month period between
the acquisition dates of both images are related to changes in
the vegetation component of the pixels. Therefore, in order to
detect anomalous pixels the ETM
+
i
=
1
(
P
ij
−
N
P
ij
)
MAE
Cj
=
(18.10)
N
where
Nis
the total number of pixels in the validation sample,
C
is the total number of target classes (4: vegetation, impervious
surfaces, water and bare soil), C
j
is target class
j
,P
ij
is the
proportion of class
j
inside validation pixel
i
, derived from
the high-resolution land-cover map (ground truth),
P
ij
is the
proportion of class
j
inside validation pixel
i
,estimatedbythe
sub-pixel classifier.
To assess the impact of cell aggregation on proportional
accuracy, all error measures were calculated at the original 30 m
resolution, as well as after aggregation of proportions to 60 m
and 90 m. As shown in Table 18.4, impervious surfaces and bare
soil are slightly underestimated, while vegetation and water are
slightly overestimated. The slight underestimation of impervious
surfaces ( - 1.8%) may be explained by the presence of shadow
in urbanized areas. When examining the estimated proportion
of water within the area, it turns out that small portions of
NDVI values of all pixels
in the overlap with the IKONOS image were plotted against
the mean NDVI value of the constituent IKONOS pixels, after
converting the raw DNs for both images to at-satellite reflectance
values. A strong linear relationship was observed between ETM
+
NDVI values and IKONOS mean NDVI values. Pixels deviating
+
TABLE 18.3
Confusion matrix for the high-resolution land-cover map, obtained by MLP-classification and improved with post-classification
techniques.
Red surfaces
87
87
1.00
Bare soil
1
146
1
1
149
0.98
Water
1
59
5
65
0.91
Grass
1
278
51
7
337
0.82
Crops
172
172
1.00
Shrub and trees
6
2
346
354
0.98
Gray surfaces
2
1
797
800
1.00
Sum
88
149
60
285
225
354
803
1964
Producer's accuracy
0.99
0.98
0.98
0.98
0.76
0.98
0.99
PCC = 95.7%
Search WWH ::
Custom Search