150
B.G.H. Gorte
- Use the iterative procedure of Sect. 7.4.4 to combine pixel probability densities
with segment prior probabilities. This approach will be applied in the case study
of Sect. 7.4.6.
In Gorte (1998) an integrated segmentation and classification procedure is developed, which first gathers evidence concerning spatial and class-membership characteristics, and then combines these to delineate and identify relevant terrain objects.
The procedure uses a pyramid of segmentations with different coarsenesses.
7.4.6 Case study in the Pantanal Area, Brazil
The purpose of the case study is to assess the rather drastic land use changes in the
Pantanal region since 1985 (see Sect. 7.2).
Available were three Landat TM images of the Pantanal study area, from 1985,
1990 and 1996 (Colour Plate 7.A). A land-use survey from 1997 was available (Disperati et aI., 1998) with classes according to level 3 of the CORINE legend (see also
Chap. 19). From the survey map, training samples were selected in areas that were
assumed to be the same in 1996 and 1997, allowing to establish spectral signatures
for those classes using the 1996 image. Seven classes were selected, according to
level 2 or the CORINE legend - in preliminary experiments, level 3 appeared too
detailed for automatic classification of Thematic Mapper imagery. Two sets of samples were chosen, a training set and a test set.
Image segmentation (Gorte, 1996) (Sect. 7.4.5) was applied to the multi-temporal
NDVI composite from Sect. 7.3.4 (Colour Plate 7.B). The result is a set of multitemporal objects. These are regions of adjacent pixels showing the similar vegetation development (Colour Plate 7.C).
Iterative local prior probability estimation (Sect. 7.4.4) was applied to the 1996
image, using one set of priors in each segment of the multi-temporal segmentation.
This gives the 1996 land-use map. Comparison with the test set shows that an overall
accuracy of 71 % was reached using local prior probabilities, as compared to 61 %
with maximum likelihood classification using the same band combination. Maximum likelihood classification with six bands gives 67% accuracy I. Regarding this
rather low accuracy, it should be noted that test set pixels were chosen at random
from the survey data, without considering their spectral values.
Since the 1996 and 1985 images are of the same season, it was assumed that the
spectral signatures derived for the 1996 image are also valid for the 1985 image (Fig.
7.5). Under this assumption, also the 1985 land-use map could be made, despite
the absence of ground truth for that year. Therefore, also a 1985 test set was not
available.
To detect changes, a straightforward procedure is post-classification comparison,
which involves overlaying the 1985 and 1996 land-use maps. The difficulty is that
both classifications contain quite a large percentage of errors (which for 1985 is
even unknown). Therefore, it is difficult to distinguish between real changes and
those that are observed as a result of misclassification.
1 Iterative local prior estimation is developed for SPOT-XS with 3 bands.
B.G.H. Gorte
- Use the iterative procedure of Sect. 7.4.4 to combine pixel probability densities
with segment prior probabilities. This approach will be applied in the case study
of Sect. 7.4.6.
In Gorte (1998) an integrated segmentation and classification procedure is developed, which first gathers evidence concerning spatial and class-membership characteristics, and then combines these to delineate and identify relevant terrain objects.
The procedure uses a pyramid of segmentations with different coarsenesses.
7.4.6 Case study in the Pantanal Area, Brazil
The purpose of the case study is to assess the rather drastic land use changes in the
Pantanal region since 1985 (see Sect. 7.2).
Available were three Landat TM images of the Pantanal study area, from 1985,
1990 and 1996 (Colour Plate 7.A). A land-use survey from 1997 was available (Disperati et aI., 1998) with classes according to level 3 of the CORINE legend (see also
Chap. 19). From the survey map, training samples were selected in areas that were
assumed to be the same in 1996 and 1997, allowing to establish spectral signatures
for those classes using the 1996 image. Seven classes were selected, according to
level 2 or the CORINE legend - in preliminary experiments, level 3 appeared too
detailed for automatic classification of Thematic Mapper imagery. Two sets of samples were chosen, a training set and a test set.
Image segmentation (Gorte, 1996) (Sect. 7.4.5) was applied to the multi-temporal
NDVI composite from Sect. 7.3.4 (Colour Plate 7.B). The result is a set of multitemporal objects. These are regions of adjacent pixels showing the similar vegetation development (Colour Plate 7.C).
Iterative local prior probability estimation (Sect. 7.4.4) was applied to the 1996
image, using one set of priors in each segment of the multi-temporal segmentation.
This gives the 1996 land-use map. Comparison with the test set shows that an overall
accuracy of 71 % was reached using local prior probabilities, as compared to 61 %
with maximum likelihood classification using the same band combination. Maximum likelihood classification with six bands gives 67% accuracy I. Regarding this
rather low accuracy, it should be noted that test set pixels were chosen at random
from the survey data, without considering their spectral values.
Since the 1996 and 1985 images are of the same season, it was assumed that the
spectral signatures derived for the 1996 image are also valid for the 1985 image (Fig.
7.5). Under this assumption, also the 1985 land-use map could be made, despite
the absence of ground truth for that year. Therefore, also a 1985 test set was not
available.
To detect changes, a straightforward procedure is post-classification comparison,
which involves overlaying the 1985 and 1996 land-use maps. The difficulty is that
both classifications contain quite a large percentage of errors (which for 1985 is
even unknown). Therefore, it is difficult to distinguish between real changes and
those that are observed as a result of misclassification.
1 Iterative local prior estimation is developed for SPOT-XS with 3 bands.