7 Land-use and Catchment Characteristics
151
Survey
data
Selection,
Training
1997
~
Samples
Reclassi,------+I Class
fication
Spectral
Signatures
Colour r-----+i
Plate 7A
LandImage
use map
1996
1996
LocalImage
prior
1990
classification
Land·
Image
use map
1985
1985
NOVI
1996
NOVI
1990
fNOVI
Combination
1985
Colour
Colour
Plate 78 Segmentation
Plete 7C
Mutl~emp.
Segment
NDVI
map
Fig. 7.5. Classification of 1985 and 1996 images with local prior probabilities according to
segmentation of multi-temporal NDVI
Alternatively, classification of a multi-temporal data set was applied (Fig. 7.6).
The 1985 and 1996 images were combined in a single 12 channel data set From
the two land-use classifications pixels were selected having high a posteriori probabilities. Moreover, only those pixels were chosen that were surrounded by pixels
of the same class, to reduce mixed-pixel effects at boundaries. The selected pixels
are the most reliable ones in the two land-use maps. The resulting maps have eight
different pixel values: 1 - 7 for the seven classes, plus 0 for not selected. Overlaying
these two maps gives 64 combinations, 15 of which contain a 0 (zero) and are no
longer considered. From the remaining 49 combinations 31 were selected: the seven
no change classes, having the same class in both years, plus 24 change classes, belonging to one class in 1985 and to another in 1996. The remaining 18 combinations
did not occur in significant amounts of pixels.
The resulting change map has very many pixels with value O. The remaining pixels (with a change-class number between 1 and 31) were used as training set for a
maximum likelihood classification of the combined - multi-temporal- image. The
assumption is that the 31 classes are distinguishable in the 12-dimensional feature
space. The result is shown in Colour Plate 7.D, where the hatched areas indicate
change: the narrow lines refer to 1985 and the wider ones to 1996. As far as this
can be judged at this stage of the Pantanal project, the results are promising. Further
evaluation will take place in the near future.
The problem of detection of landuse changes is also discussed in detail in Chap.
19 of this book.
151
Survey
data
Selection,
Training
1997
~
Samples
Reclassi,------+I Class
fication
Spectral
Signatures
Colour r-----+i
Plate 7A
LandImage
use map
1996
1996
LocalImage
prior
1990
classification
Land·
Image
use map
1985
1985
NOVI
1996
NOVI
1990
fNOVI
Combination
1985
Colour
Colour
Plate 78 Segmentation
Plete 7C
Mutl~emp.
Segment
NDVI
map
Fig. 7.5. Classification of 1985 and 1996 images with local prior probabilities according to
segmentation of multi-temporal NDVI
Alternatively, classification of a multi-temporal data set was applied (Fig. 7.6).
The 1985 and 1996 images were combined in a single 12 channel data set From
the two land-use classifications pixels were selected having high a posteriori probabilities. Moreover, only those pixels were chosen that were surrounded by pixels
of the same class, to reduce mixed-pixel effects at boundaries. The selected pixels
are the most reliable ones in the two land-use maps. The resulting maps have eight
different pixel values: 1 - 7 for the seven classes, plus 0 for not selected. Overlaying
these two maps gives 64 combinations, 15 of which contain a 0 (zero) and are no
longer considered. From the remaining 49 combinations 31 were selected: the seven
no change classes, having the same class in both years, plus 24 change classes, belonging to one class in 1985 and to another in 1996. The remaining 18 combinations
did not occur in significant amounts of pixels.
The resulting change map has very many pixels with value O. The remaining pixels (with a change-class number between 1 and 31) were used as training set for a
maximum likelihood classification of the combined - multi-temporal- image. The
assumption is that the 31 classes are distinguishable in the 12-dimensional feature
space. The result is shown in Colour Plate 7.D, where the hatched areas indicate
change: the narrow lines refer to 1985 and the wider ones to 1996. As far as this
can be judged at this stage of the Pantanal project, the results are promising. Further
evaluation will take place in the near future.
The problem of detection of landuse changes is also discussed in detail in Chap.
19 of this book.