2004
—
r the
civil
anch,
AX
| (to
the
topo-
Delta
ages
ids at
ange-
gery.
esent
). For
ation
'ed in
with
angle
es in
POT
nsure
istics
ed on
989):
urbid
and).
e, for
(1)
nage;
'ected
rence
nting
et al.
tellite
s for
ts for
tware
[s, at
g and
enter,
,5 for
im for
pling
sthod,
image
| used
à. The
:essed
omial
ample
el for
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
MSS image, £0,813 pixel for TM image and +0,275 pixel for
HRV image.
3.2 Composite images generation
Theoretically, vegetation, water and uncovered soil have very
different behaviors in visible and especially in IR range. In this
last range the separation of the three classes is easily done. A
realistic interpretation proves to be very complex. For example,
in range of 550 nm wavelength of visible spectrum, turbid water
and uncovered soil can be mistaken.
Multispectral images allow different color combination, through
which are selectively pointed out study objects. In most cases a
very expressive standard false color composite is realized if a
cambination between channels near IR, red, green - MSS 7, 5,
4: TM 4,3, 2; XS 3, 2, I is chosen (Thomson, 1990). It can also
be realized a principal standardized components analysis, to
confirm the three channels that will be used. This method brings
a significant improvement of the SNR (Signal to Noise Ratio).
In this case for color composite image generation it has been
used two principal component of PCA with Normalized
Difference Vegetation Index (NDVI). Adding this index allows
a preferential enhancement of the areas covered with vegetation
versus the built urban areas or uncovered soils.
Examining channels histograms for satellite images can be seen
that only a small part of sensor range is covered. As a result a
contrast enhancement is needed. For linear contrast
enhancement with point's saturation, are used low values of
enhancement (2.5-3.5%) with purpose of optimal use of
unsupervised classification. The nonlinear contrast
enhancement offered better results (histogram equalization) by
enhancing the contrast for the densest domain of reflectance
values from the original image. (Deutsch, Heydt, 1987; Singh,
1989). There also been used Laplace filters techniques for
directional and nondirectional edge enhancement (used for
linear details - channels, shore line).
The wider colors range on realized images allows
distinguishing waters, vegetation and soils classes. Main
morpho-hydrographic features as shore line, fluvial and
continental plains, running, stagnant and marine ecosystems,
past meanders remains, agricultural lands and villages parceled
structure, uncovered soils and different vegetation types
(compactly common reed, mixture common-mace reed, forest
vegetation, grass vegetation) are revealed. Digital processing
has been done with IDRISI 32 Release 2 software.
3.3 Multispectral classifications
For multispectral classification it has been used a hybrid
method. First, unsupervised classification of the color
composite image SPOT HRV has been realized (ISODATA
algorithm), where the selected 16 classes are defined (Lee and
Marsh, 1995). Several confusions occurred: for example in
some areas of image forest vegetation class and compactly
common reed class overlapped, or stagnant water class and
remind water class. In order to improve classes definition have
been used radiometric measurements, taken in 1984 and 1993
and a vegetation map at 1:100.000 scale, realized in 1993. After
classes regrouping followed a hybrid supervised classification
for 8 classes: running water, stagnant water, compactly
common reed, mixture common reed-mace reed, forest
vegetation, grass vegetation, uncovered soil and agricultural
soil for Caraorman-Dunavit site and sea water, stagnant water,
running water, compactly common reed, mixture common reed-
mace reed, forest vegetation, grass vegetation, uncovered soil+
urban area. Classification merges the parallelepipeds and
331
maximum likelihood methods to remove problems caused by
nonnormalised distribution of used image (Jansenand et al.
1993). This last classification served for MSS and TM
classifications checking (Kartikeyan and Gopalakrishna, 1994).
The accuracy classification for the three dates was estimated
using the standard, single-data, qualitative accuracy assessment
procedures for each image. Producer and user accuracy were
calculated for each change class, along with the overall
accuracy (error matrix and Kappa Index of Agreement analysis)
(Congalton and Green, 1998). Global accuracy for obtained
classifications varied between 89.56% and 95.73%. Kappa
coefficients varied between 0.82 and 0.91.
Figures 1, 2 for Caraorman-Dunavát site and figures 4, 5 for
Sulina site present classification results for TM and HRV
sensors.
3.4 Postclassification processing
The classified images were compared using a statistical analysis
to reveal changes. Ecological changes are focused on
emphasizing the habitat changes of Delta’s ecosystems,
depending on antropic interventions in these areas. Figures 3
and 6 present changes diagram in habitats of two sites.
It has also been calculated many vegetation indices types, in
order to realize qualitative map of vegetation classes. Among
these, Soil Adjusted Vegetation Index (SAVI) and Normalized
Difference Vegetation Index (NDVI) proved to be the best.
These have been calculated using the following equations:
_ channellR—channelR
NDVI = channelIR+channelR @)
(1+L)*(channelIR-channelR )
channelR+channelR+L
SAVI =
where coefficient L=0.5; Channel R is channel 5 for MSS
sensor, channel 3 for TM sensor and channel XS2 for HRV
sensor; Channel IR is channel 7 for MSS sensor, channel 4 for
TM sensor and channel XS3 for HRV sensor. Figure 7 presents
soil adjusted vegetation indices (SAVI) for the two sites
Analyzing these two images can be observed a significantly
separation between running water (Danube's branches,
channels) and stagnant water (lakes and past meanders remains)
ecosystems and the rest of ecosystems, because of high
absorption of red and infrared radiation by the liquid medium
and its low chlorophyll plants constituent. Separation of
different vegetation types is facile, both in ecosystem of marshy
and flooding surfaces, and in ecosystem of river banks and
marine levees or in ecosystem of polders.
For example, forest vegetation class from Caraorman bank,
northern Cretan village and Isadora Lake are very well depicted
apart the other vegetation types (common reed and mixture
common reed-mace reed), because of its dominant spectral
reflectance in near infrared range.
By comparing the results from multiple aerospace images
analysis, the ecological trends in the changing areas of most
Danube Delta ecosystems were seen to be nonlinear. During the
study period (1975-1993) nonlinearity of the ecological trend is
obvious (Figure 8). The nonlinear trend is a curve of the form:
YcanhtX, — X, ANA (4)
