ul 2004
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Two scenes HRV SPOT from May 27, 1986 and respectively
April 27 ,1993 used has been. Their characteristics in table 1.
are presented.
Acquired date May 27, 1986 | April 27, 1993
Bands 1] 2 3 ] 23.3
K /J 98 / 259 98 / 259
Resolution (meters) 20 x 20 20 x 20
Rectification (pixels) i 0.391 + 0.275
Table 3. The satellites data characteristics
3. METHOD
The change detection realised with aids of multispectral data
need: the preliminary processing, classification and specifically
proceedings post-classification (Jensen et al, 1995) In this
study was applied two methods: normalized difference
vegetation index (NDVI) image differencing and a princial
component analysis.
3.1 Preliminary processing
One from important temporal reason for change-detection is the
acquiring month of the imagery. Images acquired during the
period with power sun light, present a very good contrast
between various details (Jansen, 1993). For example, in this
case, the contrast from covered with vegetation soil and
uncovered is very strongly. Using the scene acquired in the
same years period are suggested to change-detection with
purpose to reduce the problems which appear because sun-angle
difference, vegetation-phenology changes and differences in
soil-moisture.
In this study for radiometric processing the technique based on
reference image are used (Hall et al., 1991). Image of 1986 has
been thought, because its minimum values are weakly. These
techniques apply the standardisation in respect with reference
image and adjust the calibration sensor errors. Also she
adjusting the effects due to atmospheric differences and lighting
between images. For this the fifteen pixels in each zone more
darkness (near to level 0), respectively more brightness (near to
255) has been choose. Radiometric transformation which tie the
two values of two images have the form y(i) = a(i) + b(i) * x (i),
where y (i) is pixels radiometric value from reference image
and x(i) is the correspondent radiometric value pixels of
corrected image.
After the radiometric correcting, the geometrically rectified
images have been, so that the same pixel at one date overlaps
the same pixel for the other date. The accuracy of change-
detection is directly conditioned of geometric rectification. For
this operation twenty ground control points obtained by
aerotriangulation was used. The aerotriangulation is performed
on aerial photos at 1:20,000 scale acquired in 1990. The root-
mean-square (RMS) error at rectification of the two images was
under 0.20 pixels.
For resampling of images was applied the method of the nearest
neighbour.
3.2 Composite images generation
In principal component analysis of images stand out restrain of
first principal component of each image. In means this is 7696
of input data variance. Also, and the second principal
765
component was retained, which the value represent 2196 of
variance. For each image at this two principal component was
added the Normalized Difference Vegetation Index (NDVI),
calculated with ecuation:
NDVI « ( XS3— XS2)/( XS34 XS241)* 128 (1)
Using of this allowed clear to distinguish of vegetation covered
zones of the uncovered terrain zones. This index are
advantageously for separation between the areas with very rich
vegetation and the areas with the mobile and quasi-mobile sand
correspondingly to river and sea banks (Letea, C.A.Rosetti,
Schiopu, Räducului, Caraorman, Lumina, Lat, Rosu, Puiu,
[vancea, Säräturile, Cerbului and Câsla Vädanei) or sea
beaches. Also, facilitated the separation of very little and fine
town structures from locality zones of vegetation (Sulina, Letea,
C.A. Rosetti, Sfistofca, Crisan, Mila 23, Sf. Gheorghe,
Murighiol, Dunavätu de sus and Dunavätu de Jos).
3.3 The post-clasification comparison and change-detection
Using of the post-classification comparison methods is
advantageously, because the image acquired at different times
are separately classified. This method allow minimising the
effects dues to different atmospherically conditions and using
different sensors for multispectral images acquiring (Singh,
1989). Thus different studies showed that by association of
mode filtering with classification procedures are possible the
improve accuracy of change-detection (Jansen et al., 1993).
Mode filter applied in 3 by 3 neighbouring with a threshold
value in generally three, allows to suppress the isolated pixels
or poor classified or the pixels dues to noise (Jansen et al.,
1993). It replaces central value pixels by a majority value.
Majority threshold corresponds to threshold of which going,
majority value replace central value pixels.
For classification was applied a hibryd classification process. In
this technique by clustering are optimised defining the sample
classes, which will be used in the supervised classification
process.
For spectral signature file establishing first was appllied a
clustering method (Isodata from Erdas). Afterwards, the file to
supervised classification realised by maximum likelihood has
been.
The classification of composite images from 1986, respectively
1993 was performed in accordance with seven classes:
uncovered soil, stagnant water in lakes, running water, sea
water, compactly common reed, mixture common reed-mace
reed, town structures. The mode filtering technique was applied
to the two classified images in a 3 by 3 neighbouring.
The accuracy classification from two date was estimated by
using for each the standard, single-date, qualitative accuracy
assessment procedures (i. e., an error matrix and Kappa
analysis) (Congalton, 1991).
The reference data was extracted of pedologychal map,
vegetation's map and topographical map of the Danube Deltas.
The images classification from 1986 have been one generally
accuracy of 79% and 0.46 Khat value, while for 1993 images
was optained 76% and respectively 0.41.
For change-detection the classified images was compare by
differencing.