4.3 Polygonizing and Rasterizing
4.3.1 Vector to polygon conversion. All the digitized
data were in segment mode. The polygonization process
was done to make closed polygons. In ILWIS system
we used the program called "polygonizing" to perform
this task. For this process, a unique number was
assigned to each polygon.
SEGMENTS
FINAL.DAT
FINAL.SEG polvgonizing FINAL.POL
FINAL .SLG ciao, FINAL.TOP
FINAL.CRD FINAL.PLG
4.3.2 Polygon to raster conversion. These polygonized
parcel boundaries are still in vector mode. In order to
process these data along with the satellite data, they
were converted to a raster form. In ILWIS system we
used the program called "rasterization" to perform this
task. The program assigned an arbitrary color to each
polygon as a label.
FINAL.DAT
FINAL.POL rasterization FINAL.MPD
FINAL.TOP . 9"seseconcon FINAL.MPI
FINAL.PLG FINAL.INF
By selection of the raster map in the "read pixel from
screen" menu information of all polygons can be
retrieved simultaneouslv. Moreover we should change
the name of the attribute information stored in ILWIS
table file (e.g., FINAL.INF change to FINALSAV INF).
4.4 The Final step
4.4.1 Crossing rasterized polygons and result of
classification. These rasterized polygons (FINAL.*)
were crossed with the geometrically corrected
conventional classified satellite image (e.g., ML3-237).
Crossing can be done with the program "crossing" in the
ILWIS system.
4.4.2 Aggregation. The program performs different
types of aggregation functions on the pixel values listed
in the cross table. Aggregation functions were applied to
the pixel values of the second map (e.g., ML3-23Z) per
pixel value of the first map (e.g., FINAL). At this level,
the most occurnng (predominant) pixel value within
each polygon is assigned to that polygon. This process
was carried out by using of the program "mapcalc" as
follow:
MCALC IMP := CROSS.PREDCOL([FINAL];
4.4.3 Crossing and error analysis. This IMP is
generated under MCALC and crossed by the test
sample classes ,e.g., TEST3. Confusion inatrix is shown
in figure 6.
540
1 z 3 4 5 6 2 ü uncl {| ACC
(Tw 4 9 0 2 5 1 16 9 TE
210 3 0 RO 0 0 0D 0 008
3 12 2° na 0 0 à à 0 08
410 12 n 0 0 0 0 0 0 0m
540 0 0 0 31 0 0 0 0 1 m
610 0 0 08 Ug 0 ö 40 0 [om
24d 2 0 qU B m4 0 0.30 0 00
815 0 gn 0 0 D aea 074
REL | 088 0.69 ? 0 0 0.5520 0.08
average accuracy = 40.57 %
average reliability = 34.41 %
overall accuracy - 54.67 %
figure 6 -Confusion matrix of Maximum Likelihood
classifier affer aggregation
As it can be seen, by developing the improved
classification method we could reach to overall accuracy
of 5467% in compare with 29% in conventional
method.
5. CONCLUSION
Scanned aerial photos have the advantage of showing
edges to separate different earth cover of variable
landuse. For that matter, the classification is improved
because one can very easily recognize all features
including point and line, where the 100 dpi was used as
the resolution of photo images.
For non-flat terrain, it is essential to consider the DEM
because the question of relief displacement should not
be overlooked since it adversely affects the geometry of
satellite images. If both factors are put into consideration
in the correct way, the classification is trivially improved
to a very good extent. The scanned aerial photos and
DEM have facilitated the classification of satellite
imageries though it has been difficult to be depicted in
terms of technical documents.
The problem in matching of vector data and raster data
has led us to realising that DEM had propagated errors
because topo data almost matched with SPOT raster
data on flat terrain but the some discrepancies were
visually noticed on mountainous terrain to disclose relief
displacement that increases with height.
In aggregation process in each parcel, we assigned the
label of dominant pixels to all pixels within that parcel
If we consider everything perfect, then we can expect
the classification being improved.
6. RECOMMENDATION
Hereby, we consider the factors that increase the
accuracy of land-cover type elevation. Accuracy and
attention in collecting up-to-date ground truth data have
the great importance in classification. Out-dated of
inaccurate ground truth data, besides the changes of
land-cover in different seasons, are the main problems.
The best thing is to use ground truth data, aerial
photographs, and satellite images of the same time (or at
least at the same period of successive years). Of course,
there are some serious problem to obtain this set of data
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
es Ca 7