Oliveira, Hermeson
This method differs from the conventional classification methods in a way that a clustering algorithm is
applied over a segmented image that was obtained by using region-growing technique. Then a selection of
training areas is made and finally a supervised algorithm is applied over the image to obtain a classification.
2 METHOD
The method involved 5(five) different steps: pre-processing (register and contrast), processing (segmentation
and classification), verification of the accuracy of the mapping, analysis of the results and elaboration of the
final document.
The segmentation technique proposed is object delimitation using region-growing method. Initially, the
algorithm labels each pixel as a distinct region. Pixels with similarity values lower than a threshold are united.
Later the image is fragmented into sub-regions, they will later be united using another threshold given by the
user ( INPE, 1996; Khodja et al, 1995; Khodja & Mengue, 1996).
The methodological proceedings of classification were based on two distinct processes that complement each
other. The first process used, non-supervised classification, tries to cluster pixels within the same spectral
variance (Bognola,1997). The second process, supervised classification, was controlled by the selection of
regions that represent each class. It was a sample selection process (Khodja & Mengue, 1996).
2.1 Image Segmentation
During this step the image is divided into spectral homogenous regions. It considers intrinsic characteristics
of the image like pixels gray level, texture and contrast (Woodcok et al, 1994).
Segmentation is a process that has the aim of unite regions that have the same properties (Khodja et al, 1995).
The algorithm used needs the definition of two variables, a similarity threshold and a minimum size to each
region. Imposition of lower similarity will create a great number of fragments in the original image. If the
other variable, minimum area is defined bigger than necessary, lots of heterogeneous polygons will be
created. In the other way around, high similarity threshold and small minimum areas, will also produce
heterogeneous fragments.
Optimum values, to be determined for image segmentation, will depend on the spatial division patterns of the
objects in each area. There aren't standard values for these variables. The user has to do lots of approximation
to reach an adequate fragmentation level.
2.2 Image Classification
The methodological proceedings of classification are based on two distinct processes that complement each
other. First, a non-supervised classification is used, it tries to cluster regions using as criteria the spectral
variance. The second process, supervised classification, is controlled by selection of training fields that
represent each class.
The non-supervised process is used to select spectral homogeneous areas that will be used, on a second step,
as training fields during the supervised classification.
2.2.1 Non-supervised classification. This process will help the selection of training fields in the image. Each
pixel will be an homogeneous spectral region. The region-growing algorithm will cluster pixels that have
same characteristics.
Only one supposition is associated with the data, they have to obey normal distribution. The equation that
represents the observation distribution of each pixel is equation 1:
pu se a a
1066 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.