Geography and
Geology/5. Cartography and geoinformatics
Student
Nebosenko T.V.; Candidate
of Technical Sciences, Associate Professor Garkusha I.N.
National Mining University
Research methods of segmentation in the process
of classification types of the terrestrial covering
While
thematic processing of space pictures, information classes are defined as a set of objects, which it’s
necessary to allocate
according to the requirements of a solved thematic problem. Application of segmentation algorithms that use certain conditions of
uniformity of object classes, frequently leads to allocation segments on
the image that are not corresponding to information classes of the solved
thematic problem. To overcome the arising effects the preliminary processing of
the image by classification algorithms has been included in the segmentation
process.
Research
objective - improvement of quality and efficiency of the image segmentation.
The
fragment of space picture that has been received by TM scanner (satellite Landsat - 5), displaying fields with crops
of grain, was used in the research. Image classification that has used
MultiSpecs’ applied file with reference sites has been performed by the
following algorithms: minimum distance to means, correlation (SAM), matched filter
(CEM), Fisher linear discriminant, the Gaussian maximum likelihood pixel
scheme, the ECHO spectral/spatial classifier.
The
reference sites, belonging to five classes of a terrestrial covering have been
selected in ENVI program for comparison of the classification
results. Since the data from picture has distribution close to normal, the following classification algorithms were used:
Mahalanobis Distance and Maximum Likelihood (the Bayesian classifier).
The
best result has been
received in ENVI with application of classification algorithm Mahalanobis
Distance to the image.
The
result of classification has been subjected to processing in MATLAB and
Definiens Developer. Three segmentation methods were used in MATLAB: threshold,
a morphological watershed, planimetric. Algorithms of planimetric and threshold
segmentation have yielded unsatisfactory results.
To
eliminate the superfluous segmentation in case of application morphological
watershed algorithm the following approach is offered. Before watershed
segmentation, gradient modules of the image which use linear filtration methods
(Sobel and Laplasian filters) have been calculated. The morphological watershed
result has been defined by result of gradients calculation.
Multiresolution
segmentation of classification result has been performed in Definiens Developer
program. The best result of segmentation of the investigated image has been
received by setting the Scale parameter equal to
35. Thus, integral objects with the big area were shared to the parts in the
image. Such phenomenon was not observed while performing the segmentation in
MATLAB.
As enough general quality criterion of segmentation results Hausdorff distance has been used. This criterion measures the distance between two pixel sets:
and 
.
where
If 
, this means that all the pixels belonging to 
are not father than d
from some pixels of 
. This measure indeed gives a good similarity measure between
the two images.
Using
the given indicator, it has been established that any of the existing methods
does not give exact allocation of the object borders. Therefore, segmentation method based on the object borders
allocation by brightness levels has been offered. The segmentation results that
were received by this method can be used in the following areas: in land
management for calculation of the various function areas; in agriculture to
control the crops, both all types, and separately set by the researcher.
The
offered method allows precisely allocate the object borders on the image and
raises efficiency of the segmentation.
The literature
1. R.Gonzalez,
R.Woods «Digital image processing » Moscow: The Technosphere, 2005. – 1072 p.
2. S. Chabrier, H.
Laurent, B. Emile, C. Rosenberger and P. Marché «A comparative study of
supervised evaluation criteria for image segmentation»