As
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
paper. It checks contrast between lines and background. Instead
of the approach, “Brightness consistency test” is adopted.
Brightness of end points and brightness statistic of both
segments are used. The idea behind the examination is that
brightness of both segments must be similar if segments are
parts of the same line. Suppose that brightness of the end point
on a segment is x, mean and standard deviation of another
segment is m2 and sd2 respectively. If the difference of x and
m2 is greater than 1.96 times of sd2, this line pair is rejected.
The test corresponds to 10% two-tailed significance test.
In the probabilistic screening section, the measure of junction’s
deviation (d) is calculated for each line pairs that pass rule
based screening. The pairs are sorted in ascending order of the
measure and the pairs d of which is less than threshold are
po a rte CIA. us, ELS
Figure 2. A classification result. All bands are used to
classify image. Number of class is 15.
Resulting classes are manually integrated and
categorised. Red, green, deep blue, light blue
and grey (dark & light) represent buildings,
vegetation, road, road & building and shadows
respectively.
Case 1 Case 2 Case 3 Case 4
Min. length S 10 5 10
Radiometric No use No use use use
Information
Iteration 4 4 = 3
Num. of Lines 826 187 885 180
Correctness 13.8 % 37.5% 13.7% 30.5%
Completeness | 55.4 % 36.2 % 49.7% 32.5%
Table 4. Result of grouping. Simply minified image was used
and same set of parameter was applied for all cases
in the edge extracting stage. Limitation of
minimum segments length in grouping stage
sufficiently improve correctness though it reduces
completeness. [In case 4, improvement of
correctness is significant.
connected. In the original paper, probability of chains are
experimentally calculated and used for validating segment
connectivity to avoid hard threshold value. In our study,
however, hard threshold method is applied to check
effectiveness of grouping process.
Table 4 shows a result of grouping process. The simply
minified image was used. Parameters for the edge extraction
process are 1.0 for sigma, 7.0 for edge seeding threshold and
0.5 for edge linking threshold. 1108 segments are extracted as a
result.
In all cases, segments longer than 5 pixels are used for next
iteration. In case 1, geometric relation only is considered to
investigate line connectivity. The process converges at 4
iterations. In case 2, the segments more than 10 pixels are
picked up and evaluated. In case 3 & 4, photometric relation is
also considered. It is very clear that cut-off of small segments
considerably improves correctness. If photometric information
is also considered in grouping process, degree of improvement
in correctness becomes greater. In case 4, where photometric
information is used in grouping process and small segments are
deleted after convergence, correctness becomes 50% though
completeness drops to 30%.
Table 6 shows another result of grouping process. The original
image (Figure |) is used. Parameters for the edge extraction are
1.8, 4.0 1.0 for sigma, sceding threshold and linking threshold
respectively. The segments shorter than 7.0 pixels are not
accepted. 1433 lines are extracted as the grouping process.
4. CONCLUSION
A road extraction method from ALOS PRISM image have been
investigated and tested. This process consists of three stages;
pre processing stage, edge extraction stage and grouping stage.
In pre processing stage, an automatic brightness threshold
method were applied though climination of brighter / darker
area slightly improved final result.
In edge extraction stage, a line centre diction method has been
introduced and evaluated. Better combination of parameters has
been searched but correctness was very poor. It suggests the
need of some filtering method to distinguish true road segments
from false ones. Some minified images have been also tested
but their correctness and completeness were as same as that of
original image.
In grouping stage, use of segment length and radiometric
property condition significantly improved correctness though
completeness drops by 50%.
As a whole, more improvement is needed to rise both of
correctness and completeness. One idea to achieve it is
excluding building area and vegetation area in pre-processing
stage. Another idea is using GIS road data as a guide and
classifies extracted road candidates. Hu and Tao (2002) use the
concept of saliency. They applied segment grouping to “the
most salient" lines and later “less salient” lines were added and
grouping process was continued. This method might be worth
to consider for this study.
5. REFERENCES
Baltsavias, E.P., 2002. Object Extraction and Revision by
Image Analysis Using Existing Geospatial Data and
Knowledge: State-of-the-art and Steps Towards Operational
Systems. IAPRS vol. 34, Part 2, pp. 13-22.
Baumgartner, A., Steger, Mayer, C., Eckstein, H. W., and
Ebner, H., 1999. Automatic Road Extraction Based on Multi-
Scale, Grouping, and Context. PE&RS, 65(7), pp. 777 — 785.
Crevier, D., 1999. A Probabilistic Method for Extracting
Chains of Collinear Segments. Computer Vision and Image
Understanding, 76(1), pp. 36-53.
Hu, X., and Tao, C. V., 2002. Automatic Main Road Extraction
From High Resolution Satellite Imagery. IAPRS, vol. 34, part
2, pp. 203-208.
Internatic
LEE
Park, S..
Algorithi
Proceedi
Price, K.
vol.32, P.
Steger,
Structure
Treash,
Detection
Computi
Zhao, H.
Semi-aut
Image. F
pp. 406-:
The auth
Sports, S
author a
member
opportur
useful di
