International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
The hierarchical classification method was more successful in
detecting agricultural, grass and urban areas. Except for the road
pixels which were extracted using vector coverage, the
superiority of hierarchical approach has been proved, since the
area was sub-divided into spectrally homogeneous region,
minimizing the risk of spectral confusion among classes.
5. CONCLUSIONS
Compared to the traditional multispectral classification
methods, the knowledge-based hierarchical classification did
improve the classification results. The water was found to be the
unique class, generated by both techniques, with similar
thematic output. All other classes involved some confusion due
to the spectral similarities (i.c., agricultural areas and urban,
roads and urban). No class with “coast” label was generated
with the maximum likelihood classification method because it
was created by spatial reclassification step of the hierarchical
method. Coastal regions were not included in the starting
classification scheme because high risk of confusion between
actual shore and urban and/or road pixels which would cause
some inland urban pixels to be labeled as coast and affect
resultant accuracy. Maximum likelihood classification detected
one fourth of the coniferous area, which are detected by
hierarchical classification.
The results showed the proposed hierarchical classification
approach is promising and has several advantages in
comparison to standard approaches:
e The domain spectral knowledge and other spectral
knowledge obtained from training data are provided in an
easily modified and understandable rules.
e Computationally expensive operations can be avoided by
restricting the channels involved in the classification
procedure. Although the dimensionality was increased in the
beginning of the approach, only those with the least
correlation and which would best define the target classes
were used in the rules.
e Integration of spatial characteristics of features with the
classification procedure helps to increase the understanding
of some classes confusable with others.
e [tis flexible when applying to geographically different areas.
Higher level, more general categories in the hierarchy would
remain constant across different types of terrain; only the
lower level nodes would be variable from one type of region
to another.
The disadvantage of the proposed hierarchical classification
approach is the requirement of the reliable training data. The
rules are extracted using the statistical information of the
training data so these statistics should be carefully examined
since gathering enough training statistics to adequate account in
order to be used as rules, is a difficult task.
Although the classification technique presented in this study
generally worked well, there is potential for improvement and
refinement:
e Additional ancillary data like detailed land use and land
cover maps would not only decrease omission and
commission errors for the forest cover type classifications,
but also increase the levels of classification, like classifying
the types of trees in the forest areas
e Multitemporal TM image of the study area, if combined with
additional channels like band ratios and transforms, would
help generating more reliable rules for hierarchical
classification method.
516
e More spatial rules should be generated. Most of the
frequently used spatial rules can be collected and
transformed into computer-accessible format.
REFERENCES
Anderson J.R., Hardy E.E., Roach J.T. and Witmer R.E. (1976)
"A land use and land cover classification system for use with
remote sensor data", Geological Surver Professional Paper, US.
Government Printing Office, Washington D.C.
Anil K. Jain (1989) “Knowledge-based segmentation of Landsat
Images", IEEE Transactions on Geoscience and Remote
Sensing, Vol 29, No 2.
Barnsley M.J. and Barr S.L. (1992) “Inferring Urban Land Use
from Satellite Sensor Images Using Kernel-Based Spatial
Reclassification", Photogrammetric Engineering & Remote
Sensing, Vol 62, No8, pp 949-958.
Bolstad V.P. and Lillesand T.M. (1991) “Rule-Based
Classification Models: Flexible Integration of Satellite Imagery
and Thematic Spatial Data", Photogrammetric Engineering &
Remote Sensing, Vol 58, No 7, pp 965-971.
Gruen A. and Li H. (1994) "Semi-automatic road extraction by
dynamic programming", International Archives of
Photogrammetry and Remote Sensing, Vol 30, Part 3/1, pp 324-
332.
Heipke C., Steger C., Multhammer R. (1995) "A hierarchical
approach to automatic road extraction from aerial imagery", in
McKeown JrD.M. and Dowman LJ, Integrating
photogrammetric techniques with scene analysis an machine
vision II, Proc. SPIE 2486, pp 222-231.
Johnson, K (1994) "Segment-based Land-use Classification
from SPOT Satellite Data", Photogrammetric Engineering &
Remote Sensing, Vol 60, Nol, pp 47-53.
McKeown Jr.D.M. and Denlinger J.L. (1988) "Cooperative
methods for road tracking in aerial imagery", Computer Vision
and Pattern Recognition", pp 662-672.
Ruskone R., Airault S. and Jamet O. (1994) "Road network
interpretation: A topological hypothesis driven system",
International Archives of Photogrammetry and Remote Sensing,
Vol 30, Part 3/2, pp 711-717.
Skidmore, A.K. (1989) "An expert system classifies Eucalypt
forest types using TM data and a digital terrain model",
Photogrammetric Engineering & Remote Sensing, Vol 55, No
10, pp 1449-1464.
Stewart J.S. and Lillesand T.M. (1994) “Stratification of
Landsat Thematic Mapper Data, Based on Regional Landscape
Patterns, To Improve Land-Cover Classification Accuracy of
Large Study Areas".
Wolter P.T., Mladenoff D.J., Host G.E. and Crow T.R., (1993)
“Improved Forest Classification in the Northern Lake States
Using Multi-Temporal Landsat Imagery”, Photogrammetric
Engineering & Remote Sensing, Vol 61, No9, pp | 129-1143.
