Edward M. Mikhail
: : Reformatting. The new features defined above are The
Operation CPU Time (sec.) Analyst used to create a 9 band data set consisting of the that
Time first nine of the new features, thus reducing the OV
Display Image 18 dimensionality of the data set from 210 to 9. pret
Define Classes < 20 min. boun
Feature Extraction 12 Initial Classification. Having defined the classes sche:
and the features, next an initial classification is
Reformat 67 ; ; ; :
— : : zd carried out. An algorithm in MultiSpec called The
Initial Classification ECHO (Extraction and Classification of ob
Inspect and Add 2 . 5 min. Homogeneous Objects) is used. This algorithm is on th
Training Fields a maximum likelihood classifier that first
Final Classification 33 segments the scene into spectrally homogeneous In ge
Total 164 sec = 2.7 min. . 25 min. objects. It then classifies the objects. type
Table 3.1 ; : : On tl
Finalize Training. An inspection of the initial Hill :
classification result indicates that some improvement in the set of classes is called for. To do so, two additional training
fields were selected and added to the training set.
Final Classification. The data were
again classified using the new training
set. The result is shown in Figure 11.
Note that the procedure used here
does not require complex
a * preprocessing such as correction for
Figure 11. Orthorectified Maximum Likelihood Classification atmospheric effects, absolute
calibration, transformation from
radiance to reflectance, etc. The analysis required approximately 2.5 minutes of CPU time on a Power Macintosh G3
machine and 27 minutes of analyst time. n
do
3.2 DEM Supported Hyperspectral Analysis }
This is an experimental study where using a fusion of two essentially different types of data proves significantly iis
superior to the individual use of either one or the other. The task is to identify and accurately delineate building roof-
tops in the flightline of hyperspectral data of the Washington D.C. Mall, supplemented with digital elevation model
(DEM) data for each pixel of the scene, Figure 12. thresl
desig
i Experiments using gradient-based certai
a a E P algorithms on the DEM data show that
i its use alone is not sufficient to sharply Since
delineate building boundaries. À deteri
"TA e am ET spectral classifier does not have region
Figure 12. Digital Elevation Map (DEM) boundary problems. However, building Centr
roof-tops in this urban scene are Pixels
constructed of different materials and are in various states of condition
and illumination. This and the fact that, in some cases, the material Zonin
used in roof-tops is spectrally similar to that used in streets and going
parking areas make this a challenging classification problem, even for
hyperspectral data.
It is shown here that combining hyperspectral and DEM data can
substantially sharpen the identification of building boundaries, reduce
classification error, and lessen dependence on the analyst for classifier
construction.
The information content in the DEM is in the rise in elevation of à
s XM "bis x Hd given area-element in relation to its neighbor. The use of a gradient Thres
E UE. - operator in identifying building pixels (presumably at higher elevation 1S con
p. pit Fit l'E: D enit than ground level) is thus appropriate. A Sobel gradient operator wa groun
Figure 13. Gradient Operation Section used on the DEM. A gradient-threshold was applied to the result t0 the ele
of DEM obtain Figure 13.
598 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
