International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
c d. 
rr Si Ele Spectral Profile. 
m T 7000 T T T T TT 
5000 A | \ i L | \ 
Af i [ J 
| y^ | | | 6000} US | 
4000} / Pin 4 r 
: / V [ 
ET r m 9000 + 
E. jJ B 
P 300r / 4 ü 
[ ne 4000} + 
2000 4 L 
3500 À 
L/ 
D gs E: 5 i 16 "m a 25 x 
e. f. 
Figure 3. False colour images: (a) CASI-48 with R=0.948um, 
G-0.675um, B=0.456u. (b) CASI-32 with 
R-0.914um, G-0.620um, B-0.451um, Ground truth 
data: (c) for CASI-48 and (d) for CASI-32. The 
extracted spectra of building material: (e) for CASI- 
48 bands, (f) for CASI-32 bands. 
where Mincgy and Maxcgy are respectively the minimum and 
maximum values over the image. 
The result maps for each method have been obtained after the 
segmentation of the similarity images using an adapted 
threshold. The thresholds have been found graphically from 
ROC curves. After the application of this threshold to the 
similarity image, each connected component of the resulting 
binary image is detected, and regions with a surface less than a 
given threshold are eliminated. The remaining pixels constitute 
the final decision images. The resulted similarity and target 
maps for CASI-48 and CASI-32 images are illustrated in Figure 
4 and Figure 5. 
For a quantitative evaluation of the results, we retain two 
elements derived from the confusion matrix: the overall 
accuracy (OA), and the overall kappa (OK). The overall 
accuracy is calculated by summing the number of both target 
and non target pixels correctly classified and dividing by the 
total number of pixels. Because the OA is not a very complete 
and reliable criterion, the OK is computed with other elements 
of the confusion matrix (Rosenfield 1986) and presented in 
Table 2. 
52 
6. EVALUATION OF RESULTS 
From the result images (for both data sets), we can evaluate that 
the deterministic measures can be used for material 
identification and mapping. In urban area, the spectral 
reflectance of building roofs is corrupted by topographic 
effects. But results show a relative success in detection of 
materials due to the nature of the measures. On the similarity 
images, it is visible that the topographic effect is still present 
with the CEM, especially on the rooftops, while it has 
disappeared with the MSAM and the SVM, since these two 
measures are robust to linear perturbation. On the other hand, it 
is clear that the CEM has succeeded to separate efficiently the 
target pixels, as we see nearly two classes of pixels in the CEM 
similarity image. In contrast to CEM technique, the two other 
approaches are able to distinct non-target pixels surrounded by 
target material pixels. For example, single pixels corresponding 
to chimneys and roof windows are detected. 
From a quantitative aspect the CEM technique provides better 
results for both datasets (see Table 2). 
  
  
  
  
  
  
  
  
CASI-48 CASI-32 
MSAM | SVM | CEM | MSAM | SVM | CEM 
Overall 0.96 0.96 0.98 0.95 0.95 0.96 
Accuracy 
Overa 055 | oss oso Pos lom | 0% 
Kappa 
  
  
  
  
  
Table 2. Accuracy Parameters of applied Methods. 
7. FUSION STRATEGY 
Because of limits and for benefiting of all abilities of each 
measure, we decide to use a fusion strategy in decision level. So 
we have defined a new 3-D space in which, each measure is 
defined as an axis. In this space, we have applied each measure 
as a target binary map. Then we can imagine a cube in this 
space that the interesting points are the corners. In this space, 
we have four types of corners: 
e  (0,0,0) which is corresponds to non target pixels. 
e (1,000), (0,1,0), (0,0,1) which are correspond to 
detected pixels as target at least by one technique. 
e (1,1,0), (1,0,1), (0,1,1) which are correspond to 
detected pixels as target at least by two techniques. 
e (1,1,1) which is corresponds to pixels detected as 
target by all three techniques. 
Surely; if we decide to consider the first and forth types of 
corners, we are certain that all target pixels is detected and all 
non target pixels are rejected by our techniques. But here we are 
interested in benefiting of the other possibilities. Therefore we 
define a decision tree based on three rules explained below: 
RI: the output pixel is target if at least, 1 technique detects it. 
R2: the output pixel is target if at least, 2 techniques detect it. 
R3: the output pixel is target if at least, 3 techniques detect it. 
The figure 6 is helpful to explain these rules: A3 contains only 
the white point, R2 contains the white and black points, and R1 
contains the white, black and grey points. The result target 
maps are shown on the figure 7 for both CASI-48 and CASI-32 
images. Again, from confusion matrix, the overall accuracy 
and overall kappa are calculated for the fusion images, these 
parameters present in Table 3. 
Internatio 
  
  
  
Figure 4. 
Overall 
| Accurai 
Overal 
Kapp: 
T