Roeland de Kok 
  
  
        
  
  
  
  
  
  
  
                
  
  
  
  
  
  
      
  
  
  
  
  
  
  
   
  
  
  
  
  
  
  
  
  
  
  
  
Coniferous-old *Y | RT eae) Feature Selection Coniferous-open + | nt Feature Selection | 
Std.Dev. ortho.bmp (S) Mean spot3.bmp (NN) 
[10.8 - 28.9] StdDev.: 5.95 4.2 - 121.4] StdDev.: 12.89 
jua da lu [23.5 - 66.7] StdDev.: 15.05 A ar ns AL 
00 144 288 43.1 3575 719 863 1006 115.0 Overlap : 0.00 140 435 4 1025 1320 1615 191.0 2205 250.0 Overlap : 0.21 
ue Mean spotd.bmp (NN) 
[45.5 - 71.2] StdDev.: 7.84 
I [37.5 -5177] StdDev.: 3.58 
180 43 745 1028 1310 1593 1875 2158 244.0 Overlap : 0.18 
Figure 2. Box classifier for Coniferous-old, adapted Figure 3. Fuzzy logic decision curves for SPOT-4 
box classifier for coniferous-open (user defined). Mean values of band 3 and 4. 
Image Object Information: Features and Classification IMEI - image Object Information: Features and Classification =] Ix 
Feature | without | with Current Classification Feature ] without | with Current Classification 
Msi Mb M 923 923 [2B Codoowmide x] Dn SOs Cima we me [EF mm 3 
io tho bmp © 31.86 9 >a anv vk) EC nr 
ined on M 2 is Altemative-Assignments Mean Diff. to Scene ndvibmp : 17.90 892 892 omative Assi S 
Mem sut ue ss + $3 wihou à vit class-related Features Sue. Pit ix Jog mp ated ras ve 
Mean spot2.bmp : 11.81 974 974 En e ER e Std.Dev. spot2.bmp : 1.11 1000 1000 
Mean spotl.bmp : 14.07 984 $3984 Std.Dev. spotl.bmp : 1.38 1000 1000 Fe td 
Mean ppm Bus’ 95303 00, Brosdeavedycung StdDev. ottho.bmp : 31.86 843 843 en 
ean fistmomen 5 ‚2 eds i .0 Mixed-st 
Mes MD stone Mhobap: 872 97.3 973 0) (00 Bosdeved dd Spes dione: 43 5, 0, Mos 0.0 Brosdeved-ok 
Mean Diff. to Scene ndvibmp : 17.90 983 983 Mean spot3.bmp : 77.67 706 786 ; 
Std Dev. spotd.bmp : 4.08 100.0 100.0 ; Mean spot2.bmp : 11.81 730 730 00 00 Chabl: 
Std Dev. spot3.bmp: 7.92 1000 1000 00 QU Eros Grass Mean spotl.bmp : 14.07 798 798 00 00 ÉtosionGrass 
Der Mee pai i Jog 108 00 0.0 Coniferous-voung Mean ndvibmp : 204.65 905 905 00. 00 Conlerussoung 
Pa bp: 4.94 no 100 0) no Conilemus ere terms Mean fustmoment.bmp : 214.57 999.-.999 00 0g Confer tros 
dore 00 00 Meadowehighndvi CI 00 0.0 Meadow-high-ndvi 
Stim LL uz mo wo | 30 GigE | 
and (min) 1000 1000 Level: 3.00 100.0 100.0 
Level: 3.00 100.0 100.0 Sample-Assigment Sample-Assigment 
Display Settings Browse Image Objects d Display Settings Browse Image Objects 
[ Class Evaluation _up | down| | | Class Evaluation _w |down| | | 
Figure 4. A selected object is correctly classified as Figure 5. The 38.6% membership of coniferous-open 
coniferous-middle, second possibility is membership of | depends upon the value in SPOT 4- band 4 (Infra red). 
coniferous-open with 38.6 %. The lowest value counts (‘fuzzy AND as minimum). 
in 2D Feature Space Plot lit E3lll20 Feature Space Plot 
Std Dev. spot4.bmp Std Dev. ortho.bmp 
all 4557 86.92 
on, * 
is [levés +] levd3 + 
Level in object . Level in object 
gh heerachy " . hierarchy 
is. Feature y Axis i 7 . . . Feature y-Axis 
ith (see) Sila a af 
of bho Srey Foe 
is. 
ect Coordinates x/y : Coordinates x/y : 
res ^ i: 
the a idt 
P 
ect ^ V . os » m u : 
000 . ei . : 1.32 
jor Le Jr 2 Mean spat bp an dee | 28 
Figure 6. The SPOT-4 Band 4 infra-red, mean object Figure 7. The same graphic as figure 6 using the 
values against Std.Dev. (Intensity vs. surface- panchromatic band. Gaussian distribution can not be 
roughness) There is no need to assume any correlation. | assumed. 
i Gaussian distribution can be assumed and in that case, 
the distribution should and is showing concentric 
circles of increasing density from periphery towards 
is the center. 
al 
es 
S EVALUATION 
1e 
The evaluation of the classification results is more complicated for object oriented analysis than a per-pixel 
of classification evaluation, using a confusion matrix. Actually the traditional confusion matrix or K-factor (Richards 
1999) for class evaluation is possible but too simple for object evaluation .A single large false-classified object has a 
huge impact on the K factor in comparison with many small correctly classified objects. Of course, evaluation 
according to visual check, as applied in traditional photogrammetry is always possible. Before overall accuracy 
assessment, individual objects are evaluated over their specific features. As larger overlap exist among sub-classes, here 
an example is given for a selected object in the sub-class coniferous-middle.. A specific object , not in the training set, 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 227