International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
  
a. Part of a SPOTS P image (5m) over 
Sint-Niklaas area (sub-urban) 
  
b. Part of a SPOTS P image (3m) over Brussels 
area (rural) 
Figure 1: Test Images: view of the complete test 
zone and local zoom 
Ideally, the “ground truth” image size should be 
equal to the SPOTS image. However, the scale of 
the DB is approximately 10 times finer so that ob- 
jects of different nature projected at SPOTS5 resolu- 
tion may overlap. À raster 100 times larger than the 
SPOTS image was thus necessary. 
On the other hand, the detected objects are also 
projected on the raster. Dots and lines are dilated, 
while polygons are filled up with pixels marked as 
BUA or RN. 
Traditionally, in a confusion matrix, the first row 
presents the ground truth classes (à), while the first 
column represents the detected classes (7). Usu- 
ally each cell (4,7) of such a matrix presents N;;, 
the number of pixels detected as belonging to class 
j while belonging to the ground truth class i. The 
matrix interpretation is made easier by a normaliza- 
tion by the total number of pixels. For display pur- 
We use a slight variation of such a matrix, adding in 
gach cell £7; and Ci; where Ru = Ny; / X Ni 
and Ci; — Nij/ 32; Ni;, the number of pixels nor- 
malized according to a row and to a column, re- 
spectively. When à = j, R;; and Cj; represents the 
user’s and producer’s index respectively. 
  
  
  
i 
j | Nues 
  
  
  
  
  
Table 1: À cell of our confusion matrix 
The confusion matrix of the first and second test are 
shown in table 2 a and b respectively. In each cell of 
the table 2 a the first and second line represents the 
normalized numbers generated by the expert and by 
the novice respectively. Only the expert worked on 
the second test zone. 
  
  
  
  
  
  
  
  
  
  
  
  
Truth — BUA RN NO 
Viewed | 
BUA 10752 ..075 | 358 121 4827 393 
7031... 955 | 150 — 19 | 3898 ES 
RN 283 01:2. 4. l-4797.; "S$S.di2093 5298 
12.6 ‘ 48.8 ‘ 38.6 
697 bo s [27131148 ck aus. SEE 
NO Wise 225 | 3837 2771531205 $93 
14265 207 | qoa5 118 | 32028 903 
Global index: 46.8% 
42.6% 
a. Test over Sint-Niklaas region (sub-urban) 
Truth — BUA RN NO 
Viewed | 
BUA 31310949 13-4085 500. ©. 
I 3.1 « 23.6 73.3 
RN so 711} 3990 29 928 5 73:3 
NO 1822, 195 | 1446 . 197 | 7592 . 999 
  
  
  
  
  
Global index: 8.296 
b. Test over Brussels region (rural) 
Table 2: Confusion matrix. BUA- Built-up area; 
RN- Road network; NO- Nothing. 
The false alarms are found in the last column of 
the table while the last row is made of the lacking 
elements. An ideal matrix would be diagonal. Sev- 
eral indexes of classification accuracy can be de- 
rived from such a matrix (Tso and Mather, 2001). 
To the producer's accuracy and the user's accuracy 
found in the diagonal, we have add the global in- 
dex. 
pose, a cell in our table contains N;; — Nij/ 2 Nij. A rough analysis of the matrix shows that: 
986 
Intern 
pe 
gl 
ure 
Fig 
Re 
url