John Bosco Kyalo Kiema 
  
Through the use of an appropriate threshold function, the floating wavelet coefficients are approximated into 
corresponding integer values. Incidentally, it is this quantisation process that constitutes loss of information in lossy 
compression schemes. The number of bits required to represent each quantised coefficient are then reduced through 
binarisation. 
4 SEGMENTATION RESULTS 
For the study presented here the wavelet compression software Lurawave is used. This is a software package developed 
precisely for the compression of remotely sensed imagery (Luratech, 1999). The fused imagery is systematically 
compressed (channel by channel) at compression rates (K) ranging between 5 and 100. The compressed fused imagery 
is then classified using the approach described in section 2.3. Results obtained are then analysed and their geometric 
and semantic quality evaluated. 
41 Geometric analysis 
Sixty (60) identical image points are manually measured in the classification results for both the original and 
compressed imagery. The same object points (e.g., building corners) are identified and positional errors estimated. In 
general, relatively large error values are obtained. This is basically because of the concomitant pointing problems 
encountered, especially with classification results at higher compression rates. Fig. 4 clearly illustrates this limitation. 
  
: . a y ; . . 1 Lu À = 
n 3 
. | * A 
(a) original (b) 10 (c) 50 
Figure 4: Pointing problems for various classification results 
A Fisher test is then carried out in order to establish whether the positional errors obtained for the classification results 
at the different compression rates (5) are statistically different from similar results for the uncompressed data (s, ). Tests 
are performed at 5% level of significance. Table 2 shows the general test scheme applied. 
  
Null hypothesis: H,:0=0, (= no difference) 
Alternative hypothesis: | H, :0 > 0, (= difference) 
Quantile (FISHER): F,1v2005 =1.53 (v 2v, 259) 
Test value: Z,-s Pg 
Test decision: Hy:Z,<F / H,:2;2F 
  
  
  
  
Table 2: Fisher test scheme for evaluating geometric quality 
For the original fused dataset, the positional accuracy is estimated at s , =0.44 Pixel (+ 0.66m). A comparison of this 
with similar estimates for different compression rates is illustrated in Fig. 5. From this it is apparent that the 
compression rate of 20 represents the critical compression rate. Beyond this value, the classification results for the 
compressed data are statistically different from those of the uncompressed data. 
  
492 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.