ted into
in lossy
through
veloped
natically
imagery
'ometric
nal and
iated. In
roblems
ation.
n results
;). Tests
n of this
that the
; for the
John Bosco Kyalo Kiema
no significant
difference
0.5
Positional error (pixels)
0.0
20 30 50 70 100
Compression rate
Figure. 5: Geometrical errors for different classification results
4.2 Semantic quality
The efficiency of data compression algorithms can be evaluated in several different ways. For example, it is possible to
estimate the relative complexity of the algorithm, the memory required to implement the algorithm, how fast the
algorithm performs on a given machine, the amount of compression achieved, or even how closely the reconstruction of
the compressed data resembles the original data (Sayood, 1996). For this study, the assessment of the semantic quality
of the classification results is basically done using the last of the evaluation options given above. Further, both visual
and automatic methods are compared.
4.2.1 Visual interpretation
There is hardly any significant visual difference between the classification results from the original fused imagery and
the compressed imagery for compression rates of up to about 30. However, for higher compression rates (K 250) the
smoothing effect characteristic of wavelet compression progressively increases and the quality of the classification
results quickly deteriorates.
4.2.2 Automatic interpretation
Analysis of the grey value histograms results in the popular quantitative parameter, the Peak-Signal-to-Noise-Ratio
(PSNR). This describes the relationship between the maximum grey value within the original image and the noise
resulting from the compression. Considering an image of dimension (nxm), the PSNR, expressed in decibels (dB), is
given by:
5
X peak
PSNR(dB) -101og| ——-
noise
n-l m-1
with noise = 1 SS -gu»
(n.m) i-0 j=0
where X re is the maximum grey value within the original image
ak
8, 7 grey value for the (;, j)zh element in the original image
g’, - grey value for the (i, j)th element in the compressed image
The relationship between the PSNR and the compression rate for the different classification results is shown in Fig. 6.
The level of noise increases with increasing compression rate up to a compression rate of about 15. Consequently, the
PSNR decreases accordingly. Between the compression rates of 15 and 20, no change in the PSNR is noted. At
compression rates higher than 20, the smoothing effect of wavelet compression becomes apparent. As a direct result of
this. the level of noise reduces and the PSNR increases conformably. However, after a compression rate of about 30 the
level of noise increases once again. Subsequently, this results in lower PSNR values most probably due to the instability
of the classification results.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 493
