PSM PSM 16
{ PSM:
5 0.0228
Jo 96.396
7 0.0063
% 97.39
7 0.0045
% 98.29
6 0.054
^ 93.39
9 0.0049
7 96.29
splines
lustering)
al spec-
and the
raluated
airborne
re simu-
Classification Accuracy
1 .00 T T T T T T T T T T
0.90 - A PSM —
0.80
0.70
Kappa Coefficient
0.60} B0 9 A
gi $4 J
0.40 1 } | À 1 1 x 1
200350. ZUG Gin Pi vBoIn Qe 10 i
Number of Classes
À 1
1$4142::18 14
Error in Land Cover Area Assessment
5 F T T T T T T T T T T T
[
2
j
1
" C) MS4m 1
9 | 3
s | ;
E at nit i
a | 3
3 i 1
=
© 2t 1 j
3 2}
« E i
S 4 | 3
© - 1 À
SR ug | i
E uy ud
ot | | A À L À A i À 3
2 3 4 5 6. 7, +8 9 10:21:14 412 18.14
Number of Classes
Figure 5: Classification accuracy (top), and error in land cover
assessment (bottom) for the original MS44-image (DJ) and the
fusion-sharpened PSM;,,-image (A). The error margins give
the standard deviation from 4 different runs of unsupervised
classification with random starting seeds.
6 CONCLUSIONS
High resolution satellite imagery (4 m multispectral and
1m panchromatic, MSs, and PAN,,,) was simulated us-
ing airborne scanner imagery with 1m multispectral
resolution (MSım). Three different fusion algorithms
were applied to the simulated data in order to obtain
multispectral imagery sharpened by the higher resolved
panchromatic band (panchromatic-sharpened multispec-
tral imagery, PSM). The spectral truth of the sharp-
ened PSM, ,-imagery could be directly evaluated by com-
parison to the original airborne MS,m-imagery. Also
derived features such as NDVI and local spectral vari-
ance could be compared between the original coarse ,
the fusion-sharpened and the true 1 m imagery. Finally
the accuracy of land cover classification on the fusion-
sharpened has been evaluated.
We could show that the panchromatic fusion sharpening
does not only improve the eye appraisal of the images,
but that it does also substantially improve the accuracy
of the spectral reflectance values, and subsequent image
classification.
The improvements habe been quantified for an exem-
plary vegetation scene. Although the particular values
do of course depend on the specific scene content, we con-
sider the magnitude of the improvement obtained by the
panchromatic sharpening as typical for MSam/ PAN1m-
fusion.
Findings of particular interest are:
» The MSinter-interpolation of the MS4m-image by
cubic B-splines shows a slight improvement with
respect to the spectral truth of the MS1m-image.
» Fusion preserving the relative spectral contribu-
tions (section 3.3) is significantly closer to the
spectral truth than fusion by HSV-transformation
(Prinz et al. 1997).
» The fusion by relative contributions is improved
when the MS44-image is first classified by unsu-
pervised clustering and the fusion carried out sep-
arately for each spectral class (section 3.4). The
class specific fusion improves specifically the accu-
racy of the NIR band. With this fusion method the
correlation of zz 8596 of MS44 is improved to zz 9596
for the PSMım. The mean deviation from the true
reflectance values is diminished to approximately
half.
» The feature NDVI of MS44 has a correlation of
> 90% with the true MS,n, and a mean devia-
tion of only 0.05. The NDVI is not affected by the
sharpening, since the relative contributions of the
NIR and Red bands remain unchanged by the fu-
sion method. Thus we have the same values for
PSM 1m-
» The local variance feature as the simplest tex-
ture feature is significantly improved by the fusion
from x 75% to zz 9595 correlation with the true
values.
» Comparing the results of unsupervised land cover
classification on the MS4m-imagery and the sharp-
ened PSM,nversus the true 1m MS,m-imagery
we find that the fusion improves the kappa-
coefficients of classification truth by 2096 (for the
number of classes k 7 6) and reaches « = 70-80%.
» The error made in the assessment of the total
area covered by a certain class is approximately
cut to half and as small as < 1.5% after fusion-
sharpening.
ACKNOWLEDGEMENT
This work was supported by the Volkswagen-Stiftung.
The image flights were conducted in collaboration with
the German Aerospace Center (DLR Weflling, München),
particularly with the help of Volker Amann, Peter
Hausknecht, Rudolf Richter and Manfred Schröder.
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7. Budapest, 1998 291
