5 CLASSIFICATION
One of the most wide-spread applications for multispec-
tral images is ground cover classification. The classifi-
cation of all image pixels into k classes can e.g. be per-
formed by virtue of the spectral distance between the
spectrum of a given pixel and the class reference spec-
tra. In particular for images of vegetated areas it is
also of interest to classify by virtue of derived features
such as the NDVI, the NIR-reflectance and the local
spectral variance. These three features were used for
the following check of the classification accuracy of the
sharpened PSM4-imagery in comparison to the coarser
MSa4m-imagery.
The accuracy of a land cover classification with re-
spect to the ground truth is described by virtue of
the kappa coefficient x € [—1, +1] (Congalton 1991,
Richards 1993). The kappa coefficient is x = 0 for
the pure coincidence between two totally random clas-
sifications, and reaches x = 1 for complete agreement
between classification and ground truth. In our case,
we consider the classification result of the 1m multi-
spectral MS,,-image as ground truth, and compute the
kappa coefficient for the simulated spaceborne MS4m-
image and for the fusion-sharpened PSM 1m-image. The
MS1m-, MS4m- and PSM n-images were classified by the
Minimum Euclidean Distance method with the same ref-
erence spectra for k = [2...12] classes. The reference
spectra were established by unsupervised k-means clus-
tering (Richards 1993, Wiemker 1997). The resulting
classification accuracy kappa coefficients are plotted in
Fig. 5, top.
Generally, the kappa coefficient decreases with increas-
ing number of classes; i.e., the finer the classes are cho-
sen, the higher the probability of mis-classification be-
comes. With respect to the fusion sharpening, however,
we observe a significant increase of the classification ac-
curacy for the fusion-sharpened PSM n-image compared
to the coarser MS4m-image, particularly for a higher
number of classes k.
5.1 Land Cover Area Assessment
The results of land cover classification are often ex-
pressed in terms of the total area covered by a certain
ground cover class, or a group of ground cover classes
(e.g. the area of sealed surface, vegetated surface, build-
ing occupied surface etc.). Therefore it is of interest to
measure how large the errors in land cover area assess-
ment are: according to classification of the coarse MS4n-
image, and according to classification of the sharpened
PSM m-image, both compared to the classification of the
true 1 m resolved multispectral MS,m-image.
The results in Fig. 5, bottom, show the mean deviation
from the land cover areas as assessed from the true 1m
resolved MS,n-image as percentages of the area of the
whole scene. We see that for more than k = 6 ground
cover classes, the error in area assessment is originally
at =~ 3% and cut down to half by the fusion-sharpening.
Correlation With True MS;,-Features
100 J NIR reflectance NDVI Local Variance
mma
904 | F =
zz
804 ZZ
=
= [=
701 5 mz
z | =z
Ls =
E
504 | EI Zz =
| As
404 | FA
| E
: z
304 = ZZ
ZZ z
20 zz zz
ZZ Zz
" ZZ z
ZZ ZA
0 = =
MS,, MS. PSM PSMis MS«, MS, PSM PSMis
LI MS4m MSinterp = PSM im DE PSMim, 4-16
COMPARISON TRUE VS. FUSED IMAGE
MSim MSntep PSM PSM
€ [0. 1] mean dev. | 0.0441 0.0419 0.0305 0.0228
correlation | 86.6% 88.4% 93.2% 96.3%
Red reflectance
€ [0.1] mean dev. | 0.0103 ’ 0.0099 0.0067 0.0063
correlation | 87.2% 90.4% 97.1% 97.3%
Green reflectance
€ [0.1] mean dev. | 0.0103 0.0099 0.0067 0.0045
correlation | 86.4% 89.5% 97.4% 98.2%
NDVI
€ [- 1. 1] mean dev. | 0.054 0.053 0.056 0.054
correlation | 93.0% 93.6% 93.5% 93.3%
root mean local variance
€ [0.0.5] mean dev. | 0.0168 0.0199 0.0049 0.0049
correlation | 71.1% 74.4% 95.0% 96.2%
MSam multispectral imagery with 4 m pixel size (GSD)
MS, rp © multispectral imagery interpolated to 1 m pixel size, using cubic B-splines
PSM : panchromatic sharpened (fused) multispectral imagery
PSM; : panchromatic sharpened multispectral imagery,
fusion algorithm discriminating 16 spectral classes (unsupervised clustering)
Table 2: Comparison of spectral truth, NDVI, and local spec-
tral variance between the original, the interpolated, and the
fusion sharpened multispectral imagery. The truth is evaluated
by checking against the 1m resolution multispectral airborne
imagery from which the expected satellite products were simu-
lated.
290 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
Kappa Coefficient
Figu
asses
fusior
the s
classi
High
lm
ing :
resol
were
mult
panc|
tral i
ened
paris
deriv
ance
the fi
the a
shary
We cc
does
but t]
of the
classi
