18 forward
. It can be
ge based
)ported by
are either
' extracted
a relations
a relations
manually
images of
) positions
to slope.
vhereas it
/ 2 and 11
embership
lope“, the
land
are in the
jerator can
conclusion
the same
f a single
n of each
lass is thus
1g a high
est classes
single rule
istrated in
> will keep
on further
is given a
iated with
e highest
uzzy mode
1embership
each cell.
> rules give
> user can
) functions
embership
isfy any of
Table 1. Command structure in guide (forest in the example can either be classified as a lumped category, or the two observations
can be separately classified, making it possible to include several forest types in one rule).
command Followed by Example - Boolean logic Example - fuzzy logic
WHENIMG - category "map" =3 4 soil =3 4 soil
< category "map" >0 300 DEM >040 300500 DEM
> category “map” <30 20 slope <5030 3020 slope
@ category TO category “map” @ 10TO 40 20 TO30LAI (2010 TO 4050 1020TO 3040 LAI
+ row nr TO row nr +0TO 150 100 TO 250 +00 TO 150300 50 100 TO 250 300
* column nr TO column nr * 100 TO 500 400 TO 600 * 50 100 TO 500 600 200 400 TO 600 700
SAVEIMG # category “category name” # 5 forest # S forest
3,2 Classification parameterisation
All classifications were based on the same set of training
data. Procedural rules were used for initial rule structuring in
both methods. Performance of the methods were then
iterativley improved by manual changes in rule structures.
The maximum likelihood classification was based solely on
four TM bands from the April image (3, 4, 5 and 7).
Initial guide rules were derived from a set of training data
including four bands in the April image (3, 4, 5 and 7), the
first component of a principal component analysis from 12
bands of the two TM scenes (excluding band 1, and holding
78 % of the variation from the other 12), one image of LAI
(Leaf Area Index) from the June image, LAI difference (i.e.
growth) between June and April, and wetness from the April
image (cf. McCarthy, 1996). The latter images were used
because of their physical interpretability. The result of the
initial classification was tested against the training data and
visually inspected. Wetlands turned out to be the most
problematic category to classify. Thus the DEM was used to
produce an image of updrain feeding areas to each cell
(Desmet and Govers, 1994). Wetlands were then divided in
ombrogenic raised bogs (with no or low updrain feeding
areas) and topogenic fens (with membership increasing with
updrain feeding area). Problems also occurred both between
different vegetation classes and between vegetated and non-
vegetated classes (i.e. urban areas). Thus the rules for growth
(LAI-difference) were altered to be higher for farmland and
deciduous trees and lower for urban areas, grassland and
coniferous trees respectively.
4 RESULTS
The expert classifier gave the best result, with a highest
kappa index of 0.7521 (table 2). This classification was
based on the automatically extracted training data from the
first PCA component, the LAI difference and bands 5 and 7
from the April scene. Excluding the two raw bands gave
almost the same classification accuracy (0.7506), as did also
inclusion of all images. Just using the four April TM bands
gave a kappa index 0.74. Manual changes in rule structure in
general did not improve classification accuracy. The best
result for the maximum likelihood classification using the
four TM bands was a kappa index 0.71 (table 3). The result
of the expert system classification is also shown in
simplified form in fig. 2.
Table 2 Error matrix for the expert classification (rows) against ground truth (columns) (kappa index = 0.75)
Water Wetland Coniferous Deciduous Grassland Crops Bare fields Urban Total
Water 4484 10 0 0 0 0 0 0 4494
Wetland 0 55 116 51 0 0 0 0 222
Coniferous 0 38 3642 23 0 9 0 0 3712
Deciduous 0 112 431 421 9 3 6 130 1112
Grass land 0 3 20 6 438 269 3 1 740
Crops 0 4 0 20 113 574 15 78 804
Bare field 0 2 0 1 28 2 172 5 210
Urban 0 15 13 107 148 762 10 1058 2113
Total 4484 239 4222 589 629 1619 206 1257 13407
Table 3 Error matrix for the maximum likelihood classification (rows) against ground truth (columns) (kappa index = 0.71)
Water Wetland Coniferous Deciduous Grassland Crops Bare fields Urban Total
Water 4484 2 0 0 0 0 0 0 4484
Wetland 0 78 309 100 0 0 0 8 465
Coniferous 0 0 3690 0 0 0 0 0 3362
Deciduous 0 25 129 308 2 0 0 24 458
Grass land 0 0 0 11 564 432 0 2 1009
Crops 0 0 0 24 85 661 39 41 822
Bare field 0 0 0 2 24 0 148 50 224
Urban 0 123 93 144 33 18 19 1132 1532
Total 4484 239 4222 589 629 1619 206 1257 13407
185
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
