ND45 = (TM4 - TM5)/(TM4 + TM5)*128 + 128,
The image values were generalized for the forest stands
by computing the mean image value for each stand.
These computations were performed in a raster-based
GIS environment. Boundary pixels and known roads
were omitted from the computations. The computed
values were stored for each forest stand, as an
additional attribute in the forest data base.
The correlation between crown closure and ND45 was
tested for a set of undisturbed, SE-facing (illuminated)
forest stands (N=71), and a linear regression function
was computed and tested for preliminary evaluation.
Forest stands with a potential for logging were extracted
from the database and considered as candidates for
change (N=228). Change detection was carried out by
robust regression between map attribute crown closure
and image attribute ND45. The algorithm has been
described in Baarda (1968). Outliers were automatically
detected in the regression. Outliers with unexpectedly
low ND45 values were labeled as logged forest stands.
4.3 Results and discussion
It was demonstrated that crown closure and ND45 were
correlated on a polygon level for undisturbed forest
stands (figure 7). The regression coefficient for the linear
regression function was 0.97. The two data sets had low
correlation on a per-pixel basis. This can be explained by
the natural heterogeneity of the landscape, and by the
influence of factors which could not be kept under
control, such as understory vegetation or secondary
species in the canopy.
* 1 std dev
mean unchanged
forest stands
- 1 std dev
ND(4,5)
+ 1 std dev
mean logged
forest stands
90 - 1 std dev
Crown Closure Class
Figure 7. The relationship between crown closure, as
derived from the map, and ND45, a standwise image
measure of forest density. The means and the standard
deviations for undisturbed and logged forest stands
indicate the separability between the two groups of data,
for the various crown closure classes.
10 out of 10 completely (> 60%) logged forest stands in
the test area were detected, with one error of
commission. Confusion with rock outcrops and forest of
low density was avoided. The influence from natural
spectral variation was minimized.
388
Partially logged (< 60 %) were not detected. As
discussed previously (section 3), standwise image
means should not be computed for these stands since
they contain at least two statistically separate
populations of pixel values. Some clustering algorithms,
combined with fragmentation index computations were
tested to initially detect and identify partially logged
forest stands, with promising preliminary results.
4.4 Conclusions from the case study
The polygon-based approach radically reduced the
problems related to landscape heterogeneity, in
comparison to multi-spectral classification. In
combination with the GIS-based sub-selection of forest
stands, it also reduced the confusion with features of
similar spectral characteristics.
The case study underlines the need for improved image
generalization functions, especially to deal with
spectrally heterogeneous map polygons.
A revision system is envisioned where polygon-based
change detection is the first step. The result, map
polygons which exhibit strong spectral anomalies, is
signalled to the operator, who will make the final decision.
The second step will be semi-automatic geometric
adjustment of polygon boundaries. Another advantage of
the polygon-based method is that existing polygon
boundaries can be used as first approximations in this
process. Preliminary investigations are being undertaken
in this area.
5. OPERATIONAL MAP REVISION
Currently, semi-automated revision of geographical
databases from satellite image data requires that raster
and vector GIS are operated in parallell. The map with its
attribute database is stored in a vector GIS. It is
converted from vector to raster format and imported to a
raster GIS, where it is combined with the image to
computer generalized image values. These region-based
values must be exported back to the attribute database
of the vector-GIS, for integrated analysis with the other
polygon attributes.
This situation would hardly be acceptable in an
operational situation. Too much time is spent converting
data back and forth, and it is required that parallel
databases (raster and vector) are maintained. There is
an apparent risk of introducing errors in the data.
Several authors (e.g. Star et al., 1991; Ehlers et al.,
1991; Goodenough, 1988) have suggested development
of high level data management systems or expert
systems to perform low level tasks such as format
conversion in a manner transparent to the user. This
level of integrated analysis, where two systems are
working in parallell under a common user interface Is
referred to as "seamless integration" by Ehlers et al.
(1989).
Integrated systems should certainly be a development
goal, but perhaps they lie far in the future. Most
geographic databases have been created and are being
maintained in vector systems. In addition, a large number
of data bases with environmentally related point
measurements exist (e.g. runoff, precipitation, ground
water quality), which can be linked to vector GIS based
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
on tl
proc:
on tl
provi
inco.
espe
conn
Futu
the L
value
this
GIS
imag
as s[
The
Dr. C
for hi
The !
Swec
The :
Colu
Mins
Cana
Than
Geoc
pract
BANI
J., L/
Fore:
basis
56.1
BAAF
Geod
BCM
1991
Reso
BCM
1992
Interp
EHLE
Integ
infor
Photc
55, N
EHLE
1991.
data :
Sens.
GOO
fields
Good
Lond
