tion. The
ed. In the
ular focus
data such
combined
lace them
t they can
indicated
d perform
he results
on.
993]. The
litions are
17e of the
tablished,
tion at the
the image
eeded for
has to be
nd to get
nd other
| has to be
mproving
research
ome main
3]. When
ds should
ixels and
several
rpretation
his paper.
For more details see [Nordahl, 1994 and 1995]. This
paper focuses on the use of the classification method. À
two-step post-object classification is used as classification
method as seen in Figure I (see also [Nordahl 1995 and
1996]). When the result was analyzed from this first phase
of the project it was clear that many heterogeneous stands
were wrongly classified. Signature patterns for these
stands and their respectively training fields have complex
structure that did not fit well with a statistical classifier.
At this stage a phase 2 was proposed to investigate the
possibility to handle nonlinear signatures with the use of
neural networks in the pixel-wise classification. Neural
networks for remote sensing classification have been used
in several projects with various results during recent years
and some are summarized in [Dalen 1995]. A three-layer
feed-forward neural network with standard back-
propagation as training was used and compared with the
result from the traditional statistical classification in the
first pixel-based part.
2. TEST AREA AND USED DATA
2.1 Test area
As test area the Steinkjer municipality forest,
Ogndalsbruket (21 800 hectares with 8400 hectares
productive forests) in North-Trgndelag was chosen. It has
a representative variation for a "standard" forest in
Trgndelag according to stand sizes, age and tree
variations. The productive forest is the coniferous stands
which are dominated by Norway spruce (Picea abies) and
where Scots pine (Pinus sylvestris) has a smaller
commercial volume.
In a control area of approximately 650 hectares field
measurements were conducted to establish ground truths
for analyzing the results. The height above sea level
varied in the control area from 150 to 400m and the site
classes from low to medium-high. Mainly outside of this
control area there are a total of 45 training fields
representing approximately 3500 pixels (2140 hectares)
Which had been registered [Nordahl,Kjellsen 1994].
22 Data used
As it was required to obtain an image from the maximum
growth period when the spectral reflections are at their
highest both from the conifer stands and from the broad-
leaf trees in the conifer stands, we looked for an image in
the period from mid of June to late J uly, from the same
year as a recent forest inventory which was finished in
1991. A search of existing scenes led us to only one
Suitable, a SPOT2 XS scene from 24 July 1991 which
Was only slightly covered by clouds in the test area.
Landsat TM with its higher spectral resolution specially
in the infrared wavelengths is normally better suited for
543
vegetation analysis. But SPOT XS with its higher
geometrical resolution with 20x20m was interesting for
classification at the resolution level of stands. As the test
area contains a variation from almost sea level to
mountain areas and we planned to combine the image
with ancillary GIS-data, the scene was ordered with
geometric precision correction by the use of a digital
terrain model to UTM coordinate system.
The digital terrain model (DTM) was a standard product
from the Norwegian mapping authority with medium
density point data at approximately 90x90m (3x6"). A
Kriging interpolation from this data-set was also used to
calculate values for aspects and slopes for the image
pixels. A 5m elevation map was constructed and visually
compared with a 1:10 000 map to control the
interpolation. In areas with a smooth topography the
heights were normally correct within some meters and the
aspects and slopes had negligible errors. Naturally the
errors are largest in rapidly changing topography, but in
our test area this is limited to small parts concerning only
a few stands. Due to the sampling method of the original
digital terrain model the errors in the interpolated DTM
are at a minimum along 20m level lines on the 1:50 000
topographic map.
The previous forest inventory was from 1981 in 1:10 000
paper-copies from ortho-photomaps in the NGO
coordinate system. Stand boundaries and connected
attributes were digitized for the control area and for all
field-checked stands with PC-Arc/Info. Transformation to
UTM and then to the UNIX image processing system
(ERDAS) was made. This resulted in one set with raster-
data and one with vector-data.
3. METHOD
The inventory class system had been established after a
process which started with an inductive approach [Strand
1989] for analyzing the image data. The class system got
three criteria where the first was for size/age with one
class for clear-cuts and three other classes. The second
criterion was the average crown coverage of the stands in
two classes. The third criterion was a two level division
for the amount of broad-leaf trees in the coniferous
stands. Out of these classes a total of 8-10 classes was
used. When the significance dropped too low for the
objects, the program looked if any of the above divisions
were significant and let the object be classed into a more
general super-class.
The ground truth training fields have been analyzed and
divided into slope- and aspect-subclasses of the inventory
classes. A total of approximately 40 subclasses out of the
8-10 inventory classes was used in the pixel-based
classification. To handle the topographic effect in the
image slope and aspect values have been
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
