In the TM classification these effects are averaged, allowing
apparent trends to be seen more clearly. Two levels of riparian
forest may indicate a gradual transition from creekbed to the
dominant hardwood association. On the north and west facing
slopes east of the creek, the contribution of hardwood 2 is
generally overwhelmed by that of hardwood 1; while on the other
side of the creek the east-facing stand of hardwood 2 appears
more uniform than in the aircraft version. The presence of red
spruce in the red pine plantings is still evident, but as isolated
blocks where the spruce signature apparently exceeds that of the
pine. While detail has clearly been lost, the classifications are
quite similar. Evaluating the nature and impact of this loss of
detail will have to await our next field visit once the snow melts.
The TM classification using the color infrared bands was very
similar to the one based on all six non-thermal bands. There were
a few anomalies, such as road pixels and excess hardwood 2 in
the field and some missing hardwood 2 on the west side of the
creek. It would seem that the additional blue and mid infrared
bands make a contribution in identifying forest types.
The addition of the elevation to the color infrared aircraft bands
had a dramatic effect on the classification. The slopes on the west
(and to a lesser extent east) sides of the creek are delineated quite
clearly. This is shown in an intermediate shade of grey on the
west side of the creek; the east side was merged with
riparian/brushland. Highland areas which appeared as riparian in
the previous classification (and might be wet meadows) were
generally classified as hardwood. The degree of mixture of
hardwoods 1 and 2 decreased, with hardwood 1 generally
predominant. The same was true for red pine and red spruce, but
areas of red pine appear to have been lost. In general, the amount
of vegetative detail appears to be less. Although this might reflect
a more orderly classification, it may also represent missing
information. As additional bands are added to the classification,
their effect can overshadow the variability associated with the
previous bands. This may in fact clarify and improve the
identification of forest types or it may hide interesting and
potentially meaningful variations.
The effects described above were even more pronounced when
the aspect was combined with the aircraft bands to produce a
classification. The creekbed riparian community was essentially
lost, but the east facing slopes west of the creek are clothed in
almost uniform hardwood 2. Hardwood 1 generally
predominates over hardwood 2 elsewhere but the more subtle
ridge structure on the east side of the creek is revealed (shown in
a slightly darker grey). An area of hardwood 2 south of the
pasture, which appeared mixed on the basic classification and
almost disappeared on the elevation classification, is now a major
feature. Two of the red pine plantations are clearly visible but
one shows more as red spruce than red pine.
Preliminary Accuracy Assessment
A preliminary accuracy assessment was performed using the 28
Observation sites mentioned above (Table 1). The selection of
sites needs more refinement but several observations can be
made. The line of demarcation between hardwood 1 and
riparian/brushland needs to receive more study. Shadows cast on
the road as it approaches the bridge over the creek near the top of
the image produce a darker area which is misinterpreted as spruce
or pine. Otherwise, the classifications based on the digitized
aerial photograph alone and combined with elevation provide a
better correspondence to the ground truth. A more complete
accuracy assessment with a larger number of field observations
will be developed to provide a more definitive evaluation.
Table 1
Cover Type Classification based on
(Number of Sites) Air TM (6) TM (3) Air & Elev Air & Asp
Red Spruce (3) 3 2 2 3 2
Red Pine (2) 2 2 2 0 2
Hardwood1 (10) 8 5 5 7 9
Hardwood2 (2) 2 2 2 2 2
Riparian (8) 6 3 3 8 1
Pasture (3) 3 3 3 3 3
Total Correctly
Identified 24 $7 17 23 19
"Modeling"
"Modeling" preserves the full spectral detail in the classification
while adding an additional dimension, either of elevation or
aspect (or both). While one can examine the coincidence of
certain classes with an elevation or aspect range, the assignment
of combinations is based primarily on a priori knowledge. For
instance, red spruce are generally found above 4000 feet (1219
meters). This could be quite useful in separating red spruce and
red pine, which are spectrally quite similar, at least on the July
data. It could also reveal associations which might be worth
investigating. Using this approach it was possible to separate red
pine from red spruce, hardwood 2 on the slopes west of the creek
from that by the pasture and to clearly display the lower
elevations associated with the creekbed. Using aspect and the
basic classification in a model also separated hardwood 2 east and
west of the creek but did not appear to have much beneficial
effect on separating red spruce and red pine or on delineating the
creekbed. This approach will require more knowledge to fully
appreciate its capabilities but it looks like a very powerful tool for
extending acquired information to a larger area.
CONCLUSIONS
The effect of spatial and spectral resolution on the determination
of forest communities was examined by comparing classifications
of digitized aircraft photography at 3 meter spatial resolution and
TM imagery, using the six non-thermal bands and those
equivalent to color infrared photography, at 30 meter resolution.
The aircraft-based classification provided much greater detail
demonstrating a small characteristic distance of change in these
temperate mountainous forests. The larger TM pixel averaged
much of this detail to produce a classification of more uniform
classes and graduated change, which may be less representative of
the surface vegetation but nevertheless useful for studies of larger
areas. The addition of the blue and mid infrared TM bands does
improve the classification when compared with the color infrared-
equivalent bands of TM. The inclusion of elevation and aspect
- adds additional information and capability and may be useful in
creating a more accurate forest delineation. Future work will
examine other areas within the study site and other data source
as well as exploring some of the interesting implications noted
above.
248
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
C5 nn ee p
EE tro c9 7c. + Vj €f)
£0. v0) EN. et
