; Des Cesma m e.
s llovski dol, s
ay Opeke . Spaève-
A TN van a * Srmolov CN
Figure7 The largest values of NDVI was noticed
on composite 21-30 July.
4. Annual variations in multispectral feature space
Woods classification by cluster parameters
determination is well known and well used method so
we have chosen another approach. We picked up
sample polygons for observed woods and, set them
spatially fixed, observed variations of their features in
multispectral space during twelve months. For every
polygon we computed mean vector, standard deviation
of mean vector and covariance matrix for all channels.
We expected hysteresis type curve, with turning points
at the beginnings of phenological phases but results
showed completely different annual change [2]. We
interpreted the results following mean vector changes in
first and second channels only. We did that in order to
avoid confusion between reflectance contained in these
channels and brightness temperature from channels 3,
4 and 5. We could suppose that values in 4" and 5"
channels depend on the synoptic scale meteorological
phenomena (by surface temperature) so their
contribution in interpretation of phenological
phenomena is destructive. Channels 1 and 2 cover
visible and near infra red spectrum where vegetation
changes are distinctive. Many relevant works in the
past use these spectral regions and could be used for
comparison. At the end, the limitation of two channels
frees the third axis for time, which was cruical for
phenomena separation during vegetation period. At the
same time NDVI values were high, and their changes
unpredictable, which made the usage of NDVI
variations alone for close-up detection of phenomena
during vegetation period impossible.
754
^ i "4
à MX
- Y
" x * N À
Locations of observed
oak-woods
“ Éesma
a...
^ Opeke Spaëva
ES ur
Figure 8 Low values of NDVI as noticed during
winter 1993 (11-20 January).
Figure 9 shows a sample of mean vector annual
variation for polygon which represents Cesma oak-
wood on cloudless composites. Connecting of mean
vector sequential positions and marking of leafing and
yellowing enable us to easiliy determine the cluster
which represents vegetation period. In all cases similar
behaviour is noticed: entering into vegetation cluster
from the right side, and exiting from the bottom. Before
entering, reflectances in 1% and 2™ channels are
significantly different among observed oak-woods which
is an expected reflection phenomenon for different soils
before leafing. Exiting is more similar for all five oak-
woods because the soil is covered with fallen leaves.
Inside vegetation cluster, in all cases, two distinct
subclusters are noticed. During the first half of
vegetation period mean vector stays at the right side of
vegetation cluster. From late June to early July
distinctive deposition to the left side occurs and it stays
there till the end of vegetation period. In all five wood
samples small overlaps of subclusters occur.
As the subclusters on Fig. 9 are discovered by
mean vector time variations, we call them dynamic
clusters. At this stage we are able to determine
deposition from cluster | to cluster Il only subjectively.
The phenomenon becomes more distinctive when we
introduce time as the third axis (see Fig. 10).
Unfortunately, with available ground data we are unable
to discuss physical meanings of the mean vector
deposition. Table 3 shows mean vector coordinates for
the whole vegetation period and for both subclusters
separately: dynamic cluster for spring to early summer
and dynamic cluster for late summer to autumn.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
