The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
309
As Table 2 and Table 3 (see chapter 2) already show, most of
the changes indicated by the per-pixel indicators are changes in
vegetation, i.e. either an increase or decrease of the vegetations
vitality. Therefore, a change in vegetation is given, if a segment,
which has been segmented on the basis of the Diffnorm-
channels, shows discrepancies in the NDVIs calculated for tO
and tl. By calculating the ratio between the mean NDVI tl and
mean NDVI t0 of an object no change is given if the ratio is ex
actly at 1.0. In terms of indicating a gradual change the more
the ratio between the mean NDVI tl and mean NDVI t0 of an
object is unequal to 1.0 the more a change in vegetation is given.
This expression can be depicted by a fuzzy membership func
tion as displayed in Figure 5, whereas here each NDVI has been
normalized to a range of 0.0 to 1.0 before and as upper and
lower bound for absolute membership (i.e. a definite change)
0.9 and 1.1 respectively were set.
Figure 5: Fuzzy membership function to express changes in the
vitality of vegetation by the ratio of the NDVIs of tO an tl.
In terms of expressing gradual changes, the membership func
tion in Figure 5 can be interpreted as follows: if the ratio be
tween NDVI t i and NDVI t0 of an object is exactly 1.0 no change
in vegetation is given. If it is below 0.9 or above 1.1 there is
definitely a change in vegetation given. If the ratio is between
1.0 and 1.1 or 0.9 and 1.0 a gradual change of vegetation ac
cording to the degree of membership (p) is given. This way, an
increase or decrease in vegetation vitality can be described
analogous so that a definite decrease is given with a ratio below
0.9 and vice versa for an increase. However, as demonstrated in
Figure 6, increases or decreases of the NDVI are given in agri
cultural areas the same way as for example changes from vege
tated to non-vegetated areas and vice versa. As shown in Figure
6 due to the relatively high dynamic of agricultural areas in the
NDVI (or any other measurers sensitive for vegetation vitality),
increases or decreases of the NDVI mostly occur in such areas.
However, in the most cases these increases or decrease cannot
be interpreted as changes of land use. On the other side, changes
of vegetation in agricultural areas typically occur steadily and
more or less equally distributed within a field (as field in this
context areas with equal cultivation and not of equal land tenure
are meant). Thus, in many cases such changes can be discrimi
nated from other changes by regarding the homogeneity of
change within a change area. To describe the homogeneity of a
change related to its area, e.g. the differences of the standard
deviations (Aa) for each channel at tO and tl within the change
area can be ana lyzed. However, a disadvantage of this approach
is to find suitable thresholds to determine whether a change is
homogeneous or not, since the standard deviation depends on
the size of a segment. Thus, in order to evaluate the usability of
Aa we determined the thresholds of -20 respectively +20 em
pirically in conjunction with the membership-functions as dis
played in Tab. 4. Besides, these values approximately coincide
with the overall a/2 of the per object Aa. When working with
object hierarchies, as like with Definiens™ Developer, another
approach to define the homogeneity or heterogeneity of change
within an area is to analyze the relationship of the size of a
change area to the number of sub-segments generated by an
over-segmentation as described in chapter 3.2. For this ap
proach an indication can be given by the ratio of sub-segments
to the number of pixels within the segment itself: assuming a
segment containing k pixels and n sub-segments whereas the
segmentation to generate the sub-segments aggregates pixels, so
that for the segment itself k > n > I is true. Then by dividing n
by k a maximum of homogeneity is given if n = 1 and k > 1.
Since n / k converges to 0 if n f k f 1, a minimum of homoge
neity is given if n = k and n > I and k > 1. Therefore, a linear
membership function can be defined which expresses the degree
of homogeneity in change respectively, (see Tab. 4). An ap
proach, that it capable to combine an arbitrary number of prop
erties to determine a degree of change is given by a modifica
tion of a procedure which has been used by Earth Satellite Cor
poration, named Cross-Correlation-Analysis (Koeln et al., 2000).
In this method, class boundaries from the older thematic map
separate image pixels into distinct class zones. Within these
boundaries the pixels as of a new unsupervised classification are
validated using a multivariate z-statistic. This idea has been
adopted and slightly adjusted within this project using the De-
COVERtO mapping as reference (super-objects within the
object-hierarchy) and computing the Z-Value for each sub
object within the boundaries of a tO mapping object:
with:
v i(i = mean of property i of the sub-object at point of time tl.
/u i o = mean of property i of the super-object at point of time tO.
<j i/o = standard deviation of property i of the super-object at
point of time tO.
Figure 6: Classified change segments by decrease or increase of
the NDVI and DeCOVER tO mapping superimposed to
Diffnorm-channels (R=Diff_norm nir, G=Diff_norm red,
B=Diff_norm green.)