04
as
)%
on
lon
Des
118
cal
„of
the
led
for
ble
| in
nat
0S
Veg
o 1.6%
fo 28.3%
fo 0.1%
ho 0.0%
% 0.4%
% 12.2%
b 304%
%
% 27.2%
A am
% 426%
5 3890
nat
bs veg
NEE eem
X 35%
% 0.6%
% 0.1%
% 0.3%
5% 1.3%
% 82.8%
py M5
3% 965
35 3890
best
rubs
and
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
« Better distinction between cultivated areas and ‘other’;
4.3 Convergent belief values
Distributions of CBV of the different classes are presented in
Figure 2. This distribution reflects a hierarchy among these
classes:
1. Crops present dominancy of high CBVs in which summer
crops (cotton and sunflower) exhibit very high proportions of
high belief level (PHBL; ~95%), whereas winter crops (wheat
and legumes) gained only moderate PHBLs (68% and 51%
respectively).
2. Orchards presented a mixture of high, medium and low CBV
figures.
3. Natural vegetation areas and shrubs/forest areas present
dominancy of low and poor levels, with low proportions of
PHBL.
100% à ru pump pU EAS m
14% 10%
of of
2 80% 51% 44% 22%
X 68% 34% 9
a 94% er
€ 0, 0
5 60% - 95% 11%
=
o
E 40% -
o
= 57%
[3]
© 20% 4
œ
0% Ei ; E en T “a ; ATUM ; S ; i i
Qo Q^ > © x xe S
SS S e S « A
3 à «S SS © S
CS SF X 28 oS e I
Epoor (0%-50%) Dlow (50%-65%)
D medium (6596-8096) D high (80956-10096)
Figure 2: Proportions of cumulative belief levels.
In spite of this, as presented in Table 3 orchards, shrubs and
natural vegetation achieved by the KBS accuracy of around
80% and reliability of 95%. These accuracy and reliability
figures leads to a proposition that medium and low CBVs do
not necessarily imply for wrong classification decisions by the
KBS.
S. DISCUSSION
Relationships between recognition accuracy/reliability of both
classification methods and CBVs for each class were assessed.
Two characteristics of a class have attracted attention through
the analysis of these relationships:
Heterogeneity: In similar way to ecological characterization of
species diversity, the CBVs attributed to pixels of a certain
land-cover class, represent the different variants of this class.
Heterogeneity of an ecological system is examined among other
indexes by its species diversity. Analogues to that,
heterogeneity of a class may be examined through its CBV
diversity. CBV diversity of a certain class was measured
according to Shannon-Weiner information index:
20
CBV Diversity (CBVD) - *. p; *In p; (1)
i=}
where / stands for the 5% intervals of the CBV (e.g., i=1 is 0%-
5% CBV and i=20 is 95%-100%) and p stands for the
proportion of each interval relative to the overall class.
919
CBVD PHBL US-CEM KBS-CEM
Cotton 0.75 95% 94% 95%
Sunflower 1.17 94% 91% 92%
Wheat 2.00 68% 78% 88%
Legume 2.53 51% 67% 88%
Orchards 2.19 37% 65% 77%
Shrubs 2.42 14% 29% 81%
Nat_Veg 2.76 1096 30% 83%
Table 4: Values of CBVD, PHBL, US-CEM and KBS-CEM
for each class.
deba boar ly = 0.085% + 0.0005
= 8 R? = 0.5862
TT} 8
Q » =
> = =
o
5 475 © m
=
E y = -0.2974x + 1.2361 ° o
c R? = 0.6915
9
= 05
o
= i
o o US-CEM |
a = KBS-CEM doe
0.25
0.5 1 1.5 2 2.5 3
Heterogeniety Index - CBVD
Figure 3: Relationships between heterogeneity index
(CBVD) and classification efficiency measures (CEM) of US
and KBS classifications.
A high CBVD is expected when there is a wide range of CBV
values attributed to a class, which indicates heterogeneity and,
conversely, a low CBVD is expected for cases in which there is
dominancy of a certain signature. As inferred from Table 4,
summer crops are very homogeneous, whereas wheat, orchards,
legumes and shrubs are more heterogeneous. It may be
hypothesized that as class heterogeneity increases, the
recognition ability of a classification decreases. This hypothesis
is partially supported by the classification results. When the
classification efficiency (CEM) is regarded as a measure
representing the lower value between accuracy and reliability of
each class, there was found moderate correlations (1? = 0.69)
between CBVD and CEM for the US and lower for the KBS (1
— 0.59; Figure 3). However, the US classification efficiency is
highly more affected by the heterogeneity. Slope of linear
trend-line of the US is five times higher then this of the KBS
(0.3 vs 0.065). These moderate correlations indicate that
heterogeneity alone does not fully characterize the limitedness
of the US classification, and there is a need to analyze how
unique is each class. :
Uniqueness: is represented by the PHBL obtained for each
class (Table 4). Wherever a class is composed solely of unique
variants it gains a relatively high PHBL (e.g., cotton) as there
are negligible conflicts in most of its pixels. High correlation (r^
— 0.94) was found between the CEM of the US and the PHBL
(Figure 4). In addition to the moderate correlation found with
the CBVD it can be concluded that the success of an ‘off-the-
shelf US classification diminishes with increasing
heterogeneity of a class, and to a greater extent than its
diminution with decreasing uniqueness.
In contrast, the CEM of the KBS presented moderate
correlation with PHBL (Figures 4), and with five times lower