International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
=» | unsupervised classification layer based on the NDVI
layers;
» | averaged annual rainfall data layer from the [sraeli
Meteorological Service;
= | soil types layer from the GIS of the Ministry of
Agriculture of Israel; and
= 1 land use layer from the Israeli National GIS.
Sensor Image date
10-Nov-96; 14-Feb-97; 19-April-
Landsat TM 97: 21-May-97; 22-Jun-97
Spot-
panchromati 20-Jun-96
C
Table 1. Images available to study area.
3.22 Hierarchic representation
The GSA makes it possible to use evidence, which may apply
not only to a single hypothesis (e.g., sunflower), but also to sets
of hypotheses (e.g., sunflower, cotton), that together comprise a
concept of interest (e.g., summer crops). A specific KBS
hierarchic representation should relate to semantic affinity
between classes, and to indicative information which can be
obtained from the database sources. Figure 1 displays the
hierarchic representation of crop types and their generalized
super-classes. It can be inferred from the tree that there are only
9 final classes (underlined): other (non-vegetated formations),
mixed natural vegetation, shrubs/forests, citrus, wheat, legume,
other crops, cotton, and sunflower. Each relates to different
number of super-classes.
All
p bm
Other Vegetation
|
[5 |
Cultivated Natural vegetation
Crops Citrus
peni
Summer crops Winter crops
1 1
= 1 T r 1
Sunflower Cotton Other Legume Wheat
Forest/Shrubs Mixed
Figure 2: Hierarchic tree representation of land-cover/use
and crop types.
3.2.3 Rule base Formation
In general terms, a rule here represents the support value m
given to a hypothesis A, assuming that indicators X, Y, Z are
valid:
if X and Y and Z and... then A with m
Rule base formation involves learning the relationships between
potential indicators and object classes (potential hypotheses).
The learning process was conducted through analysis of domain
literature and interviews with experts from the Ministry of
Agriculture. The results of this process were used to learn
growth rates of crop types, optimal environmental conditions
for crop growth in the various climatic areas of Israel, and the
effects of environmental modifications on crop growth rate and
quality. In addition, field survey plots were used to learn how
growth rates and quality are reflected in imagery data. This was
achieved by both visual interpretation and GIS analysis. Rules
were related to all classes from all levels. Indications of various
kinds and with various affinities (support values) were found,
and selection was applied in order to exclude indications with
poor affinity. In terms of support values, only indications with
more than 50% support were included. The resulting rule-based
composition demonstrates the priority given to imagery data, as
90% of the rules included imagery indicators. In addition, 20%
of the rules utilized soil type properties, 20% used precipitation
properties, and 13% used INGIS land-use information.
4 RESULTS
The KBS generates two outputs for each pixel: its recognition
class and its CBV. The present section will describes
classification results and the CBV distribution separately.
4.1 Classification results
Assessment of the confusion matrix for an US classification is
most important, since it indicates the locations of phenological
conflicts between crop types and thus facilitates assessment of
the resolved and unresolved confusion introduced by using the
KBS. Application of the US ISODATA classification yielded
good results for four crop categories and very poor results for
orchards, shrubs and mixed natural vegetation categories (Table
2). These results demonstrate the high information content in
the NDVI phenologies (Cohen and Shoshany, 2002).
Reference sun: nd
dass) wheat legume cotton flower orchards shrubs vl
wheat 77.8% 8.9% 0.5% 3.4% 1.6%
legume 4.0% 72.2% 4.7% 6.0% 28.3%
cotton 99.7% 5.6% 0.6% 1.1% 0.1%
sunflower 91.3% 0.3% 0.6% 0.0%
orchards 1.7% 65.0% 32.0% 0.4%
shrubs 3.5% 28.9% 12.2%
nat. veg 7.1% 5.5% 1.7% 3.6% 30.4%
other crop 3.1% 0.9% 1.8%
other 9.3% 13.3% 0.3% 22.8% 226% 2%
Reliability 87.1% 67.4% 94.0% 98.8% 55.7% 81.4% 42.6%
No. of pixels 9529 7058 9315 7653 10897 7085 3890
Table 2: Confusion Matrix of US Classification.
Reference» sun nat
Classi wheat legume cotton flower orchards shrubs veg
wheat 89.9% 7.9% * 3.9% 0.3% 3.5%
legume 64% 91.3% 2.5% 02% 06%
cotton 99.7% 5.6% 0.9% 0.2% 0.1%
sunflower 91.3% 0.6%
orchards 3.296 77.0% 05% 03
shrubs 1.5% 80.6% 1,3%
nat. veg 0.3% 14% 820%
other crop 3.1% 1.4%
other 25% 08% 0.3% 12.0% 167% ons
Reliability 88.0% 87.5% 94.6% 99.1% 96.0% 96.3% 96.1
No. of pixels 9529 7058 9315 7653 10897 7085 3890
Table 3: Confusion Matrix of KBS Classification.
The recognition achieved by applying the GSA is best
characterized by the following principle cases:
= Considerably better recognition of legume, orchards, shrubs
and natural vegetation;
= Considerably better distinction between winter crops and
natural vegetation;
1» Better distinction between orchards areas and shrubs;
918
Interna
= Bef
42 C
Distrib
Figure
classes
1. Cro
crops (
high be
and le;
respect
2. Orch
figures.
3. Nat
domina
PHBL.
1
Relative amount of pixels
NS
[n spite
natural
80% a;
figures
not nec
KBS.
Relatio
classifi
Two cl
the ana
Hetero
species
land-co
Hetero;
indexes
heterog
diversit
accordi
where
396 CI
proport
