Cohen, Yafit
images (Groome, 1998). Supervised grouping of phenological sub-classes into meaningful land covers followed the
Isodata clustering. This process was based on training plots of cultivated fields from ground reference maps supplied by
surveyors from the Israeli Agriculture Ministry and representing 1% of the total fields in the study area. Phenological
features and geographical context consideration such as location and shape were also utilized in this grouping process.
All plots were then compared with the classified image. The comparison process involved classification accuracy
estimation and locating "mixed clusters" i.e. classification clusters, which represent more than one crop or other land
cover. In addition, possible reasons were attached as comment to plots, which classified incorrectly. The coments related
to phenological behavior, plot size, plot texture and location. They were organized in tables and then used for further
learning and defining phenological behaviors as reflected from the NDVI images and their relations with image spatial
considerations. Examination of Isodata classification image shows recognition of wheat, vetch, chickpea, pea, cotton,
sunflower, orchards, high dense shrublands, other natural vegetation, non-vegetated area.
Spectral PCA
Comprehensive examination of "mixed clusters" had shown that phenological attributes are not sufficient for
distinguishing between the orchards and other natural tree formations (NTF). PCA was applied separately on winter
(April) and summer (June) spectral images. Second and third PC's were visually identified as the most discriminative for
this purpose. Utilizing the NDVI format, difference enhancement of the two PCs was calculated following amplitude
enhancement of each PC channel. These values were successfully utilized in order to differentiate between tree
formations.
Average annual rainfall map
Steep climatic gradients exist both North-to-South and West-to-East across the country. These climatic variations cause
high variability in crop seeding and harvesting periods. Climatic index could mask climatic regions and limit specific
phenological features to specific crop. Thus, identical features of two crops could be divided by the climatic mask and
different sub-features of the same crop would be identified and united on the basis of their climatic-dependence
differences. In this experimental recognition system an averaged annual rainfall map (1: 500000) from the Israeli
Meteorological Center was combined as a representing climatic indicator. The analog map was scanned and converted
into a co-registered raster form in a vector-to-raster process.
Soil type map
Soils have a strong influence on land use suitability and crop cultivation and on photosynthetic activity. Therefore, soil
types map might improve recognition quality where for example similar phenological sub- features exist for both crop
and natural vegetation types under similar climatic conditions. In the described experiment the soil contribution was
examined on a small region from the study area, as both soil types heterogeneity in the study area is high and the
connections between cultivated areas and soil types are complex.
2.1.2 *Split-and-merge" rules-base [ ^] ;
generation procedure
Rules-base generation involved an
inductive-learning of existing inter-
relations, among database sources;
relation generalization process, rules
formalization and combined sequence- 4
dependent rules-base ^ construction
(figure 1). Empirical samples chosen
from ground reference maps provided
examples of the concept to be learned.
The learning course involved domain
expert’s expertise to include from
empirical samples some generalized em
relations, that can be used to classify the
remaining data (Huang & Jensen, 1997).
In this study retrieved generalized
relations were used as splitting criteria
of unsupervised classification “mixed-
clusters” in a “split-and-merge” classifi-
cation refinement mechanism. Figure 2 Figure 2. Splitting criteria extraction procedure
describes splitting criteria extraction
stages. Firstly, relationship typology was
Relationships
examination with
land cover type
Sub clusters of
Splitting rule
generalization &
formalization
Distinctive imagery
properties
one 0 Me
relationships
Distinctive
spatial
properties
Merging rule
( Sub cluster ofi or clustémremain mixed formalization
282 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
