P hical 
ncing 
using 
crop 
ed as 
rative 
given 
'ound 
gives 
1996; 
y for 
ssure 
1nual 
isfies 
>toa 
vites 
local 
; and 
The 
f the 
1um- 
ental 
996) 
stem 
99]; 
lysis 
over 
tory 
t of 
(or 
ig to 
stem 
Cohen, Yafit 
  
separates the knowledge required for the recognition from the recognition mechanism (Kontoes et al., 1993) as the its 
algorithm passes over rules according to their confidence values, searching for the most suitable land cover (Lee et al., 
1987). 
This approach have been applied and studied for crop recognition in several researches (Janssen & Middelkoop , 1992; 
Kontoes et al., 1993; Kontoes & Rokos, 1996; Adinarayana & Krishna, 1996). In these studies knowledge-based system 
was utilized for the refinement of conventional image-based classification methods and thus was used in a 
post-classification process. Nevertheless, high spatial-temporal heterogeneity might be responsible for the limited 
utilization of this approach in Mediterranean regions (Shoshany, 2000). This paper examines the potential of the 
incorporation of remotely sensed data, GIS information and expert knowledge for crop recognition applied through 
knowledge-based classification rules in a post-classification manner in a wide heterogeneous Mediterranean region. 
2 Knowledge-Based Crop Classification Method 
Presented in this paper a knowledge-based crop recognition refinement of an unsupervised classification. Wheat, four 
types of legumes, cotton, sunflower and orchards were to identify in this framework presenting 70% of the overall crops 
in Israel. The heterogeneous characteristic of the study area invited two kinds of recognition difficulties. On one hand this 
variability entailed several spectral and phenological sub-features for a single crop type and on the other hand several 
crops had common sub-features resulting in significant confusion. The ability to unite all sub-features of a single crop and 
to overcome confusion between crops demands knowledge about crop spectral/ temporal feature variety and ways 
environmental characteristics and agricultural treatments affect them. 
  
  
  
  
  
  
2.1 Knowledge acquisition and engineering Soil type map Annual rainfall map 
  
  
  
  
Acquisition of these kinds of knowledge was achieved 
through data structuring and a comprehensive learning 
5 : ste Multi-temporal 
course of the inter-relations existing between the data P 
    
    
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
; NDVI images 
sources in relation to different crop types. Data Spatial 
structuring included integration of imagery properties Database 
and digitized soil and rainfall maps (figure 1). The Unsupervised 
relations were studied by domain literature, interviews Ground classification 
with experts and GIS overlay analysis applied on data SNe IDE m 
sources utilizing empirical samples of crop plots chosen Relations Learning & 
from ground reference maps. Engineering of retrieved Agkumml & | 3 Generalization PCA properties 
relations involved rules generalization and es n (Knowledge acquisition) 
formalization and designated for rules base 
construction. Y 
> 
Rules Formalization — Rules sequence Rules 
(knowledge engineering) arrangement Base 
  
  
  
  
  
  
Figure 1. Rule-base construction procedure 
2.1.1 Data structuring 
Second-ordered imagery information of NDVI images, multi-spectral PCA and basic classification layer integrated with 
digitized rainfall and soil maps composed the database. These components together with agricultural knowledge and GIS 
analysis used for extracting the rule-base. 
NDVI maps 
Multi-temporal NDVI maps covering the central and southern parts of Israel were generated from five Landsat TM 
images acquired during 96°-97° growing season in clear sky conditions. À Spot Panchromatic image from June 96’ was 
combined with the NDVI images in two stages. Firstly, the NDVI maps were re-sampled to the Spot spatial resolution 
(10m). Secondly, Spot edge pixels, were enhanced, coded and reduced from NDVI images in order to reduce overall 
image heterogeneity with minimal information loss. The analyzed multi-temporal NDVI images were used as a basis for 
traditional unsupervised classifier and as additional data in the knowledge acquisition process for the rule-based 
classification. 
Isodata unsupervised classified images 
Basic categorization of the multi-temporal NDVI maps was conducted by detailed Isodata (Iterative Self-Organizing Data 
Analysis) unsupervised classification algorithm. In heterogeneous areas the recognition flexibility of the unsupervised 
method makes it an efficient starting point since it enables recognition of sufficient amount of crop sub-features in the 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 281