Cohen, Yafit
Integration of remote sensing, GIS and expert knowledge in national
knowledge-based crop recognition in Mediterranean environment
Yafit COHEN, Maxim SHOSHANY
Bar-Ilan University, Israel
coheny9@ashur.cc.biu.ac.il
KEY WORDS: Agriculture, Remote Sensing, GIS, Integration, Knowledge-based rules.
ABSTRACT
This paper describes a knowledge-based crop recognition system, integrating remote sensing analysis and geographical
information in hierarchical way. Landsat TM and Spot pan images were merged to reduce heterogeneity by enhancing
field boundaries. Multi-temporal NDVI maps generated from these images were classified into eight crop types using
unsupervised classification algorithm. Using agricultural knowledge, the relations between natural vegetation and crop
types, spectral and phenological properties and precipitation and soil types were engineered. These relations were used as
binary rules in an experimental knowledge-based crop recognition system. The binary rules were activated by iterative
"split-and-merge" mechanism of the mixed unsupervised classification clusters aimed at refining the map products given
by the application of unsupervised classification algorithm alone. Experiments in a wide region of Israel using ground
data from the Isracli Agriculture Ministry have shown that the use of knowledge-based “split-and-merge” rules gives
improvement of approximately 9% compared to the unsupervised classification alone.
1 INTRODUCTION
Landuse policy in Israel attracted wide and profound attention (see for example Vitkon, 1991; Shlain & Feitelson, 1996;
Efrat, 1998) due to its severe and unreversible environmental implications and due to its potential effect on life quality for
the years to come. The rapid population and urbanization growth (Shoshany & Goldshlager, 1998) result strong pressure
for urban development on the account of agricultural and open lands. Cultivated areas in Israel have shown a mean annual
reduction of 2% in the last 15 years. Despite limited extent of agricultural lands, Israeli agricultural production satisfies
the local fresh consumption (except for cereals and other crops, which demand wide areas and massive irrigation) due to a
continuous increase in land productivity (Maor Eli, 1999). However, lack of a solid long-term national land-use invites
inadequate land utilization and a continuous decrease of cultivated areas as a result of spontaneous processes and local
economic interests (Niv, 1999).
The realization of national landuse policy aiming at achieving regulatory framework for food, agriculture, open areas and
environmental quality requires adequate sources of information providing the capability of detailed landuse mapping. The
development of a reliable national crop monitoring system is representing principal step of achieving this goal.
Remote sensing techniques have been shown to be cost efficient and useful for crop identification in wide regions of the
world. Remotely sensed image analysis for land cover recognition is commonly applied through standard ML (maximum-
likelihood) classification algorithm (Srinivasan & Richards, 1990). The synergy of remote sensing, environmental
sciences and engineering has emerged from the challenge to improve present monitoring capabilities (Wilkinson, 1996)
and to monitor heterogeneous environments at different scales (Peddle, 1995). The knowledge-based expert system
approach has been increasingly adopted over the years (Matsuyama, 1987; Nicollin & Gabler, 1987; Ton et al., 1991;
Kartikeyan et al., 1995; Pigeon et al., 1999; Soh & Tsatsoulis, 1999) in the view of the need for image analysis
procedures integrating both numerical and logical information (Hinton, 1996).
Image classification knowledge-based systems can be defined as computer programs designated for matching a land cover
type to a pixel in areas with high degree of complexity, which requires wide domain knowledge for acieving satisfactory
recognition. In general, a knowledge-based system is composed of two main elements
(Frost, 1986). Adjustments of these two elements for land cover classification are as follows:
l. A knowledge base: A set of simple facts composed of imagery and environmental data and information and a set of
rules, which describe relations between facts and ways for reliable pixel labeling from them.
2. A problem-solving mechanism: Finding a recognition path from a specific set of facts describing a pixel to one (or
more) land cover through relevant production rules.
These systems enables handling uncertainty, by attaching confidence values to each production rule according to
evaluation of expert knowledge (Alty & Coombs, 1984). By using this evidential approach the knowledge-based system
280 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
aaa —Z———————————>——————— i
