Cohen, Y., Environmental Information Laboratory, Department of Geography, Bar-Ilan University, Ramat Gan 52900, Israel, Division Geodetic Engineering, Department of Civil Engineering, Technion-Israel Inst. of Technology, Haifa 32000, Israel
Shoshany, M., Environmental Information Laboratory, Department of Geography, Bar-Ilan University, Ramat Gan 52900, Israel, Division Geodetic Engineering, Department of Civil Engineering, Technion-Israel Inst. of Technology, Haifa 32000, Israel

Population growth, urban expansion, land degradation, civil strife and war may place plant natural resources for food and agriculture at risk. Crop and yield monitoring is basic information necessary for wise management of these resources. Satellite remote sensing techniques have proven to be cost-effective in widespread agricultural lands in Africa, America, Europe and Australia. However, they have had limited success in Mediterranean regions that are characterized by a high rate of spatio-temporal ecological heterogeneity and high fragmentation of farming lands. An integrative knowledge-based approach is needed for this purpose, which combines imagery and geographical data within the framework of an intelligent recognition system. This paper describes the development of such a crop recognition methodology and its application to an area that comprises approximately 40% of the cropland in Israel. This area contains eight crop types that represent 70% of Israel agricultural production. Multi-date Landsat TM images representing seasonal vegetation cover variations were converted to normalized difference vegetation index (NDVI) layers. Field boundaries were delineated by merging Landsat data with SPOT-panchromatic images. Crop recognition was then achieved in two-phases, by clustering multi-temporal NDVI layers using unsupervised classification, and then applying 'split-and-merge' rules to these clusters. These rules were formalized through comprehensive learning of relationships between crop types, imagery properties (spectral and NDVI) and auxiliary data including agricultural knowledge, precipitation and soil types. Assessment of the recognition results using ground data from the Israeli Agriculture Ministry indicated an average recognition accuracy exceeding 85% which accounts for both omission and commission errors. The two-phase strategy implemented in this study is apparently successful for heterogeneous regions. This is due to the fact that it allows unsupervised classification to represent the high phenological variability (by utilizing 70 clusters). Utilization of the 'split-and-merge' rules derived from the entire data set of imagery and auxiliary data enabled the formalization of different interpretation contexts for each crop. This technique, which uses imagery information in both stages, is significantly different from exiting methods that are based only on auxiliary geographical and expert knowledge in the post-classification phase. © 2002 Elsevier Science B.V. All rights reserved.