Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.
Identifying Risk Conditions for FireBlight Infection using Artificial Neural Networks based on rare events
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.orgCitation: Computers in Agriculture and Natural Resources, 4th World Congress Conference, Proceedings of the 24-26 July 2006 (Orlando, Florida USA) Publication Date 24 July 2006 701P0606.(doi:10.13031/2013.21892)
Authors: Robert Moskovitch, Dima Stopel, Ilaria Pertot, Cesare Gessler
Keywords: Artificial Intelligence, Artificial Neural Networks, Clustering, Erwinia amylovora
Conditions for the infection of apple and pear by Erwinia amylovora, the causal agent of FireBlight, were investigated and modeled by plant pathologists in different places in the world. Examples include: MARYBLYT (Steiner & Lightner, 1996), BIS95 (Billing, 1996), FBCA (Shtienberg et al, 1998). Interestingly, the outputs of these models, which are targeted to predict disease risk, even based on same data, are different. Recently, fire-blight appeared on a few infected trees in Trentino, Italy. An attempt to identify dates and sites with high disease infection probability, based on known models developed in other places (Maryblyt, Bis95, FBCA) resulted in low estimation accuracy of the true events. Evidently, the models do not sufficiently consider the particularity of the region nor can they consider other factors such as the inoculum's distribution. Therefore an innovative approach was attempted to answer the challenge of extracting a model based on rare examples of infected trees and the region meteorological data, to identify the vulnerable areas within specific periods, in which Fire-Blight infection can occur in Trentino. We used artificial neural networks (Bishop, 1995) in order to cluster the given data including meteorological data, geographical properties and infected trees over two periods of time: blooming and prior-infection period. The infected stations at both periods concentrated in the 20-30% of the clusters, from which we extracted the features values in which the infection may occur. We found that all the temperature measures were relatively low, for infected clusters, and that the height was high. In the blooming period leaf-wetness and degree hours were relatively low.(Download PDF) (Export to EndNotes)