MODELING VEGETATIVE DEVELOPMENT OF BERSEEM CLOVER (TRIFOLIUM ALEXANDRINUM L.) AS A FUNCTION OF GROWING DEGREE DAYS USING LINEAR REGRESSION AND NEURAL NETWORKS

Authors: Clapham, William; Fedders, James

Submitted to: Canadian Journal of Plant Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/21/2003
Publication Date: 6/12/2004
Citation: Clapham, W.M., Fedders, J.M. 2004. Modeling vegetative development of berseem clover (trifolium alexandrinum l.) as a function of growing degree days using linear regression and neural networks. Canadian Journal of Plant Science. 84:511-517.

Interpretive Summary: Models that predict growth and development of valuable forages are very useful to producers. They help optimize yield, quality and provide guidance for scheduling harvesting. Models based upon temperature have been used for centuries but have inherent limitations due to assumptions about threshold temperatures for growth. Traditional statistical methods are not appropriate for analysis in many instances due to model complexity In this study we developed a developmental index and predictive model based upon temperature for berseem clover. We demonstrated the impact of these assumptions and demonstrated an alternative method of predicting development using neural networks. Neural networks are a tool developed by the artificial intelligence computer scientists to model complex phenomena, and are routinely used in business, military, and industrial applications. Neural networks provided a more accurate developmental model and improved the power of prediction. The neural network model we developed accurately predicted development of four diverse varieties of berseem clover under controlled conditions.

Technical Abstract: Accurate models of Berseem clover (Trifolium alexandrinum L.) development in relation to growing degree days (GDD) would be useful to both producers and researchers. Predictive ability of linear regression models of plant development may be limited by choice of threshold temperature and the non- linear nature of plant development. Neural networks provide a robust approach to dealing with non-linearity, and may therefore be useful for modeling plant development. In Experiment One, a numerical scale of plant development was created and used to describe growth of four cultivars of berseem clover ('Bigbee', 'Joe Burton', 'Saidi' and 'Tabor') under controlled environmental conditions (constant temperature of 12, 18 or 24- degrees C/12 h photoperiod) for up to 12 weeks of vegetative growth. Simple linear regression and neural networks were used to model plant development in relation to GDD using a range of threshold temperatures. Predictive ability of the models was compared with the results from a second controlled environment study (Experiment Two). The r2 of the linear and neural models produced in Experiment One were maximized at GDD threshold temperatures of 0 to 2-degrees C. Results from Experiment Two indicated that the predictive ability of neural models matched or exceeded that of the linear models for all threshold temperatures evaluated. Results of the current study suggests that neural network models are relatively insensitive to base temperatures across the range tested and may therefore be preferable when a prior knowledge of temperature thresholds is not available.