|Zhou, Marvellous -|
|Kimbeng, Collins -|
|Gravois, Kenneth -|
|Bischoff, Keith -|
Submitted to: Crop Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: October 1, 2010
Publication Date: January 3, 2011
Citation: Zhou, M.M., Kimbeng, C.A., Tew, T.L., Gravois, K.A., Bischoff, K. 2011. Artificial neural network models: A decision support tool for enhancing seedling selection in sugarcane. Crop Science. 51:21-31. Interpretive Summary: Early-stage selection in sugarcane, even when based on whole-family performance in addition to individual performance, is inadequate and visual selection among individual seedlings is subjective and inefficient. The objectives of this study were to identify statistical methods to more effectively evaluate family yield potential and to enhance seedling selection for yield and yield-related characteristics. The statistical models that were used identified the best families with higher yield potential and higher likelihood to exhibit greater potential at later selection stages. In our experiment, superior families identified with the aid of enhanced statistical procedures comprised a larger proportion of seedlings that proved to be high yielding at later selection stages than was obtained through visual selection alone. Statistical models provided an objective decision support tool for selecting high yielding seedlings and were more flexible at adjusting the number of seedlings to advance than would have occurred though visual selection alone, and thus increase the likelihood of retaining potentially high yielding varieties that might have otherwise been dropped from the program.
Technical Abstract: Currently, sugarcane selection begins at the seedling stage with visual selection for cane yield and other yield-related traits. Although subjective and inefficient, visual selection remains the primary method for selection. Visual selection is inefficient because of the confounding effect of genotype by environment interaction and competition among closely spaced seedlings. Artificial Neural Network (ANN) models are mathematical models based on biological neural networks; they are a supervised learning method and use pattern learning from training data to produce predictions of response variables. In sugarcane, ANN models would use yield components, namely, stalk number, height, and diameter as predictor variables to produce a selection decision. The objective of this study was to examine the potential of using the SAS enterprise miner ANN models to identify seedlings with higher yield potential. Seedling data for stalk numbers, height, diameter, and the decision to either select or reject a seedling were collected from five crosses grown at the Louisiana State University Agricultural Center, Sugar Research Station, and 17 crosses grown at the United States Department of Agriculture, Sugar Cane Research Unit, in Louisiana, USA. The ANN models produced greater differences between the seedling cane yields of the selected and rejected seedlings than visual selection. At the same selection rate, seedlings selected with neural network models produced higher yields than those advanced by visual selection. The ANN models selected fewer seedlings with lower cane yield and rejected fewer seedlings with higher cane yield than the visual selection.