Application of supervised feature selection methods to define the most important traits affecting maximum kernel water content in maize

Abstract

This study presents the results of applying supervised feature selection algorithms in the selection of the most important traits contributing to the maximum kernel water content (MKWC) as a major yield component. Data were obtained from a field experiment conducted during 2008 growing season, at the Experimental Farm of the College of Agriculture, Shiraz University, and from the literature. Experiments on the subject of sink/source relationships in maize were collected from twelve fields (as records) of different parts of the world, differing in 23 characteristics (features). The feature selection algorithm demonstrated that 15 features including: planting date (days), countries (Iran, Argentina, India, USA, Canada), hybrid types, Phosphorous fertilizer applied (kg ha-1), final kernel weight (mg), soil type, season duration (days), days to silking, leaf dry weight (g plant-1), mean kernel weight (mg), cob dry weight (g plant-1), kernel number per ear, grain yield (g m-2), nitrogen applied (kg ha-1), and duration of the grain filling period (0C day) were the most effective traits in determining maximum kernel water content. Among the effective traits (features), planting date (days) revealed to be the critical one. Hybrids and countries were the second most important affecting factors on the maize kernel water content. For the first time, our results showed that features classification by supervised feature selection algorithms can provide a comprehensive view on distinguishing the important traits which contribute to maize kernel water content and yield. This study opened a new vista in maize physiology using feature selection and data mining methods and would be beneficial to newcomers of this field