Scholarlink

<p>The experimental population of this work was concerned with the classification of seeds of five proso millet (<i>Panicum miliaceum</i> L.) genotypes (KY1, KY2, KY3, KY4, DS20), which were produced by single-plant selection using Turkish landraces. It was classified based on the morphometric and color features, which were obtained by using a scanner-based image-processing system, and the nutritional parameters obtained by near-infrared (NIR) spectroscopy, namely their suitability to human consumption. An equal image dataset was assembled, which had 500 seeds in each genotype. This was achieved by scanning the seeds at a resolution of 600 dpi of five technical replicates where each replicate had 100 seeds. To perform the nutritional analysis, 100 spectral measurements of an approximate amount of 300 g of ground sample of each genotype were measured using a calibration model specifically based on cereals, and this was repeated five times. In the classification task, there were ten different machine learning algorithms (J48, kNN, LMT, MLP, NB, REPTree, RF, RT, SL, SMO) and two data mining algorithms (CHAID, CART) carefully tested with ten-fold stratified cross validation. The findings clearly showed that the random forest (RF) algorithm showed the best and the most stable performances in all metrics true positive rate [TPR] = 0.950, <i>F</i>-measure = 0.950, Matthews correlation coefficient [MCC] = 0.938, receiver operating characteristic area under curve [ROC AUC] = 0.996, and precision-recall area under curve [PRC AUC] = 0.985. RF model was next followed by the CART and REPTree algorithms, whereas the kNN algorithm had relatively lower classification success. All these results support the notion that morphometric signatures and decision based on NIR-based compositional data have some synergy in proso millet that can be utilized to offer an applicable, repeatable, and cost-effective decision framework to address critical processes including the control of seed purity, quick verification of variety, and quality classification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Classification of Proso Millet (Panicum miliaceum L.) Seeds Based on Physical and Nutritional Approaches Through Machine Learning and Data Mining Algorithms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月18日

查看全文
  • Mevlüt Akçura,
  • İrem Poyraz

摘要

The experimental population of this work was concerned with the classification of seeds of five proso millet (Panicum miliaceum L.) genotypes (KY1, KY2, KY3, KY4, DS20), which were produced by single-plant selection using Turkish landraces. It was classified based on the morphometric and color features, which were obtained by using a scanner-based image-processing system, and the nutritional parameters obtained by near-infrared (NIR) spectroscopy, namely their suitability to human consumption. An equal image dataset was assembled, which had 500 seeds in each genotype. This was achieved by scanning the seeds at a resolution of 600 dpi of five technical replicates where each replicate had 100 seeds. To perform the nutritional analysis, 100 spectral measurements of an approximate amount of 300 g of ground sample of each genotype were measured using a calibration model specifically based on cereals, and this was repeated five times. In the classification task, there were ten different machine learning algorithms (J48, kNN, LMT, MLP, NB, REPTree, RF, RT, SL, SMO) and two data mining algorithms (CHAID, CART) carefully tested with ten-fold stratified cross validation. The findings clearly showed that the random forest (RF) algorithm showed the best and the most stable performances in all metrics true positive rate [TPR] = 0.950, F-measure = 0.950, Matthews correlation coefficient [MCC] = 0.938, receiver operating characteristic area under curve [ROC AUC] = 0.996, and precision-recall area under curve [PRC AUC] = 0.985. RF model was next followed by the CART and REPTree algorithms, whereas the kNN algorithm had relatively lower classification success. All these results support the notion that morphometric signatures and decision based on NIR-based compositional data have some synergy in proso millet that can be utilized to offer an applicable, repeatable, and cost-effective decision framework to address critical processes including the control of seed purity, quick verification of variety, and quality classification.