Performance of models following KOA optimization.

Accurate classification of mung bean seeds is essential for enhancing both their nutritional value and crop yields. However, current methods are limited, primarily due to the time-consuming and inaccurate classification process resulting from a lack of diverse dataset features. To overcome these challenges, this study develops a multimodal dataset that integrates Raman spectral features and image-based features through early fusion. Furthermore, the classification of mung bean seed varieties is achieved in a rapid, accurate, and non-destructive manner by optimizing a stacking ensemble learning model using the Kepler Optimization Algorithm (KOA). The multimodal dataset comprises 59 features, selected using the Competitive Adaptive Reweighted Sampling (CARS) method. Specifically, 44 key features are extracted from 700 Raman spectral data points, while 15 key features are derived from 43 image numerical features. The study also used the Kepler Optimization Algorithm to optimize the parameters of various machine learning models, including Decision Tree (DT), Support Vector Machine (SVM), k-Nearest Neighbor (KNN), Backpropagation Neural Network (BPNN), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT). By constructing a stacking ensemble learning model, the research effectively leverages the strengths of multiple classifiers, thereby enhancing the overall classification performance. Experimental results demonstrate that the proposed method significantly improves mung bean seed classification accuracy, with the Kepler-optimized stacking ensemble model achieving an accuracy of 90.71%. This represents a 3.24% improvement over KOA-RF and a 1.59% improvement over KOA-GBDT. In comparison to baseline models, the proposed method proves to be more efficient. This study underscores the potential of combining multimodal features with a Kepler-optimized stacking ensemble learning model for mung bean seed classification. It highlights the role of advanced artificial intelligence techniques in agricultural production and provides valuable technical support for the precise classification of mung bean seeds.精准分类绿豆种子，对于提升其营养价值与作物产量均至关重要。然而当前分类方法存在诸多局限，主要源于数据集特征单一，导致分类过程耗时且精度不足。为克服上述挑战，本研究构建了一套多模态数据集，通过早期融合方式整合拉曼光谱特征与图像特征。此外，本研究借助开普勒优化算法（Kepler Optimization Algorithm, KOA）对堆叠集成学习模型进行优化，实现了绿豆种子品种的快速、精准且无损分类。该多模态数据集共包含59项特征，均通过竞争性自适应重加权采样（Competitive Adaptive Reweighted Sampling, CARS）方法筛选得到。具体而言，从700条拉曼光谱数据中提取44项关键特征，从43项图像数值特征中衍生出15项关键特征。本研究还利用开普勒优化算法，对决策树（Decision Tree, DT）、支持向量机（Support Vector Machine, SVM）、k近邻（k-Nearest Neighbor, KNN）、反向传播神经网络（Backpropagation Neural Network, BPNN）、随机森林（Random Forest, RF）以及梯度提升决策树（Gradient Boosting Decision Tree, GBDT）等多种机器学习模型的参数进行优化。通过构建堆叠集成学习模型，本研究有效融合了多种分类器的优势，进而提升整体分类性能。实验结果表明，所提方法可显著提升绿豆种子分类精度，经开普勒算法优化的堆叠集成模型分类准确率可达90.71%，较KOA优化的随机森林模型提升3.24%，较KOA优化的梯度提升决策树模型提升1.59%。相较于基准模型，本研究所提方法的效率更为优异。本研究证实了将多模态特征与经开普勒算法优化的堆叠集成学习模型相结合用于绿豆种子分类的潜力，凸显了先进人工智能技术在农业生产中的重要作用，可为绿豆种子的精准分类提供宝贵的技术支撑。