Abstract

Ripe fruit provides essential nutrients for the human body. To fulfill the needs of consumers, the practice of artificial ripening has become more common. Artificial ripening not only degrades the quality of fruit but also impairs health. The potential of hyperspectral imaging coupled with machine learning to quickly and uninvasively identify differently ripened bananas was explored in this study. A total of 300 banana samples that were naturally ripened or ripened by ethephon or calcium carbide were characterized by their hyperspectral images. To improve the accuracy of classification models and to reduce the effects of noise and irregular surfaces, different preprocessing strategies were investigated. Recursive feature elimination (RFE) and the successive projections algorithm (SPA) were employed to select feature wavelengths. Four classification methods – extreme gradient boosting (XGBoost), support vector machine (SVM), multi-layer perceptron (MLP), and partial least square discriminant analysis (PLS-DA) – were applied for ripening identification. The results showed that the best classification model was XGBoost based on full wavelengths, which achieved an area under the macro-average receiver operating characteristic curve of 0.9796 and high training, cross-validation, and testing accuracy values of 95.02%, 92.31%, and 91.16%, respectively. The RFE-XGBoost model reduces the data dimension and maintains satisfactory performance. Both XGBoost models achieved 100% correct differentiation between naturally ripened and artificially ripened bananas. Hence, the method may be employed to quickly and uninvasively classify differently ripened bananas.

摘要

成熟的水果为人体提供必需的营养。为了满足消费者的需求，人工催熟的做法变得更加普遍。人工催熟不仅会降低水果的品质，还会损害健康。本研究探讨了高光谱成像与机器学习相结合快速、无创地识别不同成熟香蕉的潜力。对300个自然催熟或乙烯利或电石催熟的香蕉样品进行了高光谱图像表征。为了提高分类模型的准确性并减少噪声和不规则表面的影响，研究了不同的预处理策略。递归特征消除（RFE）和连续投影算法（SPA）被用来选择特征波长。四种分类方法——极端梯度提升（XGBoost）、支持向量机（SVM）、多层感知器（MLP）和偏最小二乘判别分析（PLS-DA）——被应用于成熟度识别。结果表明，最佳分类模型是基于全波长的XGBoost，其宏观平均接收器操作特征曲线下面积为0.9796，训练、交叉验证和测试准确率分别为95.02%、92.31%和92.31%。分别为 91.16%。RFE-XGBoost 模型降低了数据维度并保持了令人满意的性能。两种 XGBoost 模型都实现了 100% 正确区分自然成熟和人工成熟的香蕉。因此，该方法可用于快速且无创地对不同成熟的香蕉进行分类。