Abstract Determination of apple maturity in orchards is very important to determine the harvest time and postharvest storage conditions. The aim of this study is to investigate the ability of visible and near-infrared (Vis-NIR) spectroscopy to determine the maturity of bagged ‘Fuji’ apples using the starch index as the maturity index. Using the starch index, 846 apples were divided into three maturity levels (immature, harvest maturity, and eatable maturity). Principal component analysis, the random frog (RF) algorithm, and the RF algorithm combined with the successive projection algorithm (RF-SPA) were used to extract the principal components or characteristic wavelengths of the spectral data. Five machine learning algorithms, namely, the least squares support vector machine (LSSVM), the probabilistic neural network, the extreme learning machine, the partial least squares discrimination analysis, and linear discriminant analysis (LDA), were used to develop a calibration model. By comparing the results of different modeling methods, it was determined that the prediction performance of the RF-SPA-LSSVM model based on 15 characteristic wavelengths was the best. The classification accuracy of the prediction set was 89.05% and the area under the receiver operating characteristic curve of the three types of apples was greater than 0.9210. In addition, four spectral indexes related to the chlorophyll content were used to predict apple maturity. The classification accuracies of the LDA models based on the spectral indexes were 77.63%–80.95%, which were lower than that of the calibration model based on the characteristic wavelength. The results show that the maturity of bagged ‘Fuji’ apples can be accurately and nondestructive determined by Vis-NIR spectroscopy. The selected characteristic wavelengths and spectral indexes can provide a reference for development of a nondestructive device for determination of apple maturity.

中文翻译：

---

用可见光和近红外光谱结合机器学习算法测定袋装‘富士’苹果成熟度

摘要 果园苹果成熟度的测定对于确定采收时间和采后贮藏条件非常重要。本研究的目的是研究可见光和近红外 (Vis-NIR) 光谱使用淀粉指数作为成熟度指数确定袋装“富士”苹果成熟度的能力。使用淀粉指数，将 846 个苹果分为三个成熟度级别（未成熟、收获成熟和可食用成熟）。采用主成分分析、随机青蛙(RF)算法、RF算法结合逐次投影算法(RF-SPA)提取光谱数据的主成分或特征波长。五种机器学习算法，即最小二乘支持向量机（LSSVM）、概率神经网络、极限学习机、偏最小二乘判别分析和线性判别分析 (LDA) 用于开发校准模型。通过比较不同建模方法的结果，确定基于15个特征波长的RF-SPA-LSSVM模型的预测性能最好。预测集的分类准确率为89.05%，三类苹果的受试者工作特征曲线下面积均大于0.9210。此外，四个与叶绿素含量相关的光谱指标用于预测苹果成熟度。基于光谱指标的LDA模型分类精度为77.63%~80.95%，低于基于特征波长的校准模型。结果表明，可见-近红外光谱可以准确无损地测定袋装'富士'苹果的成熟度。所选择的特征波长和光谱指标可为苹果成熟度无损检测装置的研制提供参考。