红提的糖度和硬度是评价红提品质的重要指标,探究了基于高光谱成像技术的红提糖度和硬度的无损检测方法及最佳预测模型。在红提果粒的三种放置模式(横放、果柄侧朝下、果柄侧朝上)下,分别采集213个样本在400~1000 nm波长范围内的高光谱图像,对比分析光谱采集的最优模式;然后在最优采集模式下对光谱进行预处理;应用遗传算法(GA)、连续投影算法(SPA)、竞争性自适应重加权(CARS)算法和无信息变量消除法(UVE)针对原始光谱提取特征波长;结合化学计量学方法分别建立基于全光谱和特征波长的偏最小二乘回归(PLSR)、最小二乘支持向量机(LSSVM)和随机森林(RF)的红提糖度、硬度的无损预测模型。结果表明:基于RF建立的糖度和硬度模型的效果较优;预测糖度的最优模型为遗传算法优化的随机森林(GA-RF),其校正集相关系数(Rc)、预测集相关系数(Rp)分别为0.969、0.928,校正集均方根误差(RMSEC)、预测集均方根误差(RMSEP)分别为0.266、0.254;预测硬度的最优模型为基于移动窗口平滑结合连续投影算法优化的随机森林(MA-SPA-RF),其Rc、Rp分别为0.961、0.932,RMSEC、RMSEP分别2.119、1.634。研究结果表明基于高光谱成像技术预测红提的糖度和硬度是可行的。

The sugar content and firmness of red globe grapes are important indicators for evaluating their quality. This study explores nondestructive detection methods and best prediction models for determining the sugar content and firmness of red globe grapes based on hyperspectral imaging technology. The hyperspectral images of 213 samples, in the wavelength range of 400-1000 nm, are collected in three placement orientations (horizontal, fruit stalk-side down, and fruit stalk-side up). The optimal orientation for spectral imaging is compared and analyzed, and subsequently the spectrum is preprocessed in the optimal orientation. Several preprocessing methods, i.e., genetic algorithm (GA), successive projections algorithm (SPA), competitive adaptive reweighed sampling (CARS) algorithm, and uninformative variable elimination algorithm (UVE), are applied to the images to extract characteristic wavelengths from the original spectra. Using chemometrics methods, combined with either partial least squares regression (PLSR), least squares support vector machine (LSSVM), and random forest (RF) analysis based on full spectra and characteristic wavelengths, several protocols are established to mathematically predict the sugar content and firmness of red globe grapes from the images. Results show that the sugar and firmness model based on RF performs the best. The optimal model for predicting sugar content proves to be RF optimized by GA (GA-RF), with corrected-set correlation coefficient (R_c) and predicted-set correlation coefficient (R_p) values of 0.969 and 0.928, respectively, and corrected-set root-mean-square error (RMSEC) and predicted-set root-mean-square error (RMSEP) values of 0.266 and 0.254, respectively. The optimal model for predicting firmness proves to be RF optimized by moving-average method and SPA (MA-SPA-RF), with R_c and R_p values of 0.961 and 0.932, respectively, and RMSEC and RMSEP values of 2.119 and 1.634, respectively. These results prove the sugar content and firmness of red globe grapes can be nondestructively predicted via hyperspectral imaging.