CAJ | 학술논문

为了提高中国厚皮类瓜果的品质质量和出口能力，增强中国水果品质检测装备制造业的技术实力和技术水平。该文以西瓜为对象，对其糖度进行了试验研究。由于西瓜各部位存在差异，因而不同部位采集近红外光谱会对糖度预测模型精度产生影响。采用自主搭建的西瓜内部品质检测系统对不同批次西瓜瓜梗、瓜脐和赤道3个部位采集漫透射光谱信息，分别采用偏最小二乘回归法（partial least squares regression，PLSR）和最小二乘支持向量机法（least squares support vector machines，LS-SVM）2种方法对西瓜糖度建立预测模型，考察西瓜不同检测部位对西瓜糖度预测模型精度的影响。2种预测模型均显示，赤道部位采集光谱所建立的预测模型检测精度较差，而采用瓜脐部位获取光谱信息建立预测模型略好于瓜梗部位，最佳预测相关系数r pre达到0.823，预测均方根误差（root mean square error of prediction，RMSEP）为0.652%。该研究结果表明，不同部位采集光谱信息对最终的检测模型精度有影响，瓜脐部位为该文西瓜内部品质检测装置的较优采集部位。
위료제고중국후피류과과적품질질량화출구능력，증강중국수과품질검측장비제조업적기술실력화기술수평。해문이서과위대상，대기당도진행료시험연구。유우서과각부위존재차이，인이불동부위채집근홍외광보회대당도예측모형정도산생영향。채용자주탑건적서과내부품질검측계통대불동비차서과과경、과제화적도3개부위채집만투사광보신식，분별채용편최소이승회귀법（partial least squares regression，PLSR）화최소이승지지향량궤법（least squares support vector machines，LS-SVM）2충방법대서과당도건립예측모형，고찰서과불동검측부위대서과당도예측모형정도적영향。2충예측모형균현시，적도부위채집광보소건립적예측모형검측정도교차，이채용과제부위획취광보신식건립예측모형략호우과경부위，최가예측상관계수r pre체도0.823，예측균방근오차（root mean square error of prediction，RMSEP）위0.652%。해연구결과표명，불동부위채집광보신식대최종적검측모형정도유영향，과제부위위해문서과내부품질검측장치적교우채집부위。
Nondestructive detection of the soluble solid content (SSC) is very important to determine the internal quality of watermelon. To enhance the competition of the Chinese equipment manufacturing industry in fruit quality detection, and to improve the benefits of domestic fruit production and processing enterprises, the watermelon, a widely planted thick skinned variety, was selected as the study object. In the study of near-infrared spectra based detective technology of SSC, the spectra collection at different position on the watermelon could result in the variation of precision for the predictive mode by the influencing the spectral signal. In this work, 222 samples were collected at harvest time. The spectra was acquired from the calyx, equator and stem parts of each sample melon. After spectra acquisition, all watermelons were cut into halves from the stem end to the calyx end, and edible portions were removed and cut into proper pieces for obtaining watermelon juice by a juicer. The different spectral data sets were then used as the different inputs of the linear algorithm (partial least squares, PLSR) and nonlinear algorithm (least squares support vector machine, LS-SVM). The 214 samples were retained after getting rid of the abnormal samples;143 and 71 samples were set aside as the calibration set and prediction set, respectively. The predictive abilities of the different model was compared after the SSC calibration models were established. Both the PLSR models and LS-SVM models showed that the models using the spectra collected from the melon equator as input had the worst performance, while the models using the spectra collected at the calyx were the best. For the calyx collected spectra based PLSR model, the correlation coefficient (rpre) was 0.823, and the root mean square error of prediction (RMSEP) was 0.652 percent. The calyx collected spectra based PLSR model was better than the calyx collected spectra based LS-SVM model with a rpre of 0.768 and a RMSEP of 0.731 percent. For stem collected spectra based models, the predictive results were close to the calyx collected spectra based models. It was proposed that the spectra at the calyx part of watermelon should be acquired for our home-built detection system. This work illustrated how the spectra acquired at different parts of watermelons impact the final detection accuracy of the predictive model, but the methods needed to reduce or eliminate this phenomenon require further study.