Improving precision of soluble solid content predictive model by adopting suitable detective position of watermelon based on near infrared spectroscopy
Jie Dengfei,Chen Meng,Xie Lijuan,Rao Xiuqin,Ying Yibin
DOI: https://doi.org/10.3969/j.issn.1002-6819.2014.09.028
2014-01-01
Abstract: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.