西瓜可溶性固形物含量的无损检测对提升其内部品质十分重要。 为实现近红外光谱对小型西瓜表面各部位可溶性固形物含量的准确预测, 减小检测部位差异对预测模型的影响, 　以“京秀”西瓜为研究对象, 分别采集赤道、 瓜脐和瓜梗三部位的漫透射光谱信息, 利用偏最小二乘算法(PLS)建立并比较单一检测部位和混合所有检测部位的西瓜可溶性固形物近红外光谱预测模型, 并分别采用连续投影算法(SPA)和竞争性自适应重加权算法(CARS)对西瓜可溶性固形物近红外光谱变量进行特征波长筛选。 结果显示, 相比于单一检测部位的模型, 混合所有检测部位的校正集样本建立的模型取得了较优的预测结果。 同时, 利用CARS算法筛选的42个特征波长变量建模, 对三种检测部位预测集样本的预测结果分别为赤道RP=0.892和RMSEP= 0.684 °Brix, 瓜脐RP=0.905和RMSEP= 0.629 °Brix, 瓜梗RP=0.899和RMSEP= 0.721 °Brix。 模型得到了很大的简化, 且预测精度较高。 比较发现, 利用SPA算法筛选的19个特征波长变量所建模型的预测精度较低。 利用三种检测部位的西瓜样本建立的PLS混合预测模型, 结合CARS算法进行有效特征波长变量筛选, 可提高西瓜可溶性固形物预测模型的精度, 实现西瓜表面各部位可溶性固形物含量的准确预测, 减小检测部位差异对近红外光谱预测模型的影响。 结果为今后开发便携式设备检测西瓜表面各部位可溶性固形含量提供参考依据。

Non-destructive detection for soluble solids content (SSC) is important to improve watermelon’s internal quality, which attracts more and more attention from consumers. In order to realize the precise detection for SSC of mini watermelon’s whole surface by using Near-infrared (NIR) spectroscopy and reduce the influence of detective position variability on the accuracy of NIR prediction model for SSC, the diffused transmission spectra and soluble solids content were collected from three different detective positions of ‘jingxiu’ watermelon, including the equator, calyx and stem. The prediction models of single detective position and mixed three detective positions for SSC were established with Partial least square (PLS). Successive projections algorithm (SPA) and competitive adaptive reweighted sampling (CARS) were adopted to select effective variables of NIR spectroscopy for SSC of watermelon as well. The results showed that the prediction model of mixed three detective positions was better than the model of single detective position. Meanwhile, 42 characteristic variables of NIR spectroscopy selected with CARS were used to establish PLS prediction model for SSC. The prediction model was simplified significantly and the prediction accuracy for SSC was improved greatly. The correlation coefficient of prediction (RP) and root mean square error of prediction (RMSEP) by CARS-PLS were 0.892, 0.684 °Brix for the equator, 0.905, 0.621 °Brix for the calyx, 0.899, 0.721 °Brix for the stem, respectively. However, the prediction result of SPA-PLS established by 19 characteristic wavelength variables of NIR spectroscopy was bad for the equator, calyx and stem detective positions. The correlation coefficient of prediction (RP) is less than 0.752 and root mean square error of prediction (RMSEP) is relatively high. It was proposed that the PLS prediction model established by mixed three different detective positions with effective characteristic wavelength variables selected by CARS can improve the prediction accuracy for SSC. And the CARS-PLS prediction model can achieve fast and precise detection for SSC of mini watermelon’s whole surface. The influence of detective position variability on the accuracy of NIR prediction model could be reduced simultaneously. Thispaper could provide theoretical basis for calibrating NIR prediction model for SSC of mini watermelon. It also could provide reference for developing the portable and non-destructive detection equipment for soluble solids content of mini watermelon’s whole surface.