糖度(SSC)是苹果内部品质主要评价指标之一, 近红外光谱技术是预测苹果SSC的首选技术, 优化近红外光谱采集装置的参数, 可以提升模型的性能。 采用本课题组自主研发的动态在线设备采集苹果的近红外光谱(350~1 150 nm), 研究不同参数条件下(运动速度、 积分时间和光照强度)对近红外光谱预测苹果糖度模型的影响, 优化动态在线装置的参数。 210个红富士苹果被分为两批, 第一批90个苹果样品, 经过Kennard-Stone算法(K-S)算法分为建模集和预测集, 用于研究不同运动速度、 不同积分时间对苹果SSC含量在线预测模型的影响。 在0.3和0.5 m·s-1两种运动速度下, 使用多元散射校正(MSC)、 小波变换(WT)、 标准正态变量变换(SNV)对采集到的光谱进行预处理, 对不同移动速度的光谱构建糖度的偏最小二乘回归模型(PLS), 结果表明: 装置的运动速度为0.5 m·s-1所建立的预测模型性能较优, 在四种不同积分时间中, 积分时间为120 ms时, 经SNV预处理所建立的模型性能最优, 其预测集的相关系数和均方根误差分别为0.968和0.331。 第二批苹果120个, 经K-S分为建模集和预测集, 选择运动速度为0.5 m·s-1, 积分时间为120 ms的装置参数进行不同光照强度对苹果SSC预测模型影响的研究, 结果发现: 在光照强度为4.5 A时, 采集到的光谱相对其他光照强度组有较大的变化, 光谱在640和800 nm处的波峰基本消失。 在光照强度为6.5 A时, 经SNV预处理后建立的模型性能最优。 再使用竞争性自适应重加权算法(CARS)、 连续投影算法(SPA)对采集的光谱数据进行波长筛选后, 建立苹果SSC模型, 结果表明: CARS-PLS所建立的模型性能较好, 其预测集的相关系数和均方根误差分别为0.991和0.149, 同时简化了模型, 提高了模型的稳定性。 研究表明: 对动态在线设备进行参数优化, 有助于提高苹果模型的预测精度, 该研究有助于对苹果品质在线分选提供技术支持。

Soluble solids content is one of the leading evaluation indicators for internal apple quality. NIR spectroscopy is the first choice for predicting apple soluble solids. Optimizing the parameters of near-infrared spectroscopy collection devices can improve the model’s performance. In this paper, the near-infrared spectrum (350~1 150 nm) of apples was collected by the dynamic online equipment independently developed by our research group, and the effects of different parameters (movement speed, integration time, and light intensity) on the apple quality prediction model by near-infrared spectrum were studied, the parameters of the dynamic online equipment were optimized. The 210 Fuji apples were divided into two batches. The first batch of 90 apple samples was divided into a modeling set and a prediction set by the K-S algorithm, which was used to study the effect of the online prediction model on the solid soluble content of apples with different movement speeds and different integration times. At two moving speeds of 0.3 and 0.5 m·s-1, multiple scattering correction (MSC) and wavelet transform (WT) are used to preprocess the collected spectra, and the SSC model is built for the spectra with different moving speeds. The results show that the prediction model built with amoving speed of 0.5 m·s-1 performs better. Among the four different integration times, the best performance of the model built by SNV preprocessing was achieved at an integration time of 120 ms. The second batch of 120 apples was divided into modeling and prediction sets by the K-S algorithm. The influence of different light intensities on the apple’s SSC prediction model was studied using device parameters with a moving speed of 0.5 m·s-1 and integration time of 120ms. The results showed that when the light intensity was 4.5 A, the collected spectrum changed significantly compared with other light intensity groups, and the peaks at 640 and 800 nm of the spectrum disappeared. When the light intensity is 6.5A, the model after SNV pretreatment has the best performance. Competitive Adaptive Reweighting Algorithm (CARS) and Successive Projections Algorithm (SPA) were used to screen the wavelength of the collected spectral data to establish the apple SSC model. The results show that the model-based on CARS-PLS has good performance and the correlation coefficient and root mean square error of its prediction set are 0.991 and 0.149, respectively. At the same time, the model is simplified, and the stability of the model is improved. The research shows that parameter optimization of dynamic online equipment is helpful in improving the prediction accuracy of the apple model. This research is beneficial in providing technical support for online apple quality sorting.