圣女果可溶性固形物(SSC)含量对圣女果内部品质影响至关重要, 但基于高光谱成像及介电性质特征的SSC检测技术存在局限性, 且目前鲜见圣女果SSC无损检测模型。 为实现圣女果SSC的无损检测, 提出基于圣女果可见/近红外光谱特征的SCC预测模型构建, 及改进的BP神经网络算法研究, 以期解决圣女果内部品质的快速无损检测。 以圣女果为研究对象, 试验样本188个, 将其划分为训练集150个和测试集38个, 采用可见/近红外光谱采集系统获取350～1 000 nm范围内的圣女果表面反射强度, 经光谱校正得样本反射率, 为增强信噪比, 截取481.15～800.03 nm范围内的光谱波段作为有效波段进行分析。 通过对比三种预处理模型, 对有效波段进行SG平滑(Savitzky-Golay Smoothing)预处理, 建立BP神经网络预测模型, 测试集决定系数(R2)和均方根误差(RMSE)分别为0.578 5和0.563 9; 在此基础上, 对BP神经网络的网络结构进行改进, 寻求BP神经网络最优预测结构, 计算输出层与期望值间误差, 调整网络结构参数, 将隐含层学习率和神经元个数分别设置为0.01和5, 建立改进的BP神经网络模型(SG-IBP), 测试集R2和RMSE分别为0.981 2和0.102 3; 通过竞争自适应重加权采样算法(CARS)筛选出18个特征波段, 测试集R2和RMSE分别为0.997 8和0.047 9, 同时检测速度显著提升。 研究结果表明: 经过改进的BP神经网络模型性能明显提高, 通过CARS提取特征波段后, 测试集R2提高了0.419 3, RMSE降低了0.516, 检测速度明显提升。 采用CARS提取特征波段的改进BP神经网络模型(SG-CARS-IBP)具有明显的优越性, SG-CARS-IBP模型较为适合圣女果SSC无损检测研究。 该研究可为圣女果SCC的高效无损检测提供参考。

The content of soluble solids (SSC) plays an essential role in the internal quality of cherry tomatoes. However, SSC detection has some problems based on hyperspectral imaging and dielectric properties. There are few SSC non-destructive testing models for cherry tomatoes currently. Therefore, in order to realize the non-destructive detection of SSC in cherry tomatoes, a prediction model of internal quality based on the spectral characteristics of cherry tomatoes and an improved BP neural network algorithm were proposed to solve the problem of rapid non-destructive detection of cherry tomatoes’ internal quality. In this study, cherry tomatoes were selected as the research object, and there were 188 test samples divided into a training set of 150 and a testing set of 38. The cherry tomatoes’ reflective intensity in 350~1 000 nm was obtained using the visible/near-infrared spectral acquisition system, and corrected sample reflectivity was obtained and analyzed. The practical information of the cherry tomatoes’ spectral in 481.15~800.03 nm was intercepted to enhance the signal-to-noise ratio. A BP neural network prediction model was established by comparing the effective wavelengths treated by Savitzky-Golay smoothing (SG). The coefficient of determination (R2) and root mean square error (RMSE) for the test set were 0.578 5 and 0.563 9. On this basis, the network structure of the BP neural network was improved to seek the optimal prediction structure of the BP neural network. The error between the output layer and the expected value was calculated. The network structure parameters were adjusted, and the learning rate and the number of neurons were set to 0.01 and 5 to establish BP neural network model (SG-IBP). The R2 and RMSE of the test set were 0.981 2 and 0.102 3. While the R2 and RMSE of the test set were 0.997 8 and 0.047 9, with 18 feature lengths screened by the competitive adaptive reweighted sampling algorithm (CARS). Meanwhile, the speed was greatly improved. The results showed that the performance of the improved BP neural network model was significantly improved. After feature lengths were extracted by CARS, R2 of the test set was increased by 0.419 3, and RMSE was reduced by 0.516.The speed was also significantly improved. Therefore, the improved BP neural network model, which used CARS to extract characteristic lengths (SG-CARS-IBP), had apparent advantages, and the SG-CARS-IBP model was more suitable for studying cherry tomatoes’ SSC non-destructive detection. This study can provide a reference for efficient non-destructive detection of cherry tomatoes.