为了提高玉露香梨可溶性固形物的检测精度，本研究提出了一种优化反向传播（back propagation，BP）神经网络的玉露香梨SSC预测方法。使用高光谱成像仪采集玉露香梨表面的光谱信息，对剔除异常样本的光谱数据进行不同预处理，以确定最优的预处理方法。采用遗传算法（Genetic Algorithm，GA）优化BP神经网络的初始权重和阈值，建立玉露香梨SSC的GA-BP、BP、PLSR预测模型。结果表明，中值滤波（median filter，MF）预处理后的结果最好。在同一训练样本下，所建GA-BP模型性能最佳，建模集决定系数（Rc2）为0.98，均方根误差（RMSEC）为0.19；预测集决定系数（Rp2）为0.86，均方根误差（RMSEP）为0.43，剩余预测偏差（RPD）为2.45；在此基础上，采用不同数量的样本训练GA-BP网络，样本数为300时，建立的GA-BP模型的Rc2为0.99，RMSEC为0.22；Rp2为0.98，RMSEP为0.20。因此，采用GA-BP神经网络结合高光谱技术可快速、准确的检测玉露香梨可溶性固形物，当训练样本达到一定数量时，可进一步提升模型的预测精度，为基于BP神经网络检测玉露香梨SSC提供了理论基础。

In order to improve the detection accuracy of the soluble solids content of “Yuluxiang” Pear, an optimized back propagation (BP) neural network-based SSC prediction method for “Yuluxiang” Pear was proposed in this study. Spectral information on the surface of “Yuluxiang” Pear was collected by the hyperspectral imager. After the removal of the abnormal samples, different types of pre-processing were performed on the spectral data to determine the optimal pre-processing method. Genetic algorithm (GA) was used to optimize the initial weights and thresholds of the BP neural network, and establish GA-BP, BP and PLSR prediction models for the SSC of “Yuluxiang” Pear. The results showed that the median filter (MF) pre-processing method was the best. For the same training sample, the established GA-BP model performed the best, with the coefficient of determination in the established model (Rc2) being 0.98, RMSEC value being 0.19, the Rp2 value being 0.86, and RMSEP value being 0.43, and residual predictive deviation (RPD) being 2.45. On the basis of these results, different numbers of samples were used to train the GA-BP network. When the number of training samples was 300, the Rc2 value of the established GA-BP model reached to 0.99, with the RMSEC value being 0.22, Rp2 value being 0.98, and RMSEP value being 0.20. Thus, the SSC of “Yuluxiang” Pear can be quickly and accurately detected by GA-BP neural network combined with hyperspectral imaging technology. When the training samples reached a certain number, the prediction accuracy of the model can be further improved. This research provides a theoretical basis for detecting “Yuluxiang”Pear’s SSC based on BP neural network.

山西省应用基础研究计划项目（201901D211359；201801D121252）；山西农业大学科技创新基金项目（2020BQ02）；山西省优秀博士来晋工作奖励资金科研项目（SXYBKY2019049）