基于反向传播神经网络和高光谱成像的芒果可溶性固形物含量检测

Detection of soluble solids content in mango based on backpropagation algorithm neural network and hyperspectral imaging

中文关键词:  可见光-近红外高光谱成像  芒果  无损检测  可溶性固形物含量  反向传播神经网络

英文关键词:visible-near-infrared hyperspectral imaging  mango  non-destructive testing  soluble solids content  backpropagation algorithm neural network

中文摘要:

      目的 比较反向传播神经网络(backpropagation algorithm neural network, BPNN)模型与偏最小二乘回归(partial least squares regression, PLSR)模型在预测芒果可溶性固形物含量(soluble solids content, SSC)方面的性能。方法 使用高光谱成像仪和全自动折光仪采集芒果的近红外高光谱及SSC数据, 建立两种预测模型,采用多元散射校正(multiplicative scatter correction, MSC)进行光谱预处理, 利用遗传算法(genetic algorithm, GA)、区间变量迭代空间收缩算法(interval variable iterative space shrinkage algorithms, IVISSA)和变量组合群体分析算法(variable combination population analysis, VCPA)提取特征波长变量, 通过比较不同特征波长提取方法进一步优化对比预测模型。结果 与PLSR模型相比, BPNN模型在预测SSC方面更为有效。而在IVISSA特征波长变量提取后优化的BPNN模型预测能力最佳, 预测集判定系数、均方根误差(root mean square error of prediction, RMSEP)、残差预测偏差(residual prediction deviation, RPD)分别为0.8641、0.3924和2.7127。结论 该模型可快速、准确地检测芒果的SSC, 并证明可见光-近红外高光谱成像与反向传播神经网络模型相结合有望预测芒果的SSC, 为开发在线芒果SSC无损检测系统奠定基础。

英文摘要:

      Objective To compare the performance of backpropagation algorithm neural network (BPNN) model and partial least squares regression (PLSR) model in predicting soluble solids content (SSC) of mango. Methods The near-infrared hyperspectral and SSC data of mango were collected using a hyperspectral imager and a fully automated refractometer to establish 2 kinds of prediction models, and the spectral was preprocessed by using multiplicative scatter correction (MSC). Genetic algorithm (GA), interval variable iterative space shrinkage algorithms (IVISSA) and variable combination population analysis (VCPA) were used to extract the characteristic wavelength variables. The comparative prediction model was further optimized by comparing different feature wavelength extraction methods. Results Compared with the PLSR model, the BPNN model was more effective in predicting SSC. The BPNN model optimized after IVISSA feature wavelength variable extraction has the best prediction ability, with prediction set determination coefficients , root mean square error of prediction (RMSEP), residual prediction deviation (RPD) of 0.8641, 0.3924 and 2.7127, respectively. Conclusion The model can detect the SSC of mango quickly and accurately, which demonstrates that the combination of visible-near-infrared hyperspectral imaging and backpropagation algorithm neural network modeling is expected to predict the SSC of mango, which lays the foundation for the development of an on-line mango SSC nondestructive testing system.