雷竹笋硬度的近红外光谱检测模型优化

Phyllostachys violascens model optimization for bamboo shoot firmness using near-infrared spectroscopy

为了提高应用近红外光谱技术无损检测雷竹Phyllostachys violascens竹笋硬度的精度,研究了雷竹笋硬度光谱检测模型的优化方法。首先对雷竹笋原始光谱进行正态变量变换(SNV),然后采用后向间偏最小二乘法(bi PLS)去除部分与竹笋硬度无关的变量,随后进一步采用竞争性自适应权重法(CARS)剔除无关变量,最后采用连续投影算法(SPA)将光谱变量个数从1 557个减少为25个。最终,bi PLS-CARS-SPA模型的交叉验证相关系数(rcv),预测相关系数(rp),交叉验证均方误差(RMSECV)以及预测均方误差(RMSEP)分别为0.984,0.926,0.300 N·cm-2和0.625 N·cm-2,优于其他几种常见的变量选择方法及其组合。研究结果表明,bi PLS-CARS-SPA方法所选特征变量避开了水分强吸收峰的影响,具有实际的物理表征意义,为竹笋硬度在线快速检测、筛选和指导切削设备的研发提供了重要的理论依据。

To develop a calibration model for rapid, accurate and nondestructive determination of bamboo shoots firmness with Phyllostachys violascens by using near infrared spectroscopy（NIRS） technology. The diffuse reflectance spectra of bamboo shoot were obtained in the wavelength range from 800 to 2 632 nm. Different informative variable selection methods were first calculated with the full spectra being pretreated using a standard normal variate（SNV） transformation. Analyses with backward interval partial least squares（bi PLS）, synergy interval partial least squares（si PLS）, genetic algorithm（GA）, successive projections algorithm（SPA）, Monte Carlo uninformative variable elimination（MCUVE）, and competitive adaptive reweighed sampling（CARS）were compared. Then CARS and SPA were used on the spectrum to select wavelengths in proper order. The performance of the models were tested using a correlation coefficient for cross-validation of calibration（rcv）,root mean square error for cross-validation of calibration（RMSECV）, the correlation coefficient of prediction（rp）, and the root mean square error of prediction（RMSEP）. Results showed that Bi PLS combined with CARS and SPA obtained a total of 25 wavelengths or only 1.6% of the full wavelengths. The rcv, RMSECV, rp, RMSEP by bi PLS-CARS-SPA were 0.984, 0.300 N·cm^-2, 0.926, 0.625 N·cm^-2, respectively. The good performance demonstrated that NIR spectroscopy coupled with the bi PLS-CARS-SPA algorithm could be used successfully to analyze bamboo shoot firmness and revealed that the bi PLS-CARS-SPA algorithm was superior to other wavelength selection methods.