基于高光谱成像技术对番石榴可溶性固形物的快速测定

Rapid determination of soluble solids content in Psidium guava fruit based on hyperspectral imaging technology

中文关键词:  高光谱成像  番石榴  可溶性固形物含量  支持向量回归

英文关键词:hyperspectral imaging  Psidium guava  soluble solids content  support vector regression

中文摘要:

      目的 研究基于高光谱成像技术对番石榴可溶性固形物(soluble solids content, SSC)的快速测定。方法 通过高光谱成像系统和全自动折光仪获得番石榴的表面反射光谱与SSC信息, 选择Savitzky-Golay平滑和标准正态变量变换作为预处理手段, 使用主成分分析评估预处理前后光谱的聚类变化。采用区间变量迭代空间收缩法、区间随机蛙跳法、自举软收缩法(bootstrapping soft shrinkage, BOSS)、变量组合集群分析提取特征波长, 用于建立比较支持向量回归(support vector regression, SVR)和偏最小二乘回归(partial least squares regression, PLSR)预测模型。结果 预处理后光谱数据主成分聚类进一步分散, PLSR整体在预测集的回归效果比SVR更好, BOSS-PLSR为最优预测模型, 预测集决定系数为0.9216, 均方根误差(root mean square error of prediction, RMSEP)为 0.2366, 剩余预测偏差(residual prediction deviation, RPD)为3.5710。结论 利用高光谱成像技术快速实现番石榴可溶性固形物测量是可行的。

英文摘要:

      Objective To study the rapid determination of soluble solids content (SSC) of Psidium guava based on hyperspectral imaging technology. Methods The hyperspectral imaging system and a fully automated refractometer were utilized to acquire surface reflectance spectra and SSC data of Psidium guava. Preprocessing techniques including Savitzky-Golay smoothing and standard normal variable transformation were applied. Principal component analysis was employed to evaluate the clustering variations in the spectra pre- and post-preprocessing. Interval variable iterative space shrinkage algorithms, interval random frog, bootstrapping soft shrinkage (BOSS), and variable combination population analysis were used to extract characteristic wavelengths for building and comparing support vector regression and partial least squares regression prediction models. Results The principal component clustering of the preprocessed spectral data was further dispersed, and the overall model performance of PLSR showed better regression in the prediction set than SVR. The BOSS-PLSR was the optimal prediction model, with prediction set determination coefficients of 0.9216, root mean square error of prediction of 0.2366, and residual prediction deviation (RPD) of 3.5710. Conclusion It is feasible to utilise hyperspectral imaging technology to rapidly achieve Psidium guava soluble solids measurements.