Comparison of multispectral modeling of physiochemical attributes of greengage: Brix and pH values

Abstract Chemometric modeling concerns both accuracy and computational expense for the prediction of quality-indicating attributes of food materials. Modeling approaches were explored with the hyperspectral images with pH and Brix values of greengages. A two-phase architecture was applied for modeling. Firstly, waveband selection was performed using two approaches, i.e., succession projection algorithm (SPA) and its combination with genetic algorithm (SPA+GA). Secondly, multispectral models based on the two feature sets of wavebands were built via a total of six different modeling methods, i.e., partial least squares regression (PLSR) and extreme learning machine (ELM) in their respective stand-alone versions, their applications combined with genetic algorithm (GA), and their ensemble enhancements with modified Adaboost.RT (MAdaboost.RT). Analysis of accuracy and computational expense showed that supervised feature selection with SPA+GA was superior to unsupervised SPA for better modeling accuracy. MAdaboost.RT-ELM showed high accuracy at low computational expense. ELM models were the better base models than the PLSR ones, for being more randomized and diverse. It indicates that MAdaboost.RT-ELM on SPA is the best choice for a quick test on a newly available dataset, while switching the dimensionality reduction from SPA to SPA+GA may yield more accurate models with added, but well worthy, computational expense.

摘要 化学计量学建模在预测食品材料的品质指示属性时，需同时兼顾建模精度与计算成本。本研究以青梅（greengages）的高光谱图像及其pH值与白利糖度（Brix）值为研究对象，探索了多种建模方法。本次建模采用两阶段架构：第一阶段，采用两种方法开展波段筛选，即连续投影算法（Succession Projection Algorithm, SPA）及其与遗传算法（Genetic Algorithm, GA）的结合形式（SPA+GA）；第二阶段，基于筛选得到的两类波段特征集，共采用6种不同的建模方法构建多光谱模型：包括偏最小二乘回归（Partial Least Squares Regression, PLSR）与极限学习机（Extreme Learning Machine, ELM）的单模型形式、二者分别结合遗传算法的改进形式，以及基于改进型Adaboost.RT（Modified Adaboost.RT, MAdaboost.RT）的集成增强形式。精度与计算成本的分析结果表明，采用SPA+GA的监督式特征选择方法，相较于无监督的SPA方法，能够获得更优的建模精度。改进型Adaboost.RT结合极限学习机的模型（MAdaboost.RT-ELM）在较低计算成本下即可实现高精度预测。相较于偏最小二乘回归模型，极限学习机模型作为基础模型表现更优，因其具备更强的随机性与多样性。综上，基于SPA的MAdaboost.RT-ELM模型是快速测试新数据集的最优选择；而将降维方法从SPA替换为SPA+GA，虽会增加计算成本，但可获得精度更高的模型，且新增的计算成本完全值得。