高光谱结合离散二进制粒子群算法对久保桃可溶性固形物含量的检测

Determination of Soluble Solid Content in Peach Based on Hyperspectral Combination With BPSO

可溶性固形物(SSC)是评价久保桃内部品质的重要指标。传统的SSC检测有损、费时、费力;快速、无损检测久保桃的SSC含量对于其品质分级有着重要意义。离散二进制粒子群算法(BPSO)是在标准粒子群算法(PSO)的基础上,更新速度公式得到的,具有精度高,收敛快的特点,多用于离散空间的优化问题。基于高光谱技术结合BPSO算法及BPSO的组合特征波长选择算法对久保桃的SSC含量预测进行研究。首先采集198个久保桃样本的高光谱信息,获取久保桃900~1700nm范围内的光谱信息,计算感兴趣区域的平均光谱作为有效光谱数据,同时测量久保桃的SSC值。采用K-S(Kennard-Stone)算法将样本划分为校正集(147个)和预测集(51个)。使用BPSO特征波长选择算法对久保桃的原始光谱数据进行特征波长提取,并与竞争性自适应重加权算法(CARS)、连续投影法(SPA)、无信息变量选择法(UVE)等特征波长选择算法比较。同时为了避免单一算法建模中的不稳定问题,提出了基于BPSO的一次组合(BPS0+CARS、BPSO+SPA、BPSO+UVE)和二次组合[(BPSO+CARS)-SPA]、[(BPSO+SPA)-SPA]、[(BPSO+UVE)-SPA]特征波长提取方法。基于上述10种特征波长提取方法分别建立支持向量机(LS-SVM)模型和遗传算法(GA)优化的支持向量机模型(GA-SVM)模型。结果表明,基于BPSO算法提取特征波长建立的模型预测性能均高于其他单一特征波长方法,建立的两种模型预测集决定系数R_(p)^(2)均达到0.97以上;基于BPSO的组合算法中,二次组合(BPSO+SPA)-SPA算法建立的LS-SVM在特征波长数量较少的情况下对久保桃SSC含量预测性能最高,校正集和预测集决定系数R_(c)^(2)为0.982,R_(p)^(2)为0.955,均方根误差RMSEC为0.108,RMSEP为0.139。该模型预测性能略低于BPSO算法,但其仅用了22个特征波长进行建模,极大地简化了模型。说明(BPSO+SPA)-SPA是一种有效的特征波长提取方法,为水果SSC含量的无损检测提供了新的检测方法。

Soluble solids(SSC)are an important index to evaluate the internal quality of Kubo peach.Traditional SSC content detection is destructive,time-consuming and laborious.Rapid and nondestructive detection of the SSC content of Kubo peach is of great importance for its quality classification.Binary particle swarm optimization(BPSO)is obtained by updating the speed formula based on standard particle swarm optimization(PSO).BPSO has the characteristics of high accuracy and fast convergence and is mostly used in optimization problems in separate spaces.Taking Kubo peach as the research object.Basedon hyperspectral technology combined with BPSO and based on BPSO combined characteristic wavelength selection algorithm to study the SSC content of Kubo peach.Firstly,hyperspectral information of 198 Kubo peaches was collected to obtain the spectral curve of Kubo peaches in the range of 900~1700 nm.Meanwhile,theSSC value of Kubo peaches was.Used(Kennard-stone)algorithm to divide samples into a correction set(147)and a prediction set(51).The BPSO feature wavelength selection algorithm is used to extract the feature wavelength from Kubo s original spectral data.It is compared with the Competitive Adaptive Reweighting algorithm(CARS),Successive projections algorithm(SPA),and Uninformative variable selection algorithm(UVE).A method of extracting characteristic wavelength based on BPSO is proposed for primary combination(BPS0+CARS,BPSO+SPA,BPSO+UVE)and secondary combination((BPSO+CARS)-SPA),(BPSO+SPA)-SPA),(BPSO+UVE)-SPA).Based on the10 characteristic wavelength extraction methods above.Established support vector machine(LS-SVM)model and the genetic algorithm(GA)optimized support vector machine(GA-SVM)model of Kubo peach SSC content.The results show that the prediction performance of the model based on the BPSO algorithm is higher than that of other single characteristic wavelength algorithm,and the coefficient of determination R_(p)^(2) of the prediction set of the two models is above 0.97.Among the combination algorithms based on BPSO,the LS-SVM based on the quadratic combination(BPSO+SPA)-SPA algorithm has the highest prediction performance for Kubo peach SSC content when the number of characteristic wavelengths is small.The coefficient of determination between the correction set and the prediction set are 0.982 and 0.955,respectively.The root mean square errors RMSEC and RMSEP were 0.108 and 0.139,respectively.The prediction performance of the proposed model is slightly lower than that of the BPSO algorithm,but only 22 characteristic wavelengths are used for modeling,which greatly simplifies the model.These results show that(BPSO+SPA)-SPA is an effective method for extracting characteristic wavelength,which provides a new method for nondestructive detection of fruit SSC content.