耦合平均影响值-连续投影算法优化种子活力近红外检测模型

Optimization of Seed Vigor Near-Infrared Detection by Coupling Mean Impact Value With Successive Projection Algorithm

目前,近红外光谱(NIRS)可以实现种子活力的快速、无损检测,但区分的活力等级一般少于3级且精度不高。建立种子活力多等级、高精度的NIRS检测模型,解决活力等级增加与预测模型精度之间的矛盾是现阶段近红外种子活力检测的主要任务。以玉米种子为研究对象,采用人工老化的方法获得5种活力等级的种子样本并采集对应的光谱数据建立反向神经网络(BP)预测模型。为了提高模型的精度和稳健性,提出一种耦合平均影响值-连续投影特征波长提取算法(MIVopt-SPAsa)。该算法针对连续投影算法(SPA)耗时过长的问题,采用平均影响值算法(MIV)对其预降维。MIV方法实现了对波长影响值的排序,但缺乏选取波长影响阈值的指标,因此引入相对距离比对MIV算法进行优化(MIV),实现特征波长范围的有效分割。针对SPA提取特征变量数目确定的问题,设定了特征波长数目范围并在此范围内优中选优,实现了自适应的SPA(SPA)特征提取。使用耦合MIV-SPA算法对具有1845个波长的玉米种子近红外全谱数据进行特征提取,提取出特征波长37个,主要分布在玉米种子近红外光谱的7个主要吸收峰附近,表明该算法可以有效提取出与玉米种子生化物质近红外吸收特性一致的特征波长。为了测试该算法对模型性能的影响,建立了全谱BP模型、MIV-BP模型、SPA-BP模型、MIV-SPA-BP模型和竞争自适应重加权CARS-BP模型对5个等级的玉米种子活力进行分级,MIV-SPA-BP模型的预测平均准确率可达99.1%,预测精度高于其他模型;其计算平均时间为14.382 s,低于MIV-BP模型的计算时间(24.523 s)、CARS-BP模型的计算时间(97.226 s)和SPA-BP模型的计算时间(101.224 s),但高于全谱模型的平均计算时间(0.2531 s);其最佳表现交叉熵为0.007892,远远低于另外4个模型。实验结果表明:MIV-SPA算法可以有效地提高玉米种子活力近红外检测模型的精度,实现种子活力多等级、精确、无损检测,为种子活力检测模型的优化提供参考。

At present,near-infrared spectroscopy(NIRS)technology,can realize the rapid and non-destructive detection of seed vigor,but the vigor grade is generally less than 3,and the accuracy is not high.The contradiction between the increase of vigor level and model precision urgently needs to be solved in the near-infrared spectrum detection of seed vigor.Five kinds of seed samples were obtained by the artificial aging method,and the corresponding spectral data were collected to establish the BP prediction model.In order to improve the accuracy and robustness of the model,an algorithm of coupled Mean Impact Value-Successive Projection Algorithm(MIV-SPA)is presented.Aiming at the problem of determining the number of feature variables extracted by the Successive Projection Algorithm(SPA),the algorithm sets the number range of feature wavelengths and selects the best in this range to realize adaptive SPA(SPA).Aiming at the problem that SPA algorithm takes a too long time,MIV algorithm is used to reduce the dimension of SPA algorithm.Although the MIV method can sort the wavelength influence values,it lacks the threshold value for selecting wavelength influence.Therefore,the relative distance ratio is introduced to optimize the MIV algorithm to effectively segment the characteristic wavelength range.The full spectrum with 1845 wavelengths is extracted by the MIV-SPAalgorithm,and 37 characteristic wavelengths are extracted,which are mainly distributed near the 7 main absorption peaks of near-infrared spectrum of maize seeds.The results show that the algorithm can effectively extract the characteristic wavelength,which is consistent with the NIR absorption characteristics of maize seed biochemical substances.In order to verify the effect of the algorithm on the performance of the model,the full spectrum BP model,SPA-BP model,MIV-BP model,MIV-SPA-BP model and competitive adaptive reweighting CARS-BP model were established to classify the five grades of maize seed vigor.The average prediction accuracy of the MIV-SPA-BP model is 99.1%,which is higher than other models;the average prediction time is 14.382 s,which is lower than that of the MIV-BP model(24.523),CAR-BP(97.226)and SPA-BP model(101.224 s),but higher than that of full-spectrum model(0.2531);The best performance cross-entropy is 0.007892,which is far lower than other 4 models.The experimental results show that the MIV-SPAalgorithm can effectively improve the accuracy of the near-infrared detection model of maize seed vigor,realize multi-level,accurate and nondestructive detection of seed vigor,and provide a reference for optimizing the optimisation seed vigor detection model.