基于高光谱和卷积神经网络的鲜枣黑斑病检测  

Detection of black rot of fresh jujube fruits using hyperspectral imaging and a convolutional neural network

作　　者：孙海霞[1] 张淑娟[1] 刘蒋龙 陈彩虹 李成吉 邢书海 Sun Haixia;Zhang Shujuan;Liu Jianglong;Chen Caihong;Li Chengji;Xing Shuhai(College of Engineering,Shanxi Agricultural University,Taigu 030801,China)

机构地区：[1]山西农业大学工学院,山西太谷030801

出　　处：《山西农业大学学报（自然科学版）》2018年第11期15-19,26,共6页Journal of Shanxi Agricultural University(Natural Science Edition)

基　　金：国家自然科学基金(31271973);晋中市科技重点研发计划(Y172007-4)。

摘　　要：[目的]为实现鲜枣黑斑特征的识别,提高鲜枣附加价值,采用高光谱成像技术采集了不同年份完好和黑斑鲜枣的信息。[方法]基于全波段光谱,采用偏最小二乘判别分析(Partial Least Squares-Discriminant Analysis,PLS-DA)和误差反向传播神经网络(Back Propagation Neural Networks,BP-NN)建立单一年份和联合年份的判别模型。采用连续投影算法(Successive Projections Algorithm,SPA)提取联合年份的特征波长;利用主成分分析进行单波段图像的数据压缩,针对主成分图像采用BP-NN和卷积神经网络(Convolutional Neural Networks,CNN)进行黑斑鲜枣识别。[结果]联合年份所建PLS-DA、BP-NN模型的整体判别正确率均达到了99.2%,比单一年份所建校正模型的整体判别准确率高,但单一年份所建模型中BP-NN比PLS-DA的判别精度高;采用SPA提取联合年份的特征波长后所建BP-NN判别模型的正确率为100%;基于主成分图像所建BP-NN和CNN模型的判别正确率分别为78.3%和90.0%。[结论]收获年份是影响校正模型稳定性的一个重要因素,联合年份所建校正模型比单一年份所建校正模型具有更好的预测能力;同时CNN可成功应用于基于高光谱技术的鲜枣黑斑特征识别中,也为其它农产品品质检测提供了新方法。[Objective]In order to perform highly effective identification of black rot and increase the additional value of fresh jujube fruits,present study attempted to establish robust identification models using hyperspectra limaging.[Methods]Both healthy jujubes and jujubes with black rot(slight black rot and severe black rot)were collected in different years.A partial least squares-discriminant analysis(PLS-DA)and a back propagation neural networks(BP-NN)were employed to build black rot identification model with the spectral information of the full wavelengths produced by jujubes either from one single year or combined years.The wavelength characteristics of the combined years were selected by successive projections algorithm(SPA),and the data compression of the single band images was carried out by principal component analysis.The black rot was discriminated using the principal component imageof both BP-NN and a CNN.[Results]Compared to the single year model,both calibrated modelsof BP-NN and PLS-DA from combined years produced higher detection accuracy up to 99.2%,while BP-NN had a better accuracy than that of PLS-DA in model from a single year.BP-NN was used to establish the calibration model based on the selected characteristic wavelengths by SPA,and the accuracy reached up to 100%.The classified accuracy of CNN model was 90.0%,which was obviously better than that of BP-NN model at 78.3%.[Conclusion]Results indicated that the harvest year was an important impactor in term of robust model establishment,and CNN was successfully applied to the detection of fresh jujube′s disease using hyperspectral imaging.The study also provides a new quality testmethod forother agricultural products.

关 键 词：鲜枣 黑斑 卷积神经网络 高光谱成像技术 

分 类 号：S123[农业科学—农业基础科学] S665