结合高光谱成像和机器学习的棉种年份鉴别

Combine Hyperspectral Imaging and Machine Learning to Identify the Age of Cotton Seeds

棉花精量播种技术目前已经在新疆兵团全面推广,该技术能精确实现一穴一粒的农艺技术指标,但是也对高质量棉种的筛选提出了更高的要求。为了避免播种往年活力不足的棉种而导致发芽率降低的问题,结合机器学习和近红外(NIR)高光谱成像技术(HSI)进行棉种年份精确鉴别,实现棉种的快速无损筛选。采集2016年—2019年近四年外观无明显差异的棉种各360粒,共1440粒棉种(按照3∶1∶1划分训练集、验证集和测试集)作为样本,按照每批60粒采集915~1698 nm范围的棉种高光谱图像,去除首尾两端噪声大的光谱,保留1002~1602 nm范围的光谱为原始数据。利用Savitzky-Golay(SG)平滑算法对光谱进行预处理,采用主成分载荷方法(PCA-loading)选取13个特征波段,基于全部光谱数据和特征波段(±10 nm)数据建立逻辑回归(LR)、偏最小二乘判别分析(PLS-DA)、支持向量机(SVM)、循环神经网络(RNN)、长短记忆网络(LSTM)和卷积神经网络(CNN)六种分类模型。使用全光谱数据建模时,六种分类模型在测试集上的鉴别准确率分别为96.27%,98.98%,99.32%,96.95%,97.63%和100%,其中CNN和SVM模型取得了较好的结果;使用特征光谱数据建模时,六种分类模型在测试集上的鉴别精度分别为93.56%,97.29%,98.30%,95.25%,94.24%和99.66%,其中CNN和SVM模型仍有较好的分类结果。结果表明,使用全光谱数据建模时,六种分类模型都可以实现较高精度的棉种年份鉴别,使用特征光谱数据建模时CNN和SVM模型的鉴别精度仍可达到98%;其中深度学习方法优于传统机器学习方法,但是传统机器学习方法仍能保持较好的鉴别准确率。因此,结合近红外高光谱成像技术和机器学习方法能够实现棉种年份的高精度鉴别,为棉花精量播种过程中的优质棉种选种技术提供理论依据和方法。

At present,the technology of precision cotton seeding has been promoted comprehensively in Xinjiang Corps,which can accurately achieve the agronomic technical standards of one grain per hole,but it also sets higher demands for the screening of high-quality cotton seeds.To avoid the decrease of germination rate caused by the cotton seeds with lack of vitality in previous years,machine learning and near-infrared(NIR)hyperspectral imaging(HSI)technology can be used to identify cotton seed years with high precision and to screen cotton seeds quickly and nondestructively.A total of 1440 cotton seeds with no difference in appearance were collected in 2016,2017,2018,and 2019,and 360 seeds per year(According to 3∶1∶1,it is divided into the training set,validation set,and test set.)as samples.Hyperspectral images of cotton seeds in the range of 915~1698 nm were collected according to each batch of 60 seeds,and average spectra(1002~1602 nm)for removing obvious noise at the beginning and the end were extracted as the raw data.SavitzkyGolay(SG)smoothing algorithm was used to preprocess the spectra.The principal component analysis loading(PCA-loading)method was used to select 13 effective wavelengths.Six classification models,including logistic regression(LR),partial least squares discriminant analysis(PLS-DA),support vector machine(SVM),recurrent neural network(RNN),long-short memory network(LSTM),and convolution neural network(CNN),were established based on full spectra and effective wavelengths.When using full spectra to build models,the identification accuracy of the six classification models on the test set was 96.27%,98.98%,99.32%,96.95%,97.63%,and 100%,respectively,among which CNN and SVM models had achieved good results.When using effective wavelengths to build models,the identification accuracy of the six classification models on the test set was 93.56%,97.29%,98.30%,95.25%,94.24%,and 99.66%,respectively,among which CNN and SVM models still had excellent classification results.The results showed that the six classification models could achieve high precision cotton seed years identification when the full spectra were used,and the identification accuracy of CNN and SVM models was still up to 98%when the effective wavelengths were used.The deep learning methods are generally better than the traditional machine learning methods,but traditional machine learning methods can still maintain good identification accuracy.Therefore,the combination of near-infrared hyperspectral imaging technology and machine learning methods can achieve high-precision identification of cotton seed years.It provides theories foundation and methods for selecting high-quality cotton seeds in the process of precision sowing.