基于机器学习和深度学习的玉米种子活力光谱检测

Spectral detection of maize seed vigor based on machine learning and deep learning

为解决传统的种子活力检测方法存在耗时长、损伤种子等问题,实现种子活力的快速无损检测,分别利用机器学习和深度学习算法结合高光谱成像技术构建玉米种子3个活力梯度分类模型,通过人工老化方式将1 012粒玉米种子分为3个活力梯度样本,采集其高光谱数据后通过卷积平滑(SG)和多元散射校正(MSC)去除高光谱噪声,分别采用主成分分析(PCA)、连续投影算法(SPA)进行光谱特征降维,再从降维后的波段中抽取1 156、1 191和1 463 nm 3个波段合成假彩色图像,用局部二值模式(LBP)提取感兴趣区域的纹理特征,并与纯光谱特征融合。分别基于纯光谱特征构建决策树(DT)和支持向量机(SVM)模型和融合特征建立随机森林(RF)、SVM和极端梯度提升树(XGBoost)模型等机器学习模型。将假彩色图像输入ResNet18、MobileNetV2、DenseNet121、Efficientb0、Efficientb2等5个深度学习模型中进行玉米种子活力预测。结果显示,就机器学习方法而言,针对纯光谱特征表现最好的是PCA-SVM模型,其测试集准确率为92.5%;针对融合特征表现最好的是SVM模型,其测试集的分类准确率为93.1%;就深度学习方法而言,轻量化的MobileNet取得最高的测试集分类准确率99.5%;基于可解释的梯度定位类别激活映射方法表明,分类网络会重点关注玉米种子的中部或基部区域。

A three-vigor gradient classification model for maize seeds was constructed using machine learning and deep learning algorithms along with hyperspectral imaging technology to solve the problems of time-consuming and seed damage in the traditional method for detecting seed vigor and to realize the rapid,non-destructive detection of maize seed vigor.1 012 maize seeds were divided into three vigor gradient samples by artificial aging.The hyperspectral noise was removed with convolution smoothing(SG) and multivariate scattering correction(MSC) after collecting the hyperspectral data of maize seeds.Principal component analysis(PCA) and continuous projection algorithm(CPA) were used for dimensionality reduction of spectral feature,respectively.Three bands including 1 156 nm,1 191 nm,and 1 463 nm were extracted from the reduced dimensionality band to synthesize a false color image.The texture features of region of interest(ROI) were extracted using local binary mode(LBP) and fused with pure spectral features.Machine learning models including decision tree(DT) and support vector machine(SVM) models constructed based on pure spectral features and the random forest(RF),SVM and extreme gradient lifting tree(XGBoost) models constructed based on fused features were established.Maize seed vigor was predicted by inputting the false color images into five deep learning models including ResNet18,MobileNetV2,DenseNet121,Efficientb0,and Efficientb2.The results showed that the PCA-SVM model performed best for pure spectral features,with a test set accuracy of 92.5% in terms of machine learning methods.The SVM model performed best for fusion features,with a test set accuracy of 93.1%.In terms of deep learning methods,the lightweight MobileNet achieved the highest test set accuracy of 99.5%.The classification activation mapping method based on interpretable gradient indicated that the classification network focused on the bottom or basal region of maize seeds.It will provide some references for the nondestructive detection of maize seed vigor in terms of data sources,deep neural network visual interpretation and machine learning,and deep learning performance analysis.