融合高光谱图像技术与MS-3DCNN的小麦种子品种识别模型

小麦品种的纯度和小麦产量密切相关，为了实现小麦种子品种的快速识别，该研究利用高光谱图像技术结合多尺度三维卷积神经网络（Multi-Scale 3D Convolutional Neural Network，MS-3DCNN）提出了一种小麦种子的品种识别模型。首先，利用连续投影算法（Successive Projections Algorithm，SPA）对原始高光谱图像进行波段选择，以减少MS-3DCNN模型的输入图像通道数量，降低网络训练参数规模；其次，利用多尺度三维卷积模块提取特征图的图像特征和不同特征图之间的耦合特征；最后，以6个品种小麦共6 000粒种子的高光谱图像（400～1 000 nm）为研究对象，基于SPA算法选择了22个波段高光谱数据，利用MS-3DCNN、支持向量机（Support Vector Machine，SVM）分别构建了识别模型。试验结果表明，MS-3DCNN模型取得了96.72%的测试集识别准确率，相较于光谱特征SVM识别模型和融合特征SVM识别模型分别提高了15.38%和9.50%。进一步比较了MS-3DCNN与基于二维卷积核、三维卷积核、多尺度二维卷积核构建的多个识别模型性能，结果表明多尺度三维卷积核能提取多种尺度的信息，其识别模型的准确率提高了1.34%～2.70%，可为小麦种子高光谱图像品种识别提供一种可行的技术途径。

Recognizing wheat seed varieties using hyperspectral imaging technology combined with multi-scale 3D convolution neural network

Abstract: A hyperspectral image classification model was proposed to detect wheat seeds using a Multi-Scale 3D Convolution Neural Network (MS-3DCNN) in this study, in order to identify wheat seed varieties quickly and accurately. A multi-scale 3D convolution module was used to learn the characteristics of wheat seed from hyperspectral images. A deep learning model was then established to predict wheat varieties. 3D Convolutional Neural Network (3DCNN) was utilized to simultaneously extract the spatial and spectral dimension features of hyperspectral images, compared with the traditional 2D Convolutional Neural Network (2DCNN). The kernel sizes of convolution were set as 5×5×5, 3×3×3, 5×5×3, and 3×3×5, respectively, considering that the characteristics of spectral dimension occupied a higher position in the application of hyperspectral image data. A Batch Normalization (BN) layer was added after each convolution layer to reduce the over-fitting of the model. The LeaKy_ReLU was adopted in the activation function to prevent neurons from being ineffective when the input was negative. A pooling layer and a fully connected layer were stacked on the last multi-scale convolution module. Finally, the Softmax activation function was used to predict the wheat varieties in the output layer. Dropout was introduced into the fully connected layer to reduce the risk of model overfitting. As such, a total of 6 000 samples were collected for 6 varieties of seeds (1 000 seeds per variety). Specifically, 700 seeds of each variety (4 200 seeds of the 6 varieties) were randomly selected as the training set, and the remaining 1 800 seeds were used as the test set during the specific training. 6 wheat varieties were also selected with certain connections in origin and genetic relationship to evaluate the influence of these factors on the classification model. Nevertheless, there was a relatively large amount of original hyperspectral image data, and a high data redundancy between adjacent hyperspectral bands. Successive Projections (SPA) were selected to combine with the average spectral characteristics of wheat seeds for the less data dimension. Subsequently, 22 optimal bands were selected from 300 bands, where the hyperspectral image data corresponding to the optimal bands was extracted to form a new hyperspectral image space. The reduced dimension data was input into the classification model of MS-3DCNN. The traditional hyperspectral classification model using Support Vector Machine (SVM), 2DCNN, 3DCNN, and Multi-Scale 2D Convolutional Neural Network (MS-2DCNN) were selected to compare the influence of 3D convolution and multi-scale convolution on model. The experimental results showed that the classification performed a higher classification accuracy using the MS-3DCNN model. SVM model using feature fusion, 2DCNN, 3DCNN, and MS-2DCNN models for the test sets achieved the accuracies of 88.33%, 94.17%, 95.17%, and 95.44%, respectively. Particularly, the MS-3DCNN model presented a relatively higher accuracy of 96.72%. Consequently, the improved model can be applied to identify and classify wheat seeds in modern intelligent agriculture.