基于深度卷积神经网络的小麦赤霉病高光谱病症点分类方法

为快速、高效地利用高光谱成像技术诊断小麦赤霉病病症,分析了卷积层结构与光谱病症特征的关联性,并重点研究了高光谱的像元分类建模方法。首先,基于深度卷积神经网络的2种典型结构,构建了不同深度的卷积神经网络,比较了小麦赤霉病高光谱数据点集的训练和测试结果。结果显示:Visual Geometry Group(VGG)结构随着网络深度的增加,模型损失值不断下降;残差神经网络(ResNet)结构随着深度增加,损失值没有明显降低,说明ResNet网络的深度与模型性能无关。从测试集评测模型泛化性可知,具有4个基础单元模块的22层VGG网络在所有深度卷积模型中最优,其建模和验证准确率远高于传统的支持向量机(SVM),分别为0.846和0.843,测试集准确率为0.742。以VGG为基础单元构建的深度神经网络,能有效提取小麦赤霉病病症的高光谱特征。研究结果可为大尺度小麦赤霉病的智能成像诊断提供理论基础。

A classification method for hyperspectral imaging of Fusarium head blight disease symptom based on deep convolutional neural network

In order to realize rapid and early diagnosis of wheat Fusarium head blight disease via hyperspectral imaging, the correlation between the convolutional layer and the spectrum feature of disease symptom was analyzed, and the classification modeling of hyperspectral image was studied. Two typical neural network, Visual Geometry Group (VGG) and residual neural network (ResNet), were introduced to construct the convolutional neural network with different depth. By comparing the training and testing results of the hyperspectral data set for wheat Fusarium head blight disease, it was shown that the loss value decreased with the increased depth of VGG structure, yet the loss value of validation set was not significantly decreased with the increased ResNet depth. According to the evaluation results for testing set, the VGG network of 4 basic units with 22 layers showed the best performance, as its accuracy of training, validation and testing was 0.846, 0843 and 0.742, respectively. Therefore, the VGG network could effectively extract the spectrum feature of Fusarium head blight disease. These results would provide theoretical basis for the intellectual diagnosis of wheat Fusarium head blight disease by the remote sensing in a large scale.