Abstract:

In order to rapidly and accurately diagnose and identify nitrogen stress in rice， a field experiment was conducted. Taking the super rice variety ‘Liangyoupei 9’as material， 4 treatments of nitrogen application （0， 210， 300， and 390 kg·hm^-2） were set， and images of the first， second， and third leaves at the top of the rice plant were scanned and collected during the spikelet differentiation and the full heading stage. The SE block （squeeze-and-excitation block） module was added to each residual block of the ResNet34 in convolutional neural network （CNN）， and the weight parameters trained on the ImageNet （ImageNet large scale visual recognition challenge） dataset were transferred to the nitrogen nutrition diagnosis model of rice. The feature extraction layer of ResNet34 was kept unchanged， and the pooling layer at the end of the model was replaced with a global average pooling layer. The improved network was used to extract features from rice images and train the optimal weight file. The results showed that the improved network achieved a testing accuracy of 98.13% during the spikelet differentiation stage and 99.46% during the full heading stage of rice. The convergence speed of the model was faster， and the accuracy was improved by more than 7% compared to the original network. Above results showed that it was feasible to add the SE block to the residual block of ResNet34 and use transfer learning to diagnose nitrogen nutrition in rice， which could effectively diagnose and identify the nitrogen nutrition of rice during the spikelet differentiation stage and the full heading stage. Above results provided reference for the diagnosis and identification of nutrient status in crops.

Key words: convolutional neural network, rice nitrogen nutrition diagnosis, ResNet34, transfer learning, SE block

摘要：

为了快速、准确诊断和识别水稻氮素胁迫程度，对水稻进行大田栽培试验。以超级水稻‘两优培九’为试验对象，设置0、210、300和390 kg·hm^-2共4个施氮水平处理，通过扫描采集幼穗分化期和齐穗期水稻顶1、顶2、顶3叶图像，在卷积神经网络（convolutional neural network，CNN）ResNet34的每个残差块中加入SE block（squeeze-and-excitation block）模块，并将在图像数据集ImageNet（ImageNet large scale visual recognition challenge）上训练得到的权重参数迁移到水稻氮素营养诊断的识别模型中，ResNet34的特征提取层保持原结构，模型结尾的池化层替换为全局平均池化层，利用改进后的网络对水稻图像进行特征提取，训练得到最优的权重参数。结果表明，改进后的网络对水稻幼穗分化期的模型测试准确率达到98.13%，齐穗期的准确率达到99.46%，且模型的收敛速度更快，相比于改进前的网络准确率均提升了7%以上。以上结果表明，通过在ResNet34残差块中加入SE block并基于迁移学习的方法对水稻氮素营养诊断方法是可行的，能有效对水稻幼穗分化期和齐穗期的氮素营养进行诊断识别，为农作物的营养诊断识别提供了参考。

关键词: 卷积神经网络, 水稻氮素营养诊断, ResNet34, 迁移学习, SE block