基于Stacking集成卷积神经网络的水稻氮素营养诊断

【目的】为实现水稻氮素营养状况的快速、准确诊断，提出了基于集成卷积神经网络的水稻氮素营养诊断模型，为建立高性能的氮素营养诊断模型提供思路和方法。【方法】水稻田间试验以超级杂交水稻‘两优培九’为材料，设置4个施氮水平(0、210、300、390 kg/hm²)。扫描获取水稻幼穗分化期顶部3片完全展开叶的叶片图像，将图像裁剪至只包含叶尖片段的图像，进行水稻叶片图像数据采集。分别以单一卷积神经网络模型DenseNet121、ResNet50、InceptionResNet V2为基学习器，多层感知机(MLP)为元学习器，集成卷积神经网络模型，比较了集成模型与单一卷积神经网络模型以及不同基学习器组成的集成模型的氮素营养诊断结果。【结果】4个单一模型中，DenseNet121的氮素诊断准确率最高，为96.41%。二元集成模型和三元集成模型的准确率均高于任意一个单一模型的准确率，由3个基学习器组成的集成模型的准确率最高，达到98.10%，相比准确率最高的单一模型准确率提高了1.69个百分点。【结论】采用DenseNet、ResNet50、InceptionResNet V2集...

Rice nitrogen nutrition diagnosis based on stacking integrated convolutional neural network

  【Objectives】  To achieve rapid and accurate diagnosis of rice nitrogen nutrition status, we established a rice nitrogen nutrition diagnosis model involving stacking integrated convolution neural networks.  【Methods】  In a rice field experiment, a super hybrid rice cultivar “Liangyoupeijiu” was used as the test material, and four N application levels (0, 210, 300, 390 kg/hm2) were the treatments. At the young panicle differentiation stage of rice, the images of the top three fully unfolded leaves were taken by scanning, and the images were cut into images containing leaf tip parts, and the rice leaf image dataset was established after preprocessing. A stacking integrated convolutional neural network model with different four combinations of three base learners (i.e., DenseNet121, ResNet50, InceptionResNet V2) and MLP as the meta learner was constructed. The results of the integrated models on nitrogen nutrition diagnosis task were compared with that of the single convolutional neural network model of different single base learners (i.e., DenseNet121, ResNet50, InceptionResNet V2, and VGG16).  【Results】  Among the four single models, DenseNet121 had the highest accuracy of 96.41%. The accuracy rate of the binary integration model and the ternary integration model were higher than that of the single model. The accuracy rate of the stacking integration model was the highest, reaching 98.10%, with an increase of 1.69 percentage points compared with the single model which had the highest accuracy.  【Conclusions】  The nitrogen nutrition diagnosis model established by stacking integrated convolution neural network has strong generalization ability and learning ability, and can accurately identify nitrogen nutrition status.