基于小规模数据集的柑橘树冠层施药情况的分类模型

目的 提升柑橘果园的智能化管理水平，快速无损获取柑橘树冠层的施药情况，改善小规模数据集导致施药情况分类模型易发生过拟合的问题。方法 提出一种基于卷积神经网络的柑橘树冠层施药情况分类模型——VGG_C模型。模型以VGG模型核心思想为基础进行构建，通过交叉熵损失函数优化，加速概率分布与真实分布的迭代过程，并在输出端引入不确定性度量计算以及在下采样模块中插入Droupout方法，降低由于数据较少而发生过拟合的概率。结果 VGG_C模型针对训练集的分类损失值为0.44%，比ResNet和VGG模型分别降低了87%和91%；准确率为95.3%，比ResNet和VGG模型分别提高了5%和10%；验证集的预测平均准确率为96.4%。结论 VGG_C模型通过多层卷积模型协同实现柑橘冠层热红外图像特征的高效提取，通过优化输出端结构提高了柑橘冠层施药情况分类模型在小数据集规模上的训练测试优度，可为柑橘树施药情况的智能化判断提供有效参考。

Classification model of spraying deposition on citrus canopy based on small-scale data set

Objective The study was aimed to improve the intelligent management level of citrus orchards, quickly and non-destructively evaluate the spraying quality on citrus canopy, and solve the overfitting problem of the spraying quality classification model caused by small-scale data set.Method We proposed a classification model of spraying quality on citrus canopy based on convolutional neural network: Visual geometry group citrus model (VGG_C model). The model was constructed based on the core idea of the VGG model. Through optimization of the cross-entropy loss function, the iterative process of probability distribution and true distribution was accelerated. The uncertainty measurement calculation was introduced at the output end and the Droupout method was inserted in the downsampling module to reduce the probability of overfitting due to small amount of data.Result The loss value of VGG_C model for the training set was 0.44%, which was 87% and 91% lower than that of ResNet and VGG respectively. The accuracy of VGG_C model for the training set was 95.3%, which was 5% and 10% higher than that of ResNet and VGG respectively. The average accuracy of the verification set was 96.4%.Conclusion VGG_ C model can effectively extract the features of citrus canopy thermal infrared image through multi-layer convolution model and improve the training and testing superiority of citrus canopy application classification model in small data set by optimizing the output structure. VGG_ C model can provide an effective reference for the intelligent judgment of pesticide application on citrus trees.