多尺度特征融合的柑橘冠层施药沉积量分类模型  

Classification model for citrus canopy spraying deposition based on multi-scale feature fusion

作　　者：陆健强 林佳翰 邓小玲 兰玉彬 邱洪斌 杨瑞帆 陈平福 Lu Jianqiang;Lin Jiahan;Deng Xiaoling;Lan Yubin;Qiu Hongbin;Yang Ruifan;Chen Pingfu(College of Electronic Engineering,College of Artificial Intelligence,South China Agricultural University,Guangzhou 510642,China;National Precision Agriculture International Joint Research Center of Aerial Application Technology,Guangzhou 510642,China;Guangdong Laboratory for Lingnan Modern Agriculture,Guangzhou 510642,China)

机构地区：[1]华南农业大学电子工程学院、人工智能学院,广州510642 [2]国家精准农业航空施药技术国际联合研究中心,广州510642 [3]岭南现代农业广东实验室,广州510642

出　　处：《农业工程学报》2020年第23期70-76,共7页Transactions of the Chinese Society of Agricultural Engineering

基　　金：广东省重点领域研发计划(2019B020214003);国家自然科学资金(61675003);广东省普通高校“人工智能”重点领域专项(2019KZDZX1001)。

摘　　要：针对传统农作物冠层施药沉积量分类模型分类准确率低、网络模型参数量大且运算速度慢的问题,该研究提出一种改进的SPP-Net-Inception-v4模型。该模型通过构建稀疏网络结构平衡各个模型子网间的计算量,利用3个Inception模块生成施药沉积量在柑橘冠层热红外图像的稠密有效特征数据;在模型的卷积层与全连接层间创新性接入空间金字塔池化网络(Spatial Pyramid Pooling Network,SPP-Net),进行一次历遍提取热红外图像特征信息,再通过空间池化操作融合3种池化方式提取的多尺度特征,实现柑橘冠层热红外图像施药沉积量表现特征的提取与融合。搭建多环境因素自主控制试验环境,模拟无人机低空采集柑橘冠层热红外图像,应用3个分类模型进行对比试验,试验结果表明,SPP-Net-Inception-v4模型与Inception-v4和ResNet-152两种模型相比,准确率分别提高1.58%和3.26%,模型训练完成冻结后占用计算机存储空间大小分别降低13%和24%,表明SPP-Net-Inception-v4模型在降低模型规模的基础上,提高了柑橘树冠层施药沉积量分类的准确率,可为精准农业航空中无人机植保技术的进一步发展提供参考。There are relatively few studies on the spray quality of citrus tree canopy in China.In most cases,the method is that farmers observe the spray quality of citrus tree canopy up close in the orchard,which wastes manpower and material resources.Moreover,the observation effect is not ideal and may cause harm to human safety.Domestic conditions allow some orchards to judge the spray effect by observing the water-sensitive paper.This method does save a lot of manpower,and efficiency is significantly improved compared with the previous ones.However,this method is affected by the external environment,and the naked eyes cannot correctly determine the spray quality,which will cause certain errors.Therefore,this study explored new ways to solve these problems.In recent years,thermal imaging technology had shown great research promise in some emerging research areas,especially in agricultural production.In the field of precision agriculture,for example,high-resolution thermal imaging cameras,with the aid of advanced aerial photography technology,can quickly capture thermal images of the canopy before and after spray,which provides new ideas for precision agriculture spray detection technology.This study combined thermal imaging technology and computer vision technology to identify and classify the spray conditions of the plant canopy and accurately detected the spray quality of the leaves,which avoided the waste of pesticides and reduced the manual re-examination steps.Thereby reducing agricultural production costs and improving the economic benefits of agricultural products.In this study,the canopy of a citrus tree was used as a thermal image acquisition area.Inevitably,there are problems such as high noise,low contrast,and blurred feature information in the thermal image acquisition process.In response to the above problems,this study preprocessed the acquired thermal images and eliminated unreasonable data,and set the original data set to two labels(sprayed and unsprayed),and divided them into the training set and the te

关 键 词：农药 模型 喷雾 柑橘 热红外图像 多尺度特征 SPP-Net 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]