基于三维重建的秸秆地单株水稻生长形态检测研究  

Detection of single rice early growth morphology based on 3D reconstruction under straw returning condition

作　　者：魏天翔 汪小旵 施印炎 王延鹏 王凤杰 张先洁 卜俊怡 杨四军[2] WEI Tianxiang;WANG Xiaochan;SHI Yinyan;WANG Yanpeng;WANG Fengjie;ZHANG Xianjie;BU Junyi;YANG Sijun(College of Engineering/Engineering Laboratory of Modern Facility Agriculture Technology and Equipment in Jiangsu Province,Nanjing Agricultural University,Nanjing 210031,China;Institute of Agricultural Resources and Environment,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China)

机构地区：[1]南京农业大学工学院/江苏省现代设施农业技术与装备工程实验室,江苏南京210031 [2]江苏省农业科学院农业资源与环境研究所,江苏南京210014

出　　处：《南京农业大学学报》2021年第6期1197-1208,共12页Journal of Nanjing Agricultural University

基　　金：江苏省农业科技自主创新资金项目。

摘　　要：[目的]为评估长江下游稻麦轮作区小麦秸秆全量还田对下茬水稻早期生长的影响,设计1种基于深度学习和三维重建的单株水稻生长形态参数测量方法。[方法]首先,以交并比(IoU)、像素精度(PA)和F值(精准率和召回率的调和平均数)作为精度衡量指标,对比传统图像分割方法(ExG+Otus、Grabcut)和深度学习模型(DeepLabv3+、SegNet、PSPNet)优选出适用于田间水稻植株的图像分割方法。然后,将植株分割图像作为输入,使用多视图空间雕刻算法对水稻进行三维重建,并从模型中提取植株形态参数,与人工测量值对比验证精度。最后,利用筛选出的方法对不同秸秆还田方式下的水稻早期生长发育情况进行量化对比。[结果]DeepLabv3+网络可用于田间水稻图像分割,在水稻分蘖期的不同阶段均能实现较好的分割效果,且能克服田间的杂草、浮萍、倒影等干扰因素,其性能指标分别为IoU值0.801,PA值0.986,F值0.822,优于其他图像分割方法。使用空间雕刻算法结合深度学习图像分割技术可以实现田间水稻的三维模型重建,根据植株三维模型计算得到株高、叶长、分蘖数和叶片数4个形态参数,与人工测量参数对比,决定系数(R2)分别为0.99、0.95、0.89和0.95,均方根误差(RMSE)分别为1.03、1.19、0.82和1.39。利用该方法对“泡田+旋耕(SR)”“旋耕+泡田(RS)”和“翻耕+旋耕+泡田(PRS)”3种不同秸秆还田方式下的水稻早期生态参数进行为期4周的测量,从生长参数的增量对比来看,整体上水稻受秸秆还田的负面影响从小到大依次为PRS、RS、SR。[结论]本文方法可以实现较高精度的田间单株水稻三维重建,可用于秸秆还田后水稻早期生长形态检测研究。[Objectives]In order to evaluate the effect of full wheat straw returning on the early growth of next-stubble rice in the grain rotation area in the lower reaches of the Yangtze River,a measurement method of single-plant scale rice growth morphological parameters based on deep learning and 3D reconstruction was designed.[Methods]Firstly,the methods of image segmentation suitable for rice plants in the field were compared and optimized by comparing the traditional image segmentation methods(ExG+Otus,Grabcut)and deep learning models(DeepLabv3+,SegNet,PSPNet),using intersection over union(IoU),pixel accuracy(PA),F-score as precision indexes,the three-dimensional reconstruction of rice at the single plant scale was carried out by multi-view space carving algorithm with the plant segmentation image as the input.The plant morphological parameters were extracted from the model,and the accuracy was verified by comparing with the manual measurement values.Finally,the quantitative comparison of early growth and development of rice under different straw returning methods was made.[Results]The performance values of DeepLabv3+network in the field rice image segmentation task were IoU value 0.801,PA value 0.986,and F value 0.822 respectively,which were better than the traditional image segmentation method and the previous generation segmentation networks SegNet and PSPNet.It could achieve better segmentation results at different stages of rice tillering stage,and could overcome the interference factors such as weeds,duckweed,and reflections in the field.The spatial carving algorithm combined with deep learning image segmentation technology was used to reconstruct the three-dimensional model of rice in the field.According to the three-dimensional model of the plant,the four morphological parameters of plant height,leaf length,tiller number and leaf number were calculated.Compared with the manual measurement parameters,the determination coefficient R2 was 0.99,0.95,0.89,0.95,and the root mean square error RMSE was 1.03,1.19,0.82

关 键 词：三维重建 深度学习 生长形态 秸秆还田 水稻 

分 类 号：S126[农业科学—农业基础科学]