基于深度学习的大田小区水稻无损测产研究

The yield of a large number of field plots needs to be measured in rice breeding and cultivation. Traditional manual measurement of yield is destructive, tedious, time-consuming, labor intensive and error-prone. Developing yield measuring model based on image analysis has the advantage of non-destructiveness and high-efficiency. Panicle is the organ where grains grow on, so panicle traits are directly related to yield. The ratio of panicle and other organs in rice may be different among different samples. However, the existing models were built on image traits extracted from the whole canopy, which were not capable of characterizing differences in yield between different samples effectively, so these models were generally only fitted to specific varieties. Meanwhile, these models were based on linear regression analysis, thus the accuracy of the models were limited. In this research, we plan to investigate yield measuring method for field plot rice based on deep learning technology. Firstly, we propose a segmentation approach incorporating super-pixel segmentation and super-pixel discrimination based on convolutional neural network. Benefiting from this proposed segmentation approach, panicles of different rice varieties under complex field environment can be segmented accurately and efficiently. Secondly, we introduce a feature extraction method for obtaining yield-related panicle features at two-scale: rice population and single plant. Thirdly, deep belief network is adopted to express the complex relationship between image features and yield, utilizing large amount of unlabelled data and limited labelled data. This research will facilitate the development of high-throughput field rice yield measurement devices, promote rice phenotyping in our country, and support researches related to rice breeding and cultivation.

水稻育种与栽培相关领域研究需要对众多大田小区测产。目前有损测产操作繁琐，耗时耗力，极易在收获、脱粒等过程引入误差。基于图像分析可准确无损测产。现有研究从整个冠层图像中提取特征，构建测产模型，其性能受稻穗在冠层中占比的影响较大，模型适用性差，且基于线性回归分析的建模精度有限。本项目利用深度学习理论，研究适用性强的大田小区水稻无损测产方法。结合超像素分割和卷积神经网络，实现复杂大田环境下多品种稻穗的准确分割，以超像素而非像素作为处理单元，提高处理速度；在此基础上，结合水稻农艺学知识，提取与产量直接相关的群体/单株双尺度稻穗图像特征；引入深度信念网络，充分利用研究中有限数量的有标签样本和大量无标签样本，确定图像特征与产量的数学关系，建立测产模型。该项目的实施将为高通量大田水稻测产平台提供技术储备，对提高我国水稻表型数字化测量水平、支撑水稻育种与栽培相关领域研究具有重要科学意义及潜在应用价值。

水稻育种与栽培相关领域研究需要对众多大田小区测产。目前有损测产操作繁琐，耗时耗力，极易引入较大的人为误差。基于图像分析可准确无损测产。现有研究从整个冠层图像中提取特征，构建测产模型，当水稻样本中稻穗在冠层的占比差别较大时，测量误差较大，模型适用性差。本项目以大田水稻为研究对象，利用深度学习理论，设计了适用于复杂大田环境下的多品种稻穗分割方法Panicle-SEG,该算法通过简单线性迭代聚类方法对图像进行超像素分割，由卷积神经网络判别超像素类别，实现对大田水稻图像的粗分割，然后由基于熵率的超像素方法对粗分割结果进行优化，测试结果表明该方法能克服不同品种水稻稻穗边缘严重不规则、外观表现差异大、穂叶颜色混叠和复杂大田环境中光照、风场、遮挡等因素的干扰，准确分割稻穗。该算法已被打包为exe软件安装包，供用户自由下载。本研究还设计了基于对称编解码结构的PanicleNet算法和基于空洞卷积的PanicleNet v2算法，进一步优化稻穗分割算法。测试结果表明，Panicle-SEG算法的交并比(IoU)为0.588,F值为0.739，PanicleNet的IoU为0.755,F值为0.859，PanicleNet v2的IoU为0.774,F值为0.872。对673个品种的高温实验水稻进行了测产实验，所拍摄图像生育时期涵盖抽穗期至完熟期，产量分布为62.17-444.98g，水稻存在杂草、倒伏、病虫害等现象。在稻穗和冠层准确分割的基础上，提取与产量直接相关的稻穗及冠层图像特征（群体图像特征），每个小区取代表性单株进行考种（单株考种特征），RFECV选取最优特征子集后由SVM构建测产模型，5-fold交叉验证重复100次评估模型性能，结果表明群体图像特征模型R2为0.407，单株考种特征模型R2为0.377，而结合群体图像特征和单株考种特征的模型R2为0.555。本项目还开发了适用于高分辨率大田水稻图像的稻穗检测与计数算法，穗数检测的R2为0.9382。本项目建立的测产模型能适用于不同生育时期、不同品种及不同生长条件，具有较强的适用性，可为高通量大田测产平台提供技术储备，对提高我国水稻表型数字化测量水平、支撑水稻育种与栽培相关领域研究具有重要科学意义及潜在应用价值。