| 基于无人机多源图谱融合的水稻稻穗表型监测 |
| |
| 引用本文： | 万亮,杜晓月,陈硕博,于丰华,朱姜蓬,许童羽,何 勇,岑海燕.基于无人机多源图谱融合的水稻稻穗表型监测[J].农业工程学报,2022,38(9):162-170. |
| |
| 作者姓名： | 万亮 杜晓月 陈硕博 于丰华 朱姜蓬 许童羽 何 勇 岑海燕 |
| |
| 作者单位： | 1. 浙江大学华南工业技术研究院，浙江大学生物系统工程与食品科学学院，杭州310058; 2. 农业农村部光谱检测重点实验室，杭州310058;;4. 沈阳农业大学信息与电气工程学院，沈阳110866; 5. 辽宁省农业信息化工程技术研究中心，沈阳 110866;;1. 浙江大学华南工业技术研究院，浙江大学生物系统工程与食品科学学院，杭州310058; 2. 农业农村部光谱检测重点实验室，杭州310058; 3. 浙江大学现代光学仪器国家重点实验室，杭州 310027; |
| |
| 基金项目： | 广东省重点领域研发计划项目（2019B020216001）；国家重点研发计划课题（2016YFD0200600，2016YFD0200603）；浙江省重点研发计划项目（2020C02002） |
| |
| 摘 要： | 稻穗表型是表征水稻生长状况和产量品质的关键参数，稻穗表型的准确监测对于大田精准管理和水稻智慧育种具有重要意义。无人机图谱数据已被广泛用于水稻生长监测，然而大部分研究主要集中在水稻的营养生长阶段，针对抽穗期和成熟期稻穗表型监测方面的研究非常有限。因此，该研究利用无人机多源图谱数据进行水稻稻穗表型监测研究，分析了不同氮肥梯度和生长时期对稻穗表型的影响，构建了稻穗覆盖度、生物量以及倒伏等监测模型。结果表明，不同生长时期和氮肥梯度的稻穗表型呈现显著差异，稻穗覆盖度与图像特征高度相关。利用粒子群优化算法（Particle Swarm Optimization，PSO）和支持向量机（Support Vector Machine，SVM）回归模型能够从可见光图像中准确识别稻穗，计算的穗覆盖度与实际标记值高度相关，决定系数（coefficient of determination，R2）为0.87，将此结果与多光谱图像反射率融合，利用随机森林（Random Forest，RF）回归模型可以提高稻穗覆盖度的评估精度，R2为0.93，相对均方根误差（relative Root Mean Square Error，rRMSE）为9.47%。融合可见光图像的颜色和纹理以及多光谱图像的光谱反射率改善了穗生物量的评估精度，R2高达0.84，rRMSE为8.68%，此模型能够在不同种植年间迁移，进一步利用模型更新添加10%新样本能够改善模型迁移能力。基于PSO-SVM分类模型，联合可见光图像的颜色和纹理以及多光谱图像的光谱反射率也准确地识别稻穗倒伏，准确率达99.87%。上述研究结果证明了无人机遥感用于水稻稻穗表型监测的可行性，可为作物精准管理和智慧育种提供决策支持。
|
| 关 键 词： | 无人机 模型 可见光图像 多光谱图像 水稻稻穗 倒伏 机器学习 表型 |
| 收稿时间： | 2022/3/25 0:00:00 |
| 修稿时间： | 2022/4/27 0:00:00 |
Rice panicle phenotyping using UAV-based multi-source spectral image data fusion |
| |
| Wan Liang,Du Xiaoyue,Chen Shuobo,Yu Fenghu,Zhu Jiangpeng,Xu Tongyu,He Yong,Cen Haiyan.Rice panicle phenotyping using UAV-based multi-source spectral image data fusion[J].Transactions of the Chinese Society of Agricultural Engineering,2022,38(9):162-170. |
| |
| Authors: | Wan Liang Du Xiaoyue Chen Shuobo Yu Fenghu Zhu Jiangpeng Xu Tongyu He Yong Cen Haiyan |
| |
| Institution: | 1. Huanan Industrial Technology Research Institute and College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou 310058, China; 2. Key Laboratory of Spectroscopy Sensing, Ministry of Agriculture and Rural Affairs, Hangzhou 310058, China;;4. College of Information and Electrical Engineering, Shenyang Agricultural University, Shenyang 110866, China; 5. Liaoning Agricultural Information Engineering Technology Research Center, Shenyang 110866, China;;1. Huanan Industrial Technology Research Institute and College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou 310058, China; 2. Key Laboratory of Spectroscopy Sensing, Ministry of Agriculture and Rural Affairs, Hangzhou 310058, China; 3. State Key Laboratory of Modern Optical Instrumentation, Zhejiang University, Hangzhou 310027, China; |
| |
| Abstract: | The phenotypic trait of panicle is the key parameter to characterize the growth status, yield, and quality of rice. Accurate phenotyping of panicle is of great significance for field precision management and rice breeding. Unmanned Aerial Vehicle (UAV) image data have been widely used to monitor rice growth status. However, most of studies focused on the vegetative growth stages of rice, and only limited work explored rice panicle phenotyping at the heading and mature stages. Therefore, this study used UAV-based multi-source image data to phenotype rice panicle, analyze the effects of different nitrogen (N) fertilizer levels, growth stages and cultivars on rice panicle phenotyping, and develop the models for monitoring rice panicle coverage, biomass, and lodging. Three field experiments were conducted in Zhuji and Shenyang, China from 2017 to 2018, and a multi-rotor UAV platform equipped with RGB (Red-Green-Blue) and multispectral images was applied to collect rice canopy images. Meanwhile, the ground true values of panicle coverage and lodging were obtained from the RGB images by marking manually object regions, and panicle biomass was measured based on the destructive sampling. The results showed that the panicle phenotyping of rice at different growth stages and N fertilizer treatments was significantly different, and panicle coverage was highly correlated with image features, such as normalized green-red difference index and normalized difference vegetation index. The Support Vector Machine (SVM) combined with the Particle Swarm Optimization (PSO-SVM) accurately identified the panicles from RGB images, and the calculated panicle coverage was highly correlated to the actual marked value with the coefficient of determination (R2) of 0.87. Such a classification model for rice panicles could be applicable to different experimental datasets with the good generalization. Further combination with multispectral reflectance improved the estimation of panicle coverage with the R2 and relatively Root Mean Square Error (rRMSE) of 0.93 and 9.47%, respectively, using the Random Forest (RF) regression model. Fusion of color and texture from RGB images and spectral reflectance from multispectral images improved the estimation of panicle biomass (R2 = 0.84, rRMSE = 8.68%), which outperformed the single RGB or multispectral image data. Further, when the model established from UAV-based multi-source image data in 2017 was used to estimate panicle biomass in 2018, a good estimation result was obtained with the R2 and rRMSE of 0.61 and 15.98%, respectively. Further, the model updating by adding 10% new samples from 2018 to 2017 greatly improved the transferable estimation of panicle biomass, and the R2 and rRMSE were 0.69 and 13.59%, respectively. This indicates that the proposed model for assessing panicle biomass has a high robustness across different planting years. Based on the PSO-SVM classification model, combining color and texture features from RGB images and spectral reflectance from multispectral images accurately identified the panicle lodging with the accuracy of 99.87%. It indicates that UAV-based image features could identify panicle lodging of difference rice cultivars with a close classification accuracy and threshold. The results confirm the feasibility of UAV remote sensing for rice panicle phenotyping, which can provide the decision support for precise crop management and breeding. |
| |
| Keywords: | unmanned aerial vehicle models RGB image multispectral image rice panicle lodging machine learning phenotyping |
|
| 点击此处可从《农业工程学报》浏览原始摘要信息 |
| 点击此处可从《农业工程学报》下载免费的PDF全文 |
|
