华中农业大学工学院光机电共性技术与装备团队与湖北省农业科学院麦类育种团队联合在Computers and Electronics in Agriculture发表题为“Dynamic change and rating of wheat and barley waterlogging based on multimodal fusion and analysis of key phenotypic features”的研究论文。该研究成果为小麦和大麦渍害等级的快速、客观评价提供了新的技术途径，有望服务于大规模种质资源耐渍性筛选和农业智能表型监测。
我国长江中下游地区春季连绵多雨天气较多，且稻麦轮作为该地区最主要的种植模式，小麦和大麦整个生长季节内都容易遭受田间渍水胁迫，对产量和品质都产生了不利影响。开展作物耐渍性精准评价，对耐渍种质资源筛选和新品种选育具有重要意义。目前，小麦和大麦田间渍害等级评价主要依赖人工调查，存在工作量大、评价效率较低、易受主观因素影响等问题，难以满足大规模种质资源高通量筛选的需求。
针对上述问题，研究团队首次构建了基于无人机RGB成像、多光谱成像及激光雷达（LiDAR）点云的多模态表型采集与渍害智能评级体系。通过RGB影像提取冠层纹理对比度Cons及黄叶占比YLR，利用LiDAR点云获取植株高度PH，并结合多光谱影像筛选出6个关键植被指数VIs，实现作物图、形、谱的特征互补融合。在此基础上，研究团队对比了基于机器学习和深度学习模型的麦类作物渍害智能评级模型。研究发现，多模态特征融合能够有效提高不同渍害等级的识别能力，其中随机森林模型表现最优。在渍水处理后第12天的评级实验中，该模型在不同田块上的分类准确率达到80.48%以上。此外，研究团队提出的黄叶占比自动提取方法与人工标注结果具有较高一致性，决定系数R?达到0.98。基于SHAP算法的模型解释性分析进一步表明，黄叶占比YLR是影响渍害等级判别的重要单一特征，且与多种光谱植被指数存在协同作用，同时表明在充分发挥多模态信息融合优势的同时，仅利用RGB影像提取黄叶占比也具有开展低成本、大范围田间表型筛查的应用潜力。
为进一步降低无人机数据采集成本，研究团队结合四个观测时期的主成分分析、代表性材料表型变化及分日期随机森林建模结果，确定渍水开始后第12天为本试验条件下的最佳渍害评级窗口。通过将多期监测集中于表型差异最明显的关键时点，可在保持较高评级精度的同时减少无人机航测次数，为大规模耐渍种质资源筛选提供更加高效的技术路径。
华中农业大学工学院谢静副教授和光电信息科学与工程专业2023级本科生刘佳桐为论文共同第一作者，华中农业大学工学院刘泉副教授和湖北省农业科学院秦丹丹研究员为论文共同通讯作者。华中农业大学工学院本科生谭威皇、工学院硕士研究生冯桦熊、工学院谭佐军教授以及湖北省农业科学院硕士研究生付大龙和董静研究员等共同参与了本项研究。该研究得到了国家重点研发计划、国家自然科学基金和华中农业大学本科生科技创新专项的支持。
【英文摘要】
High-throughput rating of waterlogging severity in wheat and barley is of significant importance for breeding waterlogging-tolerant varieties and monitoring crop stress. To address the limitations of traditional field surveys, an unmanned aerial vehicle （UAV）-based acquisition platform was utilized to collect RGB images, multispectral imagery, and LiDAR point cloud data from wheat and barley germplasm resources. Data were acquired during the jointing–booting stage before waterlogging and at 5, 12, 20, and 30?days after waterlogging initiation. A multimodal phenotypic feature system fusing visual, morphological, and spectral data was constructed by extracting texture features and the yellow leaf ratio （YLR） from RGB images, plant height （PH） from LiDAR point clouds, and 45 spectral vegetation indices （VIs） from multispectral images. Waterlogging-sensitive indicators, including CON, YLR, PH, and specific VIs, were identified following one-way analysis of variance （ANOVA） and Pearson correlation analysis. Five machine learning models （SVM, RF, BPNN, ELM and CNN） and a deep learning model （YOLOv8-OBB） were developed to determine the optimal model for rating waterlogging. The fusion of tri-modal features significantly enhanced the identification of waterlogging severity. The RF model exhibited superior performance, with a peak accuracy of 94.38% on the test set from Field I and 80.48% in Field II, which served as an internal field validation set under the same experimental environment. YLR extracted automatically from RGB images showed high agreement with manual annotation （R2?=?0.98） and was identified by SHAP as the most important individual predictor, demonstrating its potential for low-cost large-scale screening. Furthermore, interaction effects were observed between YLR and spectral indices, forming an effective combinatorial indicator for stress discrimination. Temporal analysis quantitatively identified approximately 12?days after waterlogging as the optimal UAV monitoring window, when phenotypic differentiation among waterlogging categories was greatest. Concentrating data acquisition at this single time point reduced UAV surveys from four dates to one without compromising classification accuracy. Together, the multimodal fusion framework and the identified optimal monitoring window provide a robust and resource-efficient strategy under comparable experimental conditions for waterlogging tolerance evaluation, intelligent breeding, and field-scale stress monitoring.
论文链接：https://doi.org/10.1016/j.compag.2026.112597
日期：2026-11-03
