Zhou Lei, Zhang Huichun, Bian Liming, Tian Ye, Zhou Haopeng
College of Mechanical and Electronic Engineering, Nanjing Forestry University, Nanjing 210037, P. R. China.
Jiangsu Co-Innovation Center of Efficient Processing and Utilization of Forest Resources, Nanjing Forestry University, Nanjing 210037, P. R. China.
Plant Phenomics. 2024 Jul 29;6:0205. doi: 10.34133/plantphenomics.0205. eCollection 2024.
Drought stress is one of the main threats to poplar plant growth and has a negative impact on plant yield. Currently, high-throughput plant phenotyping has been widely studied as a rapid and nondestructive tool for analyzing the growth status of plants, such as water and nutrient content. In this study, a combination of computer vision and deep learning was used for drought-stressed poplar sapling phenotyping. Four varieties of poplar saplings were cultivated, and 5 different irrigation treatments were applied. Color images of the plant samples were captured for analysis. Two tasks, including leaf posture calculation and drought stress identification, were conducted. First, instance segmentation was used to extract the regions of the leaf, petiole, and midvein. A dataset augmentation method was created for reducing manual annotation costs. The horizontal angles of the fitted lines of the petiole and midvein were calculated for leaf posture digitization. Second, multitask learning models were proposed for simultaneously determining the stress level and poplar variety. The mean absolute errors of the angle calculations were 10.7° and 8.2° for the petiole and midvein, respectively. Drought stress increased the horizontal angle of leaves. Moreover, using raw images as the input, the multitask MobileNet achieved the highest accuracy (99% for variety identification and 76% for stress level classification), outperforming widely used single-task deep learning models (stress level classification accuracies of <70% on the prediction dataset). The plant phenotyping methods presented in this study could be further used for drought-stress-resistant poplar plant screening and precise irrigation decision-making.
干旱胁迫是杨树生长的主要威胁之一，对植物产量有负面影响。目前，高通量植物表型分析作为一种快速、无损分析植物生长状况（如水和养分含量）的工具已得到广泛研究。在本研究中，将计算机视觉和深度学习相结合用于干旱胁迫下杨树幼苗的表型分析。培育了四个品种的杨树幼苗，并施加了5种不同的灌溉处理。采集植物样本的彩色图像进行分析。进行了两项任务，包括叶片姿态计算和干旱胁迫识别。首先，使用实例分割来提取叶片、叶柄和叶脉区域。创建了一种数据集增强方法以降低人工标注成本。通过计算叶柄和叶脉拟合线的水平角度实现叶片姿态数字化。其次，提出了多任务学习模型以同时确定胁迫水平和杨树品种。叶柄和叶脉角度计算的平均绝对误差分别为10.7°和8.2°。干旱胁迫增加了叶片的水平角度。此外，以原始图像作为输入，多任务MobileNet实现了最高准确率（品种识别为99%，胁迫水平分类为76%），优于广泛使用的单任务深度学习模型（预测数据集上胁迫水平分类准确率<70%）。本研究提出的植物表型分析方法可进一步用于抗旱杨树植株筛选和精准灌溉决策。
