基于小样本学习和骨架提取算法的干旱胁迫杨树苗表型解析

Plant phenotyping for drought-stressed poplar saplings using few-shot learning and skeleton extraction algorithm

杨树是木材生产加工的重要来源和防护林建设的重要树种,中国的杨树人工林栽培规模居世界首位。然而,全球变暖加剧使得干旱成为杨树培育中最为严重的典型的非生物胁迫,给木材原料形成、森林资源保护带来挑战。该研究采用低成本的机器视觉成像研究干旱胁迫杨树苗表型快速解析方法。首先,提出了基于YOLOv8-pose的杨树骨架提取算法,识别植株上的叶片个体及内部的关键节点,实现植株整体形态结构信息的提取。其次,引入小样本学习技术,提出基于极少量人工标注的训练数据集扩充方法,大幅降低YOLOv8-pose模型训练的人工标注成本。然后,依据杨树骨架信息计算每个叶片主叶脉、叶柄的倾角,并将倾角信息转换为频数分布,便于人工智能算法建模。最后,研究并对比了传统机器学习及深度学习分类器,选择最优算法建立干旱胁迫等级分级模型。结果表明,小样本学习YOLOv8-pose模型在叶片识别、关键点提取任务中表现优异(交并比阈值为0.5时的平均精度分别为0.798和0.914);基于一维卷积神经网络分类模型和倾角频数分布特征的杨树苗干旱胁迫等级分级模型优于其他对比方法,分类准确率为0.850。该研究提出的杨树表型解析方法可为缺水杨树识别、抗旱杨树筛选提供新的技术支持。

Poplar tree is one of the most important tree species in wood production and protective forest construction.The poplar artificial forest cultivation is ranked first in the world.Current severe drought and typical abiotic stress under global warming can pose a great challenge to the production of wood raw materials and the protection of forest resources.A series of planting strategies(such as the identification of water-deficient plants)have been used to cultivate the poplar species of high drought tolerance against drought stress.These approaches are required for the rapid observation and analysis of the growth status of poplar saplings,in order to execute irrigation decisions or drought-resistant variety screening.Previous research has focused mainly on the detection of drought stress,where the macroscopic morphology of plants can be classified to predict the drought level.Only a few studies have been focused on the digital expression of water deficiency symptoms in plants.Alternatively,the water content of plants has been directly predicted using relatively complex and expensive sensing devices.In this article,a low-cost machine vision was adopted to realize the lant phenotyping of poplar saplings under drought stress.Firstly,an extraction framework of the poplar skeleton was proposed to identify the individual leaves and key nodes inside the plant using the YOLOv8-pose detection model,in order to extract the overall morphological and structural information of the plant.Secondly,the few-shot learning techniques were introduced,including the training dataset expansion(or considered as an automatic synthesis for datasets)using minimal manual annotation.The annotated leaves and stems were extracted from the images,and then added into a‘component pool’.New images of the poplar plant were also selected randomly.All the individual leaves and the key points were detected(including the leaf tip,connection point between blade and petiole,and connection point between the petiole and the trunk).Among them,the few-shot learning was trained using 600 synthetic poplar images with only 10 manually annotated samples.Thirdly,the angles of the main vein and petiole in each leaf were calculated from the skeleton information of the poplar sapling.Furthermore,the frequency distribution vector was converted(ranges from-90°to 90°,with an interval of 10°)for the artificial intelligence-based modeling.Finally,partial least square discrimination analysis(PLSDA),support vector machine(SVM),multilayer perception(MLP),and convolutional neural network(CNN)classifiers were compared to evaluate the drought stress level.The optimal model was selected for the final drought-stress grading.The results showed that the YOLOv8-pose model with the few-shot learning performed the best in the leaf recognition and key point extraction tasks,where the mean average precision with 0.5 as the threshold for intersection over union(mAP50)values were 0.798 and 0.914,respectively.The effectiveness and reliability few-shot learning-based YOLOv8-pose model were also validated using an independent test dataset composed of real samples and accurate manual annotations.Meanwhile,the manual annotation cost was significantly reduced for YOLOv8-pose model training.Subsequently,the classification model with one-dimensional CNN and leaf angle frequency distribution was superior in the drought stress grading task,with the highest classification accuracy of 0.850.In addition,the improved models shared relatively small parameters(model size less than 7 MB)with high speed.Low-cost computing can be realized in cloud service platforms with limited sources or embedded systems(such as Nvidia Jetson series devices).The phenotyping can also provide new technical support to identify the water-deficient poplar saplings and screen the drought-resistant plants.