基于1DCNN融合多源表型数据的杨树干旱胁迫评估方法

Assessment of Poplar Drought Stress Level Based on 1DCNN Fusion of Multi-source Phenotypic Data

目前关于不同杨树品种抗旱性的研究主要集中在利用传统测量方法获取形态结构和生理生化表型参数进而分析杨树的抗旱性,依据多源成像传感器提取的表型参数指标确定杨树干旱胁迫等级的方法较为少见。为了阐明杨树耐旱的表型机制、筛选抗旱性树种和明确杨树抗旱等级,本文以杨树不同性别的喜水和耐旱品种为研究对象,在杨树苗期进行梯度干旱胁迫处理,通过热红外以及RGB多源成像传感器获取杨树冠层温度参数与颜色植被指数表型数据,并建立基于1DCNN的多任务分类模型划分杨树苗期品种抗旱等级与干旱胁迫等级等2个分类任务,探究杨树性别与生长时间对杨树干旱胁迫响应机制的影响。结果表明,以27组数据变量降维后的4个特征作为模型变量,与传统机器学习算法SVM、RF、XGBoost相比,本文提出的1DCNN多任务分类模型在杨树品种抗旱等级分类与单株干旱胁迫等级分类2个任务中的模型分类精度皆达到最优,分类准确率分别为81.8%和62.3%;引入杨树的性别和生长时间后共6个特征作为模型的输入变量后,杨树苗期品种抗旱等级与干旱胁迫等级的分类精度显著提高,1DCNN多任务分类模型在2个分类任务中的准确率分别达到93.5%与76.6%,模型分类准确率分别提高11.7个百分点与14.3个百分点。研究结果表明,通过热红外与RGB成像传感器获取多源表型数据,并建立1DCNN多任务分类模型对实现杨树干旱胁迫等级评估的可行性,同时表明杨树的性别和生长时间作为模型输入变量能够有效提升模型的分类精度,可为筛选杨树抗旱性品种提供新的思路与方法。

At present,the research on drought resistance of different poplar varieties mainly focuses on using traditional measurement methods to obtain morphological structure and physiological and biochemical phenotypic parameters to analyze the drought resistance of poplars.The method of determining the drought stress level of poplars based on phenotypic parameter indicators extracted by multi-source imaging sensors is relatively rare.In order to clarify the phenotypic mechanism of poplar drought resistance,screen drought-resistant tree species and clarify the drought resistance level of poplars,taking water-loving and drought-resistant varieties of poplars of different genders as the research objects,gradient drought stress treatment at the seedling stage of poplars was conducted.The phenotypic data of poplar canopy temperature parameters and color vegetation index were obtained by thermal infrared and RGB multi-source imaging sensors,and a multi-task classification model based on 1DCNN was established to divide the two classification tasks of poplar seedling variety drought resistance level and drought stress level,so as to explore the influence of poplar gender and growth days on the response mechanism of poplar drought stress.The results showed that compared with the traditional machine learning algorithms SVM,RF and XGBoost,the proposed 1DCNN multi-task classification model achieved the best classification accuracy in the two tasks of poplar variety drought resistance classification and individual drought stress classification,with classification accuracy rates of 81.8%and 62.3%respectively,using the four features after dimension reduction of 27 groups of data variables as model variables.After introducing the sex and growth days of poplars as the input variables of the model,the classification accuracy of the drought resistance and drought stress levels of poplar seedling varieties was significantly improved,and the accuracy of the 1DCNN multi-task classification model in the two classification tasks was 93.5%and 76.6%,respectively,and the classification accuracy of the model was improved by 11.7 percentage points and 14.3 percentage points,respectively.The research results showed that it was feasible to obtain multi-source phenotypic data through thermal infrared and RGB imaging sensors and establish a 1DCNN multi-task classification model to realize the evaluation of poplar drought stress level.At the same time,it was showed that the sex and growth days of poplars as model input variables can effectively improve the classification accuracy of the model,which can provide ideas and methods for screening poplar drought-resistant varieties.