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

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. © 2024 Chinese Society of Agricultural Machinery. All rights reserved.