基于特征提取的CNN-BiLSTM长白落叶松树干液流密度预测

Sap Flow Prediction of Larix olgensis based on Feature Extraction CNN-BiLSTM

【目的】树干液流密度是影响植物蒸腾的重要因素,其大小受太阳辐射、空气湿度、土壤水分等环境变量影响,准确测量或估计树干液流密度对于了解森林的水分利用效率、研究全球气候变化具有重要意义。【方法】本文提出一种基于特征提取的长白落叶松树干液流预测方法。首先,选取黑龙江省佳木斯市孟家岗林场的长白落叶松为试验对象,将林场气象监测站测得的9个环境变量和Granier法所测长白落叶松树干液流密度进构建数据集。其次,通过皮尔逊相关系数方法对环境变量与树干液流密度之间存在的线性关系进行分析。然后,应用转移熵方法对数据进行特征提取,提取出四个主要环境变量作为模型的输入。最后,搭建基于特征数据的CNN-BiLSTM混合模型,将数据集输入模型进行训练与测试。【结果】通过对比试验,将相关系数(R^(2))、均方根误差(RMSE)和平均绝对误差(MAE)作为预测精度的评价指标,CNN-BiLSTM方法在性能表现上优于BP、CNN、CNN-LSTM模型的预测方法。【结论】转移熵方法可以很好地分析环境变量与长白落叶松液流引入时滞后的因果关系,基于转移熵的特征提取构建的CNNBiLSTM模型能有效的提高长白落叶松树干液流的预测精度。

【Objective】The stem sap flow density is a crucial factor affecting plant transpiration.Its magnitude is influenced by environmental variables such as solar radiation,air humidity,and soil moisture.Accurate measurement or estimation of stem sap flow density is of great significance for understanding forest water use efficiency and studying global climate change.【Method】In this paper,a feature extraction-based prediction method for sap flow in Larix olgensis is proposed.Firstly,Larix olgensis from Mengjiagang forest in Jiamusi city,Heilongjiang province,are selected as the experimental subjects,and data are collected from meteorological monitoring stations in the forest,including nine environmental variables and the sap flow of Larix olgensis measured by the Granier method,to construct the dataset.Secondly,the linear relationship between environmental variables and sap flow is analyzed using Pearson correlation coefficient method.Then,feature extraction is performed using transfer entropy method to extract four major environmental variables as inputs for the model.Finally,a CNN-BiLSTM hybrid model based on feature data is constructed,and the dataset is input into the model for training and testing.【Result】Through comparative experiments,with correlation coefficient,root mean square error,and mean absolute error as evaluation indicators of prediction accuracy,the CNN-BiLSTM method outperforms the prediction methods of BP,CNN,and CNN-LSTM models in terms of performance.【Conclusion】The transfer entropy method can effectively analyze the causal relationship between environmental variables and lagged introduction of Larix olgensis sap flow.The CNN-BiLSTM model constructed based on transfer entropy-based feature extraction can effectively improve the prediction accuracy of stem sap flow density in Larix olgensis.