Modeling canopy transpiration of young poplar trees (Populus × euramericana cv. N3016) based on Back Propagation Artificial Neural Network

PDF HTML阅读 XML下载 导出引用 引用提醒 BP人工神经网络模拟杨树林冠蒸腾 DOI: 10.5846/stxb201308262155 作者: 作者单位: 中国科学院沈阳应用生态研究所;中国科学院大学,中国科学院沈阳应用生态研究所,中国科学院沈阳应用生态研究所,中国科学院沈阳应用生态研究所,中国科学院沈阳应用生态研究所,中国科学院沈阳应用生态研究所 作者简介: 通讯作者: 中图分类号: 基金项目: 国家"十二五"科技支撑计划项目(2011BAD38B0203) Modeling canopy transpiration of young poplar trees (Populus × euramericana cv. N3016) based on Back Propagation Artificial Neural Network Author: Affiliation: State Key Laboratory of Forest and soil Ecology,Institute of Applied Ecology,Chinese Academy of Sciences,State Key Laboratory of Forest and soil Ecology,Institute of Applied Ecology,Chinese Academy of Sciences,State Key Laboratory of Forest and soil Ecology,Institute of Applied Ecology,Chinese Academy of Sciences,State Key Laboratory of Forest and soil Ecology,Institute of Applied Ecology,Chinese Academy of Sciences,State Key Laboratory of Forest and soil Ecology,Institute of Applied Ecology,Chinese Academy of Sciences,State Key Laboratory of Forest and soil Ecology,Institute of Applied Ecology,Chinese Academy of Sciences Fund Project: 摘要 | 图/表 | 访问统计 | 参考文献 | 相似文献 | 引证文献 | 资源附件 | 文章评论 摘要:利用2008和2010年的气温、饱和差、总辐射和叶面积指数作为模型输入,液流法观测的蒸腾速率作为模型输出,建立了用于杨树林冠蒸腾模拟的BP人工神经网络模型,利用2009年的观测数据对模型的模拟能力进行了检验,并应用连接权值计算得到的输入变量对输出变量的相对贡献进行了敏感性分析。结果表明:建立的BP人工神经网络蒸腾模型可以很好的模拟林冠蒸腾大小和季节变化,模拟的绝对误差和绝对相对误差的平均值分别为0.11 mm/d和9.5%,纳什效率系数为0.83;输入变量对蒸腾的相对贡献以及蒸腾与输入变量之间的相关性大小顺序相同,均为总辐射 > 叶面积指数 > 饱和差 > 气温。 Abstract:Artificial neural network (ANN) is a practical tool and a powerful alternative to mechanism models in operation of hydrology modeling. In this paper, a three layer back propagation (BP) artificial neural network model was developed to estimate the canopy transpiration of young poplar trees (Populus × euramericana cv. N3016) in Northeast China. The combination of air temperature (Ta), vapor pressure deficit (VPD), solar radiation (Rg) and leaf area index (LAI) was chosen as the input variables, while the transpiration measured by sap flow was chosen as output variable. Observational data in growing season of 2008 and 2010 was used to develop model. The number of neurons in the input layer and output layer was 4 and 1, respectively based on the number of input and output variables. Levenberg-Marquardt (LM) algorithm was selected as the learning algorithm to train the network. Tansig and Logsig function were selected as the transfer function in the hidden layer and output layer, respectively. The learning rate and momentum factor were set as 0.1 and 0.01, respectively. The number of neurons in the hidden layer was optimized as 9 by a trial and error method. So the network structure of the developed model was determined as 4:9:1. After 49 times training, the optimal BP ANN transpiration model was determined. The data samples in 2009 were chosen to evaluate the developed model. Results showed that BP ANN transpiration model can successfully simulate the seasonal variation of transpiration. The slope of the regression equation between the simulated and measured transpiration was 0.99, while R2 was 0.85. Maximum and minimum absolute error were 0.28 mm/d and 0.003 mm/d. Mean absolute error and mean absolute relative error were 0.11 mm/d and 9.5%, and Nash-Sutcliffe coefficient of efficiency were 0.83, which all indicated the high accuracy and efficiency of developed BP ANN model. However, compared with the model performance during training process, the accuracy decreased slightly, which turned out the existence of over-fitting. At last, a sensitivity analysis of input variables on transpiration was performed using the connection weights of the developed ANN model to assess the relative importance of input variables. Results showed that the relative contribution of radiation to simulated transpiration (33.46%) was maximal, while that of temperature (16.58%) was minimal. The relative contribution of LAI (30.19%) was larger than that of VPD (19.77%), but less than that of radiation. Magnitude order of correlation coefficient between input variables and transpiration and relative contribution of input variables to transpiration presented the same order of Rg > LAI > VPD > Ta, which provided the physical interpretation of why the developed BP ANN model can well simulate the transpiration despite it did not explain the physical process of transpiration. It must be realized that the data employed for developing ANN model contain important information about the physical process of transpiration. The BP ANN can well learn and remember this kind of information by adjusting its weights during training process, and represent it when new variables in evaluation samples were inputted into the model. 参考文献 相似文献 引证文献