基于神经网络的杉木人工林碳通量影响因素的选择

Factors Selection for Affection of Carbon Flux Based on Artificial Neural Network

对碳循环影响因素的研究是预测碳通量的重要环节,也是研究碳循环的重要基础。利用亚热带地区湖南省会同杉木人工林生态系统国家野外科学观测研究站2008年7—9月的碳通量和环境因子观测数据,采用遗传神经网络模型对碳通量预测因素进行优化选择,并与传统的相关分析方法进行对比分析。结果表明:模型CIV.8(输入参数包括空气温度Ta、光合有效辐射Par、大气CO2浓度ρc、空气相对湿度Rh、风速Ws、土壤温度Ts)是所有模型中模拟效果最好的。光合有效辐射与碳通量的相关性最强,相关系数是-0.704(P=0.000);降雨量与碳通量的相关性最弱,相关系数是0.002(P=0.854)。最多输入变量或最复杂的神经网络结构并不能得到最好的模型。

Selecting the affection factors for carbon cycle is a critical step for predicting carbon flux and the foundation to study the mechanism of carbon cycle.In this paper,C flux and meteorological data were collected o-ver a three months period between July and September in 2008 at Huitong National Research Station of Forest Eco-system.This paper proposes a method of driving factors selection based on artificial neural network （ANN）.What is more,ANN is used to select the driving factors according to the criteria that model showed.Finally,the results based on ANN were compared to correlation analysis.The results showed that CIV.8 （input parameters included Ta,Par,ρc ,Rh,Ws,Ts）model was the best of all models.The strongest correlation was between Par and carbon flux （R=-0.704,P=0.000）;but Precipitation （Prec）had weak positive correlation with carbon flux （R=0.002,P=0.854）.The best results pertaining to ANN modeling do not derive from structure complexity or a grea-ter number of input parameters.