基于贝叶斯改进的人工神经网络毛竹林碳通量估算  

Estimation of Phyllostachys heterocycla cv.pubescens Carbon Flux Based on Artificial Neural Networks Improved by Bayesian

作　　者：汪雪[1] 周国模[1] 周健[2] 徐小军[1] 顾志华[2] 李楠[2] 

机构地区：[1]浙江农林大学环境与资源学院浙江省森林生态系统碳循环与固碳减排重点实验室,浙江临安311300 [2]浙江农林大学信息工程学院,浙江临安311300

出　　处：《西北林学院学报》2017年第1期203-209,共7页Journal of Northwest Forestry University

基　　金：国家自然科学基金项目(31500520);浙江省自然科学青年基金项目(LQ15C160003)

摘　　要：碳通量的估算对全球碳循环有很重要的意义,目前精度最高的涡度相关法并不能完全满足大区域估算的需求,因此寻求一种估算大区域碳通量的方法具有重要的意义。对毛竹(Phyllostachys heterocycla cv.pubescens)林碳通量相关气候因子进行敏感性分析,定量分析气候因子对碳通量的影响程度,鉴于贝叶斯在处理不确定性方面的优势,提出了一种基于贝叶斯(Bayesian)改进的BP神经网络(Back Propagation Neural Network,BPNN)方法对浙江省安吉县毛竹林碳通量进行估算,并利用毛竹林通量塔实测值进行验证。结果表明:1)经敏感性分析,对毛竹林碳通量估算影响程度较高的因子为潜热通量、入射辐射、土壤温度、摩擦风速等,这与实际生态过程中对于碳通量贡献最大的因素基本一致。2)利用本研究方法估算的毛竹林碳通量的变化范围为-1.25~0.57μmol·m^(-2)·s^(-1),年平均值为-0.057 85μmol·m^(-2)·s^(-1),表现为碳汇,与实测情况相符。3)通过B-BPNN方法估算的碳通量值与碳通量实际值相关性达到0.932,均方根误差(RSME)为0.103,偏差(Bias)为0.001 06,该结果比单纯的BPNN方法估算结果的相关性(0.827)提高12.70%,均方根误差(0.155)降低33.55%。说明B-BPNN算法能改进单纯算法模型中估算碳通量的不确定性,提高碳通量预测的精度。Carbon flux estimation is of very important significance for the global carbon cycle. Currently, the eddy covariance method with the highest precision can not fully meet the demand for the estimation of large area. Therefore,to find a carbon flux estimation method suitable for large area is of great significance. In this paper,the carbon flux sensitivity analysis of Phyllostachys heterocycla cv. pubescens plantations related to climatic factors was conducted which quantitatively analyzed the influence of climatic factors on the carbon flux in Anji,Zhejiang. A method based on artificial neural networks improved by Bayesian was proposed. The estimated results were verified by using eddy correlation method. The results showed that 1） sensitivity analysis demonstrated that factors that had significant effects on carbon flux prediction were latent heat flux,incident radiation, soil temperature, wind friction velocity,it was consistent with actual ecological processes. 2） Estimation of P. heterocycla cv. pubescens carbon flux in this method had the range of 1.25-0.57μmol·m^-2·s^-1 ,and annual average was -0. 057 85μmol·m^-2·s^-1 ,expressed as carbon sinks,in line with the actual situation. 3） The correlationship between estimated carbon flux with B-BPNN and actual carbon flux reached 0. 932, with root mean square error （RSME） of 0. 103, bias （Bias） of 0. 001 06,compared to the results from the simple method of BPNN, the correlation （0. 827） increased 0. 105,RSME （0. 155） reduced 0. 052,bias （0. 001 3） reduced 0. 000 24. It explained that algorithm of B-BPNN could improve the uncertainty in the estimation of carbon flux by the simple algorithm of BPNN, and further improve the accuracy of the prediction.

关 键 词：碳通量估算 微气象因子 敏感性分析 B-BPNN算法 

分 类 号：S795.9[农业科学—林木遗传育种]