首页 | 本学科首页   官方微博 | 高级检索  
     

基于人工神经网络与空间仿真模拟的区域森林碳估算比较——以龙泉市为例
引用本文:秦立厚,张茂震,袁振花,杨海宾. 基于人工神经网络与空间仿真模拟的区域森林碳估算比较——以龙泉市为例[J]. 生态学报, 2017, 37(10): 3459-3470
作者姓名:秦立厚  张茂震  袁振花  杨海宾
作者单位:浙江农林大学, 浙江省森林生态系统碳循环与固碳减排重点实验室, 临安 311300;浙江农林大学, 环境与资源学院, 临安 311300,浙江农林大学, 浙江省森林生态系统碳循环与固碳减排重点实验室, 临安 311300;浙江农林大学, 环境与资源学院, 临安 311300,浙江农林大学, 浙江省森林生态系统碳循环与固碳减排重点实验室, 临安 311300;浙江农林大学, 环境与资源学院, 临安 311300,浙江农林大学, 浙江省森林生态系统碳循环与固碳减排重点实验室, 临安 311300;浙江农林大学, 环境与资源学院, 临安 311300
基金项目:国家自然科学基金项目(30972360,41201563);浙江农林大学农林碳汇与生态环境修复研究中心预研基金;浙江省林业碳汇与计量创新团队项目(2012R10030-01);浙江省林学一级重中之重学科学生创新计划项目资助(201515)
摘    要:森林是生态系统的重要组成部分,准确估算森林碳储量及其分布对于评价森林生态系统的功能具有重要意义。以龙泉市为研究区,利用2009年99个森林资源清查样地数据和同年度Landsat TM影像数据,采用高斯序列协同仿真(SGCS)与BP神经网络方法(BPNN)分别模拟森林地上部分碳密度及其分布,并进行了对比分析。随机将样本数据分成70个建模样本和29个检验样本。通过模型检验,BP神经网络预测值与实测值的相关性达到0.67,相对均方根误差为0.63,空间仿真方法预测值与实测值的相关性为0.68,相对均方根误差为0.63,空间仿真方法预测能力略高于神经网络方法。仿真结果表明,基于BP神经网络模拟的森林碳总量为11042990 Mg,平均碳密度为36.10 Mg/hm2,总体森林碳密度均值高于样地平均值8.82%。基于空间仿真模拟的森林碳总量为11388657 Mg,平均碳密度为37.23 Mg/hm2,总体森林碳密度均值高于样地平均值9.40%。对比分析可知:高斯协同仿真模拟和BP神经网络虽然在碳总量估算值上与抽样数据估计值相近,但两种方法在估测值的频率分布以及研究区碳分布上有较大的差异。与BP神经网络相比,序列高斯协同模拟结果更接近系统抽样样地实测值,全部样地预测值与实测值的相关性达到0.75,在估计区域森林碳空间分布上有明显优势。在碳密度值域与频率分布方面,序列高斯协同模拟结果分布更合理。综上所述,序列高斯协同模拟在森林碳空间估计方面要优于BP神经网络。

关 键 词:森林碳储量  高斯协同仿真模拟  BP神经网络  森林资源清查数据  TM影像
收稿时间:2016-03-01
修稿时间:2016-11-14

Comparison of regional forest carbon estimation methods based on back- propagation neural network and spatial simulation: A case study in Longquan County
QIN Lihou,ZHANG Maozhen,YUAN Zhenhua and YANG Haibin. Comparison of regional forest carbon estimation methods based on back- propagation neural network and spatial simulation: A case study in Longquan County[J]. Acta Ecologica Sinica, 2017, 37(10): 3459-3470
Authors:QIN Lihou  ZHANG Maozhen  YUAN Zhenhua  YANG Haibin
Affiliation:Zhejiang Provincial Key Laboratory of Carbon Cycling in Forest Ecosystems and Carbon Sequestration, Zhejiang Agriculture & Forestry University, Lin''an 311300, China;School of Environmental & Resource Sciences, Zhejiang Agriculture & Forestry University, Lin''an 311300, China,Zhejiang Provincial Key Laboratory of Carbon Cycling in Forest Ecosystems and Carbon Sequestration, Zhejiang Agriculture & Forestry University, Lin''an 311300, China;School of Environmental & Resource Sciences, Zhejiang Agriculture & Forestry University, Lin''an 311300, China,Zhejiang Provincial Key Laboratory of Carbon Cycling in Forest Ecosystems and Carbon Sequestration, Zhejiang Agriculture & Forestry University, Lin''an 311300, China;School of Environmental & Resource Sciences, Zhejiang Agriculture & Forestry University, Lin''an 311300, China and Zhejiang Provincial Key Laboratory of Carbon Cycling in Forest Ecosystems and Carbon Sequestration, Zhejiang Agriculture & Forestry University, Lin''an 311300, China;School of Environmental & Resource Sciences, Zhejiang Agriculture & Forestry University, Lin''an 311300, China
Abstract:Quantifying the carbon stocks of forest is critical for understanding the dynamics of carbon fluxes in terrestrial ecosystems and the atmosphere as well as monitoring ecosystem responses to environmental changes. However, due to the lack of methods and data, results of forest carbon estimation from different studies shown large difference, which presents a great uncertainty in the evaluation of forest carbon sink. Different methods can be used to estimate the carbon storage in the same study area, which can be compared with the advantages and disadvantages of each method and provides guidance for forest carbon estimation. On the basis of National Forest Inventory (NFI) data and the Land-sat TM image data collected in Longquan County, Zhejiang Province in 2009, we applied two methods, namely error back-propagation neural network (BPNN) and sequential Gaussian co-simulation (SGCS) to reproduce the distribution of above-ground forest carbon. We randomly divided plots into two sets, a 70-plot set for modeling and a 29-plot set for testing. For the model test, the correlation coefficient of predictive value and the plot data was 0.67 and 0.68 for BPNN and SGCS, respectively. Both of the two methods have the same RRMSE value (0.63). The predictive ability of SGCS was slightly higher than that of BPNN. The estimation results using BPNN showed that the sum of above-ground carbon is 11042990 Mg and the mean carbon density was 36.10 Mg/hm2 which was higher than the average from the sample plots with a relative error of 8.82%. The SGCS showed that the sum of above-ground carbon was 11388657 Mg with a mean carbon density 37.23 Mg/hm2 which was higher than the average from the sample plots with a relative error of 9.4%. Comparative analysis showed the carbon densities estimated using these two methods are both close to that calculated from the NFI data. However, there were some differences between the two methods with respect to the estimation of the frequency distribution and the carbon distribution in the study area. Predictive value of sample plot obtained using the SGCS method was closer to the plot data value than that obtained using the BPNN. And the correlation between predictive value and the plot data was 0.75, which proved that there were obvious advantages in estimating the spatial distribution of forest carbon. In addition, in terms of carbon density range and frequency distribution, SGCS was more reliable. This study further verifies the effectiveness of the SGSC which could provide effective methods for the estimation of regional forest carbon storage.
Keywords:forest carbon storage  sequential Gaussian co-simulation  back-propagation neural network  National Forest Inventory  TM image
本文献已被 CNKI 等数据库收录!
点击此处可从《生态学报》浏览原始摘要信息
点击此处可从《生态学报》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号