| 基于支持向量回归模型的水稻田甲烷排放通量预测研究 |
| 摘要点击 3185 全文点击 1843 投稿时间：2012-09-24 修订日期：2012-12-13 |
| 查看HTML全文
查看全文 查看/发表评论 下载PDF阅读器 |
| 中文关键词 支持向量回归 BP人工神经网络 甲烷排放通量 水稻田 k折交叉验证 |
| 英文关键词 support vector regression back propagation-artificial neural network methane emission paddy fields k-fold cross validation |
| 作者 | 单位 | E-mail | | 陈强 | 北京师范大学资源学院, 北京 100875
北京师范大学环境演变与自然灾害教育部重点实验室,北京 100875 | chenqiang@mail.bnu.edu.cn | | 蒋卫国 | 北京师范大学环境演变与自然灾害教育部重点实验室,北京 100875 | jiangweiguo@bnu.edu.cn | | 陈曦 | 北京师范大学环境演变与自然灾害教育部重点实验室,北京 100875 | | | 袁丽华 | 北京师范大学环境演变与自然灾害教育部重点实验室,北京 100875 | | | 王文杰 | 中国环境科学研究院,北京 100012 | | | 潘英姿 | 中国环境科学研究院,北京 100012 | | | 王维 | 中国环境科学研究院,北京 100012 | | | 刘孝富 | 中国环境科学研究院,北京 100012 | | | 刘海江 | 中国环境监测总站,北京 100012 | |
|
| 中文摘要 |
| 利用静态箱和气相色谱仪法获取水稻田甲烷排放通量数据,选取大气温度、土壤5 cm深温度、土壤pH、土壤Eh、土壤含水量和地表生物量作为影响因子. 应用建立在结构风险最小化优化上的支持向量回归(ε-SVR)模型,采用留一法交叉检核网格搜索法(LOOCV)优化ε-SVR 预测模型的参数,采用k折交叉检验的方法依据平均相对误差(MRE)和均方根误差(RMSE)对模型的精度进行验证,并与BP人工神经网络(BP-ANN)模型比较,评价ε-SVR预测模型的准确性. 结果表明,通过LOOCV选择最优的惩罚因子C和损失系数ε,并由此构建的 ε-SVR 预测模型预测值和实测值具有很好的一致性,通过11折交叉验证后,测试样本的平均MRE为44%,平均RMSE为16.21 mg·(m2·h)-1. 通过与BP-ANN模型比较,预测值和实际值相关系数达0.863,各项指标均优于BP-ANN预测模型,说明 ε-SVR 模型能够适用于水稻田甲烷排放通量的预测. |
| 英文摘要 |
| The methane emission data of paddy fields was obtained by using the static chamber and gas chromatography, and six parameters including atmospheric temperature, soil temperature at 5 cm depth, pH of soil, Eh of soil, soil moisture and ground biomass were selected as the primary influencing factors of methane emission. The support vector regression (ε-SVR) model was built on the optimization of structural risk minimization, and the parameters of the ε-SVR model were optimized using Leave-one-out Cross Validation (LOOCV). The prediction accuracy of model was evaluated by k-fold cross validation with the mean relative error (MRE) and the root mean square error (RMSE). In addition, the accuracy of the ε-SVR model was analyzed by comparison with the Back Propagation-Artificial Neural Network (BP-ANN) model. The results indicated that the predicted value of the ε-SVR model with the parameters C and ε optimized by LOOCV was in good agreement with the measured value, and the average MRE of test samples was 44% and the average RMSE was 16.21 mg·(m2·h)-1 in the process of 11-fold cross validation. Compared with the BP-ANN model, the correlation coefficient was 0.863, and all the indicators were better. It demonstrated that the ε-SVR model could be applied to the prediction of methane emission of paddy fields. |
|
|
|
