摘要
根据湖泊监测的特点,采用支持向量机(SVM)方法,反演太湖叶绿素a的浓度分布.将2005年8月太湖29个现场水质监测点数据分为训练测试样本集和验证样本集,利用训练测试样本集以及与其时间同步的MODIS遥感影像,分别构建了4种SVM模型.对比分析表明,直接以波段反射率以及水深信息构成输入向量的SVM模型预测效果最好.利用训练测试样本构建了线性回归模型、主成分分析模型(PCA)以及神经网络模型(ANN),并利用验证数据比较了上述3种模型与SVM模型的预测结果.结果表明ANN模型和SVM模型预测能力明显优于另外2种模型,其中SVM模型对低值和高值均有较好的预测精度,平均相对误差仅为15.91%,预测精度比ANN模型提高了10%.利用SVM模型和ANN模型分别反演了2005年8月15日太湖叶绿素a浓度分布,比较了2种模型反演结果的异同,分析了太湖叶绿素a分布特征及其成因.
Considering limited monitoring points in lakes, support vector machines （SVM） was chosen to retrieve chlorophyll-a （chl-a） concentration with MODIS data. In this research, 29 monitoring points of Taihu Lake in August, 2005 were divided into training and testing group and validating group. 4 SVM models, a linear regression model, a principle component analysis （PCA） model and an artificial neural net （ANN） model were constructed using traning and testing group. Validation data was used to compare with other 3 models and SVM model The retrieving precision of ANN model and SVM model were both better than the other two models, especially the average relative error of SVM model was only 15.91%, and 10% lower than ANN model. Chl-a distribution of Taihu Lake on Aug 15, 2005 was retrieved synchronously using SVM model and ANN model. At last, the main factors affecting the distribution of chl-a were discussed.
出处
《中国环境科学》
EI
CAS
CSCD
北大核心
2009年第1期78-83,共6页
China Environmental Science
基金
国家“973”项目(2008CB418003)
“十一五”国家科技支撑计划(2006BAC02A15)
