| 基于多维度遥感影像的洪河国家级自然保护区沼泽湿地分类方法研究 |
| |
| 引用本文： | 解淑毓,付波霖,李颖,刘兆礼,左萍萍,蓝斐芜,何宏昌,范冬林.基于多维度遥感影像的洪河国家级自然保护区沼泽湿地分类方法研究[J].湿地科学,2021(1). |
| |
| 作者姓名： | 解淑毓 付波霖 李颖 刘兆礼 左萍萍 蓝斐芜 何宏昌 范冬林 |
| |
| 作者单位： | 桂林理工大学测绘地理信息学院;中国科学院湿地生态与环境重点实验室中国科学院东北地理与农业生态研究所 |
| |
| 基金项目： | 国家自然科学基金项目(41801071);广西自然科学基金项目(2018GXNSFBA281015);广西科技计划项目(桂科AD20159037);桂林理工大学科研启动基金项目(GUTQDJJ2017096);广西八桂学者团队项目资助。 |
| |
| 摘 要： | 以洪河国家级自然保护区为研究区,选取了多时相的Sentinel-1B和Sentinel-2A影像为数据源,制定出9种多时相主被动遥感数据组合方案,用于沼泽湿地遥感分类;分别对根据9种方案整合的多维数据集,进行基于尺度继承的多尺度分割,得到面向对象的分割影像,建立与不同方案对应的特征数据集;采用随机森林机器学习算法,对多维特征数据集进行特征优化,并进行参数调优,构建沼泽植物的最优遥感识别模型,实现对沼泽湿地中地物的识别与分类。研究结果表明,采用递归特征消除(recursive feature elimination,RFE)算法的变量优选和采样随机森林的参数调优,可以优化随机森林模型,显著减少数据冗余,整合多时相主被动遥感数据方案九的随机森林模型的最佳参数mtry和ntree分别为4和1500,模型训练精度为93.06%,Kappa系数为0.916,其模型训练精度在所有方案中最高;经过变量优选得到最佳特征变量组合,交叉极化方式的后向散射系数(Mean_VH)对于沼泽分类的重要性比同向极化方式的后向散射系数(Mean_VV)高;光谱特征是遥感图像分类的最主要特征,其中,红光和绿光波段(Mean_R和Mean_G)、多光谱红边波段(Mean_REG1和Mean_REG2)、近红外波段1(Mean_NIR1)、红边植被指数(CIgreen和CIreg)都对分类具有较高的重要性;纹理特征熵(GLCM_Entro、GLCM_Ent2)和位置特征像元坐标(X_Min_Pxl和Y_Max_Pxl)对沼泽湿地分类也起到重要作用,相对于其它特征,形状特征对分类的贡献性较小;利用方案九的数据得到的分类结果的总体分类精度为94.42%,主被动遥感数据的特征变量和不同时相的特征变量都对分类做出重要贡献,其中,6月的多光谱特征变量的贡献大于9月的变量;9月的SAR特征变量的贡献大于6月的变量;对分类贡献最大的9月的特征变量为多光谱红边波段2(Mean_REG2),对分类贡献最大的6月的特征变量为绿光波段(Mean_G);融合多时相主被动遥感数据源,可以将不同数据源、不同时相对分类具有较大贡献的特征变量集合在一起,充分利用光谱信息和雷达数据反映的结构信息,构建多维度最佳特征变量组合方案,有效提高沼泽湿地分类精度。
|
| 关 键 词： | 植物 沼泽 多尺度继承分割 特征优化 随机森林 主被动遥感影像 |
Classification Method on Marsh Wetlands in Honghe National Nature Reserve based on Multi-dimensional Remote Sensing Images |
| |
| XIE Shuyu,FU Bolin,LI Ying,LIU Zhaoli,ZUO Pingping,LAN Feiwu,HE Hongchang,FAN Donglin.Classification Method on Marsh Wetlands in Honghe National Nature Reserve based on Multi-dimensional Remote Sensing Images[J].Wetland Science,2021(1). |
| |
| Authors: | XIE Shuyu FU Bolin LI Ying LIU Zhaoli ZUO Pingping LAN Feiwu HE Hongchang FAN Donglin |
| |
| Institution: | (School of Surveying,Mapping and Geographic Information,Guilin University of Technology,Guilin 541006,Guangxi Zhuang Autonomous Region,P.R.China;Key Laboratory of Wetland Ecology and Environment,Northeast Institute of Geography and Agroecology,Chinese Academy of Sciences,Changchun 130102,Jilin,P.R.China) |
| |
| Abstract: | Taking Honghe National Nature Reserve as the research area,selecting the multi-temporal Sentinel-1B and Sentinel-2A images as the data sources,9 schemes of multi-temporal active and passive remote sensing data combination were developed for the remote sensing classification of marsh wetlands.Multi-scale segmentation based on scale inheritance on a multi-dimensional data set of 9 schemes,obtain object-oriented segmentation images,and establish feature data sets corresponding to different schemes;using random forest machine learning algorithm to optimize the characteristics of multi-dimensional feature data sets,and optimize the parameters,construct the optimal remote sensing recognition model of the marshes,and realize the recognition and classification of ground objects in marsh wetlands.The research results showed that the variable optimization of the recursive feature elimination(RFE)algorithm and the parameter tuning of the sampling random forest could optimize the random forest model and significantly reduce data redundancy.The best parameters mtry and ntree of the random forest model integrating multi-temporal active and passive remote sensing data Scheme 9 were 4 and 1500,the model training accuracy was 93.06%,and the Kappa coefficient was 0.916,its model training accuracy was the highest among all the schemes.After variable optimization,the best combination of feature variables was obtained,the backscattering coefficient(Mean_VH)of the cross-polarization mode was more important for marsh classification than the backscattering coefficient(Mean_VV)of the co-polarization mode.Spectral features were the most important features of remote sensing image classification.Among them,red and green bands(Mean_R and Mean_G),red-edge bands(Mean_REG1 and Mean_REG2),near-infrared band(Mean_NIR1),red-edge band vegetation index(CIgreen and CIreg)were of high importance to classification.The entropy of texture feature(GLCM_Entro,GLCM_Ent2)and the pixel coordinates of location feature(X_Min_Pxl,Y_Max_Pxl)also played important role in the classification of marsh wetlands.Compared with other features,shape features had less contribution to classification.The overall classification accuracy of the classification results obtained by using the data of Scheme 9 was 94.42%.The feature variables of the active and passive remote sensing data and the feature variables of different time phases all made important contributions to the classification.Among them,the contribution of the multi-spectral feature variables in June was greater than those in September,the contribution of the synthetic aperture radar(SAR)characteristic variables in September was greater than those in June.The feature variable of September that contributed the most to the classification was red-edge band(Mean_REG2),and the feature variable of June that contributed the most to the classification was the green band(Mean_G).Fusion of multi-temporal active and passive remote sensing data sources could bring together feature variables that have relatively large contributions from different data sources and relative classifications at different times,and make full use of spectral information and structural information reflected by radar data,construct a multi-dimensional optimal feature variable combination scheme to effectively improve the classification accuracy of marsh wetlands. |
| |
| Keywords: | plant marsh multi-scale inheritance segmentation feature optimization random forest active and passive remote sensing images |
| 本文献已被 CNKI 维普 等数据库收录！ |
|
