基于WorldView-2数据的基塘系统遥感分类研究  

Remote Sensing Classification for Dike-pond System based on Worldview-orldview-2 Data

作　　者：黎丰收 刘凯 刘洋 唐焕丽 柳林 彭力恒 LI Fengshou;LIU Kai;LIU Yang;TANG Huanli;LIU Lin;PENG Liheng(School of Geography and Planning,Sun Yat-sen University,Guangdong Key Laboratory for Urbanization and Geo-Simulation,Center of lntegrated Geographic Information Analysis,Guangzhou 510275,Guangdong,P.R.China;Department of Geography,University of Cincinnati,Cincinnati,OH 45221,USA;Guangzhou Zengcheng District Urban and Rural Planning,Guangzhou 511300,Guangdong,P.R.China;College of Geographical Science,Guangzhou University,Geo-Informatics of Public Security,Guaugzhou 510006,Guangdong,P.R.China)

机构地区：[1]中山大学地理科学与规划学院综合地理信息研究中心、广东省城市化与地理环境空间模拟重点实验室,广东广州510275 [2]Department of Geography, University of Cincinnati, Cincinnati, OH 45221, USA [3]广州市增城区城乡规划与测绘地理信息研究院,广东广州511300 [4]广州大学地理科学学院公共安全地理信息分析中心,广东广州510006

出　　处：《湿地科学》2018年第5期587-596,共10页Wetland Science

基　　金：国家自然科学基金项目(41001291);广东省科技计划项目(2017A020217003);广州市科技计划项目(201510010081)资助

摘　　要：分布在热带和亚热带低洼地区的基塘系统是一种典型的人工复合农业生态系统。通过水陆交互作用和能量多级利用,基塘系统创造出较高的生态价值和经济价值。掌握基塘系统土地利用的类型、面积和比例关系,有助于了解和分析基塘系统的现状及其生态模式和功能。以广东省佛山市顺德区西南部为研究区,利用高空间分辨率WorldView-2卫星影像数据,采用结合面向对象分类和基于像元的混合分类方法,提取基塘系统中的塘面和基面（基面又细分为道路、植被和裸土等）;从分类特征组合与分类器两个方面,通过对比分析,筛选出基塘系统遥感影像分类的最优方法;根据最优分类结果,计算出塘基与塘面的面积比例。研究结果表明,采用面向对象分类方法分离的塘面和塘基的总体精度为82.2%;在基于像元进一步对塘基进行分类的过程中,结合8个多光谱波段光谱反射率平均值、影像亮度值和8个红边波段指数特征组合方式的随机森林分类器的分类精度最高,达到85.6%;得到的基-塘面积比例约为4.9∶5.1,植被-塘面的面积比例约为3.5∶6.5。提出的结合面向对象和基于像元的混合分类方法能够对基塘系统进行精细的遥感影像分类,能准确估计基塘系统中各地物的面积和基-塘比例等指标,并为基塘系统生态功能的调查和基塘系统的合理保护与开发提供参考和数据支持。Dike-pond system is a typical artificial mixed agricultural ecosystem, mostly distributed in tropical and subtropical areas. Dike-pond system creates great ecological and economic values due to its water-land interactive agriculture and multi-level energy exploitation. It is essential to investigate the land use types, their areas, and proportions in dike-pond system as a base to further analyze its current situation, and ecological mechanism and functions. In this study, a study area was selected in the southwestern Shunde District in Foshan city, Guangdong, China. A multi-level image classification approach that combines object-oriented and pixel-based classification methods was proposed to extract the information of pond, dike, and land cover types on the dike, through high-resolution WorldView-2 satellite image data. By optimizing feature space and comparing Bayes, KNN, and Random Forest classifier, the classification result with the highest overall accuracy was generated and further applied to estimate the area of each kind of land cover type and the ratio between the pond and dike areas. The result showed that the overall classification accuracy of pond and dike by the first-level object-oriented classification was 82.2%. In the second-level pixel-based classification further applied only to the dike, random forest classifier with the feature space comprised of the brightness of pixels, 8 multi-spectral bands, and 8 derived indices from red-edge bands generated the classification result with the highest overall accuracy, approximately 85.6%. A further estimation indicated that the dike-pond area ratio and the vegetation-pond area ratio were approximately 4.9∶5.1 and 3.5∶6.5, respectively. The multi-level image classification approach that combines object-oriented and pixel-based classification methods in this study generated the land use maps with high-accurate dike-pond system. The area statistics and pond-dike ratios estimated from the classification results could provide data support and decision ref

关 键 词：基塘系统 面向对象分类 WorldView-2 遥感分类 

分 类 号：S181.6[农业科学—农业基础科学] TP79[自动化与计算机技术—检测技术与自动化装置]