CART集成学习方法估算平原河网区不透水面覆盖度  

Estimation of impervious surface percentage of river network regions using an ensemble leaning of CART analysis

作　　者：李晓宁[1] 张友静[1,2] 佘远见 陈立文[1] 陈静欣[3] 

机构地区：[1]河海大学地球科学与工程学院,南京210098 [2]河海大学水文水资源与水利工程科学国家重点实验室,南京210098 [3]江苏省测绘产品质量监督检验站,南京210013

出　　处：《国土资源遥感》2013年第4期174-179,共6页Remote Sensing for Land & Resources

基　　金：国家重点基础研究发展计划(973计划)(编号:2010CB951101);水利部公益性行业科研专项项目(编号:201101024)共同资助

摘　　要：快速扩展的不透水面已成为影响高密度河网生态系统的主要因素。以平原河网城市的典型区域苏锡常地区为研究区,提出了一种基于分类与回归树(classification and regression tree,CART)集成学习的不透水面覆盖度(impervious surface percentage,ISP)遥感估算方法,利用Landsat TM数据构建多源特征集,采用变精度粗糙集进行数据约简,以获取CART决策树的最佳属性变量,结果优于传统的单一CART方法,但得到的初始估算结果中ISP高值区低估现象较为严重,借助温度植被干旱指数(temperature vegetation dryness index,TVDI)与ISP的相关性,寻找后处理规则对其进行改善。实验结果表明,经变精度粗糙集进行属性约简和TVDI后处理的CART集成学习方法估算精度明显提高,ISP估算值与ISP参考值之间的均方根误差为10.0%,决定系数为0.89,可用于平原河网地区ISP的估算。The rapid expansion of impervious surface has become a major factor affecting ecosystem health of the high density river network. This paper provides an approach to estimate impervious surface percent （ISP） through the ensemble leaning of CART analysis based on variable precision rough sets （VPRS）. Landsat TM and ALOS imagery were utilized to construct the ISP predictive model; then, in order to get the best attribute variables of CART decision tree, the authors adopted VPRS to extract optimum feature subset from multi - source feature sets. The results illustrate the validity of this ensemble leaning, and prove that this method can obtain estimated accuracy better than the traditional single CART method. However, in the initial estimation results, ISP＇ s high value area is underestimated relatively seriously. The authors have discovered that the temperature vegetation dryness index （TVDI） and ISP have an intensive relationship with each other： the increase of ISP will cause the increase of local TVDI significantly. Therefore, the post -processing rule extracted from the relationship is used to improve the results. According to the verification results, the method combined with VPRS reduction and post -processing rule in CART algorithm has fairly higher analysis precision than the traditional single CART learning algorithm. The root mean square error between estimated ISP value and reference ISP is 10.0%, with the correlation coefficient being 0.89, so it can be used to estimate the ISP in plain river network region.

关 键 词：不透水面 分类回归树 变精度粗糙集 TVDI 平原河网地区 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]