基于J48决策树的面向对象方法的土地覆被信息提取  

Land cover information extraction from remote sensing images using object-based image analysis method integrated with decision tree

作　　者：孙宇翼 赵军利[1,3] 王苗苗[1] 刘勇[1] 

机构地区：[1]兰州大学资源环境学院 [2]甘肃省地图院 [3]中国人民解放军61175部队

出　　处：《国土资源遥感》2016年第4期156-163,共8页Remote Sensing for Land & Resources

基　　金：国家自然科学基金"遥感影像多尺度分割质量评价与参数优选方法研究"(编号:41271360);甘肃省湿地自然边界确定试点项目共同资助

摘　　要：过去10多a来,面向对象的影像分析方法在高分辨率影像信息提取中表现出了明显优势,得到了快速发展。该方法中一个难题是,如何有效地建立满足健壮性和通用性准则的分类规则集。基于数据挖掘原理的决策树方法有望提供有效的解决方案。选用WEKA J48算法从影像光谱、纹理和地形特征等诸多参数中优选出部分参数构建决策树分类模型,以此建立分类规则集,并集成于面向对象的影像分类方法中。利用Landsat5 TM影像和ASTER数字高程模型数据进行的甘肃省会宁县白草塬地区土地覆被分类的结果表明,本方法所建立的分类规则集具有较佳的健壮性和通用性,其分类精度明显优于基于像元的最大似然法和基于试错性规则集的面向对象法。Object-based image analysis, which has been advantageous over classic pixel - based image classification. developed rapidly over the last decades, performs One of the key problems within this paradigm is to automatically build robust and transferable rule sets for segment classification. It has been identified promisingly to develop rule sets by means of decision tree based on data mining. The authors suggest a decision tree model integrated with J48 algorithm embedded in Weka to select parameters from a set of spectral, textural and terrain features relevant to rule sets for segment classification. Based on this method, the authors used Landsat5 TM image data and ASTER digital elevation model to establish land cover classification in the study area, i. e. , Baicaoyuan area in Huining county, Gansu Province. Rule sets developed in this way perform acceptable robustness and transferability. Accuracy assessment proves that this method has significantly higher classification accuracy than other pixel -based methods based on employing maximum likelihood and objected -based nearest neighbor logic.

关 键 词：面向对象的影像分析 J48算法 决策树 土地覆被分类 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]