基于属性差决策树的全极化SAR影像海冰分类  

Sea Ice Classification of Polarimetric SAR Imagery based on Decision Tree Algorithm of Attributes' Subtraction

作　　者：王常颖 田德政[1] 韩园峰 隋毅 初佳兰[3] Wang Changying;Tian Dezheng;Han Yuanfeng;Sui Yi;Chu Jialan(School of Data Science and Software Engineering,Qingdao University,Qingdao 266071,China;Institute of Big Data Technology and Smart City of Qingdao,Qingdao 266071,China;National Marine Environmental Monitoring Center,Dalian 116023,China)

机构地区：[1]青岛大学数据科学与软件工程学院,山东青岛266071 [2]青岛大学智慧城市与大数据技术研究院,山东青岛266071 [3]国家海洋环境监测中心,辽宁大连116023

出　　处：《遥感技术与应用》2018年第5期975-982,共8页Remote Sensing Technology and Application

基　　金：国家自然科学青年基金项目"基于复杂网络的国产高分卫星影像海岸线自动提取方法研究"(41506198);国家自然科学面上基金项目"海洋灾害大数据分析的系统模型研究及应用"(41476101);"基于多时相高分遥感影像的筏式养殖藻类分类识别方法"(41706105)和"基于复杂网络的国产高分影像围填海类型早期识别方法研究"(41706198);全国统计科学研究项目"基于系统级模型构建的关系复杂大数据分析方法及其应用"(2017LY14)资助

摘　　要：全极化SAR影像往往具有多个极化属性,不同海冰类型在不同极化方式下的成像亮度通常具有明显的差异。提出了一种适合于SAR影像海冰分类的属性差决策树分类分析(SDDT)方法,即在给定n个属性特征的样本基础上,通过计算任何两个属性的属性差特征的分类能力,选择出具有最优分类能力的属性差特征及其最优分裂值,实现海冰分类决策树的构建。采用这种策略,相当于在n+C2n个属性(原始n个属性与C2n个属性差)中寻找最优分类能力的属性,不仅充分考虑了影像中原始多极化属性特征,而且增加了属性差特征的有效利用,进而提高了分类精度。另外,针对计算属性分类能力的衡量指标,在C4.5算法中提出的信息增益比GainRatio基础上,进一步考虑了分裂点的宽度ΔWidth以及分裂点属性总宽度TotalWidth,定义了分类能力指数ClassifyAbility=GainRatio*ΔWidth/TotalWidth。实验表明:采用同样的训练样本,应用SDDT算法挖掘出的海冰分类规则,比C4.5算法挖掘出的分类规则的检测精度至少提高10%以上。Polarimetric SAR image usually has multiple polarization attributes, and the imaging brightness of different sea ice types in different polarization modes is obviously different.A decision tree （SDDT） anal- ysis method on attributes＇ subtractions suited for sea ice classification of polarimetric SAR imagery is pro- posed in this paper.The subtractions between any two attributes based on a given n attributes are calculated.Then their classification ability and optimal divided threshold are calculated.The most effective attribute is discovered and used to construct classification tree.According to this strategy,it equal to find the optimal subtraction attributefrom n ＋Cn^2 features for classification,which include original n attributes and Cn^2 sub- traction attributes.In addition,we use GainRatio to compare the classification ability between different at- tributes firstly.When there are several attributes with a same GainRatio,we consider the width of the split point （AWidth） and the total width of the attribute （TotalWidth） and define a classification ability index Classi f yAbility=GainRatio - AWidth/TotalWidth. By calculating and comparing Classi f yAbility in- dex,an optimal attribute with the largest attribute classification abilitycan be selected. Experiments show that the accuracy of SDDT algorithmhas ten percent higher than that of C4.5 algorithm using a same train- ing samples.

关 键 词：SAR影像 属性差 决策树 海冰分类 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]