基于CART算法的纳木措湖泊面积精确提取  

Accurate extraction of Namtso Lake area based on decision tree CART algorithm

作　　者：郑晨键 刘炜[1] 薛永福 冯珂 ZHENG Chenjian;LIU Wei;XUE Yongfu;FENG Ke(Xizang Key Laboratory of Optical Information Processing and Visualization Technology,Xizang Minzu University,Xianyang 712000,China)

机构地区：[1]西藏民族大学西藏光信息处理与可视化技术重点实验室,陕西咸阳712000

出　　处：《哈尔滨商业大学学报（自然科学版）》2024年第3期288-295,共8页Journal of Harbin University of Commerce：Natural Sciences Edition

基　　金：国家自然科学基金(62062061;61762082);西藏自然科学基金项目(XZ2019ZRG-43);西藏自治区科技厅项目(XZ202001ZY0055G)。

摘　　要：为了能够更精确地提取到湖泊水体的范围,对比MLC、SVM和基于多特征的全新遥感影像CART决策树分类方法,对西藏自治区的纳木措湖泊进行自动提取研究.选择Landsat 8OLI卫星遥感影像数据作为数据源.将得到的不同特征的影像进行组合,组合成全新的多特征遥感影像.决策树方法具有结构清晰、快速、简单、有效的优点,而CART算法可以根据选取的训练样本获取节点和阈值,不需要反复试验来确定阈值,避免了基于传统专家知识方法的主观性,因此采用CART算法构建决策树模型对研究区域进行湖泊水体的提取.结果表明CART决策树方法总体精度为99.82%,Kappa系数为0.996,MLC总体精度为96.814%,Kappa系数值为0.929,SVM总体精度为98.045%,Kappa系数值为0.956,总体精度相较于SVM和MLC分别提高了3%、1.775%,Kappa系数提高了0.067、0.04.CART决策树、MLC、SVM所得到的湖泊面积分别为2009.43、2014.93、2026.9 km^(2),MLC和SVM得到的结果比CART决策树分类法存在更多的错分和漏分现象,主要是将山地中的阴影信息错认为是水体,CART决策树方法识别到的细小水体更加连续,对于湖泊边界识别的效果也更好.In order to extract the scope of lake water more accurately,MLC,SVM and the CART decision tree classification method based on multi-feature remote sensing images was compared to automatically extract Namtso Lake in Tibet Autonomous Region.Landsat 8OLI satellite remote sensing image data was selected as the data source.The obtained images with different features were combined to form a new multi-feature remote sensing image.Decision tree method has the advantages of clear structure,fast,simple and effective,while CART algorithm can obtain nodes and threshold values according to the selected training samples without repeated trials to determine the threshold value,avoided the subjectivity based on traditional expert knowledge methods.Therefore,CART algorithm was used to build decision tree model to extract lake water in the study area.The results showed that the overall accuracy of CART decision tree method was 99.82%,Kappa coefficient was 0.996,MLC overall accuracy was 96.814%,Kappa coefficient was 0.929,SVMoverall accuracy was 98.045%,Kappa coefficient was 0.956.Compared with SVM and MLC,the overall accuracy was increased by 3%and 1.775%respectively,and the Kappa coefficient was increased by 0.067 and 0.04.The lake area obtained by CART decision tree,MLC and SVM was 2009.43 km^(2),2014.93 km^(2) and 2026.9 km^(2) respectively.The results obtained by MLC and SVM had more misclassification and missing phenomenon than that obtained by CART decision tree classification.The main reason was that the shadow information in mountain mistaken as water body.The small water bodies identified by CART decision tree method were more continuous,and the effect of lake boundary identification was also better.

关 键 词：纳木措 多特征 CART算法 决策树 水体提取 湖泊面积 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]