地球信息科学学报 ›› 2020, Vol. 22 ›› Issue (10): 2088-2097.doi: 10.12082/dqxxkx.2020.190483

• 遥感科学与应用技术 • 上一篇    下一篇

“四同”条件下周口城区高分一号遥感影像分类对比研究

叶杰1(), 孟凡晓1,*(), 白潍铭1, 张斌1, 郑金明2   

  1. 1.河南省航空物探遥感中心,郑州 450053
    2.西北核技术研究所,西安 710024
  • 收稿日期:2019-09-02 修回日期:2019-10-28 出版日期:2020-10-25 发布日期:2020-12-25
  • 通讯作者: 孟凡晓 E-mail:21807621@qq.com;mengfx1030@163.com
  • 作者简介:叶杰(1979— ),男,河南驻马店人,高级工程师,主要从事遥感技术研究与应用。E-mail:21807621@qq.com
  • 基金资助:
    全国矿山环境恢复治理状况遥感地质调查与监测(DD20190705);河南省航空物探遥感中心自主科研项目(2018-05);2018年度省财政厅地质科研项目(豫国土资发[2018]98号)

A Comparative Study on the Classification of GF-1 Remote Sensing Images for Zhoukou Urban under the Four Identical Condition

YE Jie1(), MENG Fanxiao1,*(), BAI Weiming1, ZHANG Bin1, ZHENG Jinming2   

  1. 1. Henan Aero Geophysical Survey and Remote Sensing Center, Zhengzhou 450053, China
    2. Northwest Institute of Nuclear Technology, Xi'an 710024, China
  • Received:2019-09-02 Revised:2019-10-28 Online:2020-10-25 Published:2020-12-25
  • Contact: MENG Fanxiao E-mail:21807621@qq.com;mengfx1030@163.com
  • Supported by:
    National Remote Sensing Geological Survey and Monitoring for Mine Environmental Restoration and Control in China(DD20190705);Independent Scientific Research Project of Henan Aero Geophysical Survey and Remote Sensing Center Research on Multi-Element Remote Sensing Information Extraction Technology of Urban Geological Environment in 2018(2018-05);Henan Provincial Department of Finance Geological Research Project in 2018(豫国土资发[2018]98号)

摘要:

目前大多数面向像元、面向对象遥感影像分类对比研究算法、软件、样本均不同,引入多方面系统误差导致结果一定程度上不严谨。为更准确比较2种分类方法,本文采用面向像元、面向对象2种分类方式,在同软件平台、同分类器、同训练样本、同验证样本,即“四同”条件下对2018年4月17日高分一号周口城区融合影像进行分类对比研究,并完成主、客观评价精度评价。结果表明:① “四同”条件下2种分类方式、CART(Classification and Regression Tree)、SVM(Support Vector Machine)、RF(Random Forests)3种机器学习算法均能识别周口城区主要地物类型,而面向对象的分类效果明显优于面向像元分类,与前人研究结论一致。其中面向像元分类效果最好的是RF算法,总体分类精度为78.02%,Kappa系数为0.72;面向对象分类效果最好的是RF算法,总体分类精度为93.40%,Kappa系数为0.92;② 尽管由于光谱特征相似、分布交叉,单类别建筑用地、交通用地用户精度与生产者精度较低,但面向对象分类较面向像元分类效果明显提升,以RF分类为例,建筑用地生产者精度由56.18%提高至92.13%,用户精度由69.44%提高至87.23%;交通用地生产者精度由72.15%提高至89.87%,用户精度由72.15%提高至92.20%;③ 与前人研究成果比较,本文在“四同”条件下实现了更科学、更严谨的面向像元、面向对象遥感分类方法对比,对后续高分辨率遥感影像分类具有一定参考意义。

关键词: “四同”条件, 面向像元, 面向对象, 高分一号, 机器学习, 对比分析, 单类别, 总体精度

Abstract:

At present, due to different classification methods, softwares, and samples used for classification which could introduce various systematic errors, the majority of studies for comparing the advantages and disadvantages of pixel-based and object-based classification are unprecise to a certain degree. To make a better comparison between the pixel-based and object-based approaches, pixel-based and object-based classification methods were adopted to classify the fused image of panchromatic and multispectral images provided by GF-1 satellite in the main urban district of Zhoukou on April 17, 2018, using the same hardware and software environments, classifier, training samples, and verification samples, namely four identical conditions. Subjective and objective evaluations of the pixel-based and object-based classification methods were made. For comparison, three machine learning algorithms including Classification and Regression Tree (CART), Support Vector Machine (SVM), and Random Forest (RF) were used as the classifiers in the pixel-based and object-based classification procedure. Results show that (1) both pixel-based and object-based approaches could recognize the main urban targets, which was consistent with previous research results. However, the object-based method had a better overall accuracy than the pixel-based method on average. For pixel-based image classification, RF produced the highest overall accuracy (78.02%) and the Kappa coefficient (0.72); for object-based image classification, RF also achieved the highest overall accuracy (93.40%) and the Kappa coefficient (0.92), which demonstrated that RF was the best machine learning algorithm for classifying Zhoukou urban targets; (2) due to similar spectral signature and cross-distribution, the Producer's Accuracy (PA) and User's Accuracy (UA) of building land, and traffic land were lower. However, the object-based classification produced much higher PA and UA than pixel-based classification in classifying building land and transportation land. Taking RF as example, the PA of building land increased from 56.18% to 92.13%, with the UA increasing from 69.44% to 87.23%, and the PA of traffic land increased from 72.15% to 89.87%, with the UA increasing from 72.15% to 92.20%; (3) compared with previous related researches, this paper conducts a more scientific and rigorous evaluation for pixel-based and object-based classification methods under the four identical conditions, which provides valuable references to classify urban targets using high resolution satellite remote sensing images in the future.

Key words: the four identical condition, pixel-based, object-based, GF-1, machine learning, comparison analysis, each category, overall accuracy