测绘通报 ›› 2018, Vol. 0 ›› Issue (3): 43-48.doi: 10.13474/j.cnki.11-2246.2018.0073

• 行业观察 • 上一篇    下一篇

顾及多尺度分割参数的FNEA面向对象分类

孙坤1,2, 鲁铁定1,2   

  1. 1. 东华理工大学测绘工程学院, 江西 南昌 330013;
    2. 流域生态与地理环境监测国家测绘地理信息局重点实验室, 江西 南昌 330013
  • 收稿日期:2017-06-19 出版日期:2018-03-25 发布日期:2018-04-03
  • 通讯作者: 鲁铁定。E-mail:tdlu@whu.edu.cn E-mail:tdlu@whu.edu.cn
  • 作者简介:孙坤(1992-),男,硕士生,主要从事遥感数据处理与应用研究。E-mail:251206743@qq.com
  • 基金资助:

    国家自然科学基金(41464001;41374007);测绘地理信息公益性行业科技专项(201512026);国家重点研发计划(2016YFB0501405);国家重大科学研究计划(2016YFB0502601-04);江西省教育厅科技项目(KJLD12077;GJJ13457)

Research on FNEA Object-oriented Classification Based on Multi-scale Partition Parameters

SUN Kun1,2, LU Tieding1,2   

  1. 1. Faculty of Geomatics, East China University of Technology, Nanchang 330013, China;
    2. Key Laboratory of Watershed Ecology and Geographical Environment Monitoring, NASG, Nanchang 330013, China
  • Received:2017-06-19 Online:2018-03-25 Published:2018-04-03

摘要:

首先对分形网络演化(FNEA)算法做了介绍,通过试验对比4种分割方法棋盘分割、四叉树分割、多尺度分割、光谱差异分割的效果。此外,通过eCognition Developer 8.7分析软件对影像进行多尺度分割预处理,从5开始,以5为单位向上递增,共选取12组参数进行分割试验,采用面向对象CART分类器对分割后影像分类。对比分类效果图可知,小尺度分割参数对分类效果能有较好的提升;对比总体精度及Kappa系数可知,小尺度分割参数分类精度优于大尺度分割参数,且当分割参数Scale为10时,分类精度达到最好的级别。

关键词: FNEA, 分割方法, 多尺度分割, 参数选取, CART分类器

Abstract:

In this paper,the fractal net evolution approach (FNEA) is introduced,and then the effects of four segmentation methods,chessboard segmentation,quadtree-based segmentation,multiresolution segmentation and spectral difference segmentation are compared.In addition,the eCognition Developer 8.7 analysis software is used to perform multi-scale segmentation preprocessing of the image,starting from 5 and increasing in 5 increments.A total of 12 sets of parameters are selected for segmentation experiments.The object-oriented CART classifier is used to classify the segmented image.Compared with the overall accuracy and Kappa coefficient,it can be seen that the classification accuracy of small-scale segmentation parameters is better than that of large-scale segmentation parameters.When the segmentation parameter scale is 10,the small-scale segmentation parameters can be better than the large-classification accuracy to achieve the best level.

Key words: fractal net evolution approach, segmentation method, multi-scale segmentation, parameter selection, CART classifier

中图分类号: