粗糙集高分辨率遥感影像面向对象分类外文翻译资料

 2022-12-31 13:01:08

Rough set theory based object-oriented classification of high resolution remotely sensed imagery

CHEN Jie1, DENG Min1, XIAO Pengfeng2, YANG Minhua1, MEI Xiaoming1

  1. Department of Surveying and Geo-informatics, Central South University, Hunan Changsha 410083, China;
  2. Department of Geographical Information Science, Nanjing University, Jiangsu Nanjing 210093, China

Abstract: Object-oriented classification has been paid more attention in the field of remote sensing. In this paper, a novel object-oriented algorithm based on rough set theory is proposed to classify different objects extracted from high-resolution remotely sensed imagery. The method consists of three steps. Firstly, image segmentation is achieved by watershed transform based on phase congruency gradient and foreground marking to extract image objects. Secondly, texture vector of each object is obtained by Gabor wavelet, and clustering rules is further formed based on the knowledge reduction theory. Finally, according to the restriction of the preliminary clustering result derived from spectral feature of objects, the ultimate classification is achieved referring to the rules. Meanwhile, a new technique to discretize continuous interval-valued attributes is developed, which is very suitable for the object-oriented classification, because the rough set is inadequate for dealing with continuous attributes. The experiments demonstrate that the proposed method can achieve better results and better accuracies.

Keywords: object-oriented classification, rough set, watershed transform, phase congruency, Gabor wavelet, discretization

1 Introduction

With the emergence of high-resolution satellite imagery, such as IKONOS and QuickBird, traditional classification techniques only based on the color and tone of the pixel confront with many challenges (Van der Sande et al., 2003). The high-resolution imagery which have abundant meaningful information is integrating visible spectral response with additional elements, such as shape, texture, and context (Thomas et al., 2003). The pixel-based approaches for high-resolution data will possibly cause the low reliability of landscape objects recognition or classification (Blaschke amp; Hay, 2001; Benz et al., 2004). As early as 1996, Lobo argued that the result obtained by the object-based method was more understandable and has integrality of image objects comparing with the pixel-based one. The analysis of earth observation data has shift from multi-spectral pixel-based to the multi-resolution object-based, and the result derived from the latter is closer to human visual interpretation (Hay et al., 2005).

The extraction of homogeneous segments is fundamental for object-oriented classification, and the accuracy of classification is directly affected by the quality of segmentation. The traditional segmentation approaches include threshold-based, spatial clustering based, region-based, edge-based (Cheng et al., 2001), and some intelligent methods such as the fuzzy theory based segmentation (Lin amp; Tian, 2002). Since that the segmentation technique by merging regions considers the color, texture, shape and scale of imagery objects is developed by Baatz and Schauml;pe (2000), and later is introduced in software eCognition, the object-oriented image analysis approach becomes more and more popular. Jyothi et al. (2008) pointed out that Object Based Image Analysis (OBIA) is facing to strengths, opportunities, weakness and threats simultaneity, and the eCognition technique still has some problems, such as complicated options. After an image is segmented into regions, a following problem is to recognize the objects through merging disciplines and mapping relationship (Zhou amp; Luo, 2009). As an artificial intelligence tool to deal with knowledge, rough set has been applied for pattern recognition, image processing, data analysis and other fields. Rough set has been applied well in remotely sensed imagery processing, such as feature selection (Pan et al., 2002), bands selection (Sun amp; Gao, 2003) and classification (Zhang amp; Wang, 2008).

This paper employs rough set as a tool for acquiring classification rules and combines the way of object-based image analysis in order to achieve satisfied classification of the Quickbird imagery. Because interval-valued attributes of object making up of abundant pixels cannot be discretized by traditional methods (e.g. Guan et al., 2009), a new algorithm to discretizing continuous interval-valued attributes is proposed to fit for the object-oriented classification in this paper. The method to segment objects is described below firstly.

2 EXTRACTION OF OBJECTS FROM REMOTELY SENSED IMAGERY

In order to merge neighboring pixels into homogeneity regions, the multi-scale gradient images are firstly obtained by applying phase congruency model to the original image with Log Gabor wavelet filters, and then a marker-driven watershed transform is used to segment the image. The method based on phase congruency model is better than the Sobel and Canny operator for the extraction of the edges of the objects, because the former has not need low-pass filtering, and not sensitive to local brightness and contrast changes. Indeed, it can get single lines with high precision location (Xiao et al., 2007). The measure of phase congruency is defined as (Kovesi, 1999):

PC ( )2 x = sum;nW x( )⎢⎣An( )x Delta;Phi;n( )x minus;T⎦⎥ (1)

sum;n An( )x ε

Delta;Phi;n( )x = co

剩余内容已隐藏，支付完成后下载完整资料

粗糙集高分辨率遥感影像面向对象分类

摘 要：面向对象的高分辨率遥感影像分类已受到研究者们的广泛关注。本文提出一种基于粗糙集理论的面向对象分类方法以区分高分辨率遥感影像上的不同地物。首先， 利用基于相位一致梯度与前景标记的分水岭变换进行影像分割， 提取图像斑块； 然后， 利用 Gabor 小波提取斑块的纹理特征， 进而根据粗糙集理论提取纹理分类规则； 最后， 在对象光谱特征的初步分类结果， 根据纹理分类规则得到最终结果基础上。依据粗糙集理论只能处理离散属性数据， 本文重点提出一种适用于面向对象分类的连续区间属性离散化方法。实验表明本文方法可取得较好分类结果与较高分类精度。

关键词： 面向对象分类， 粗糙集， 分水岭变换， 相位一致， Gabor 小波， 离散化

1引言

随着 IKONOS、QuickBird 等高分辨率遥感影像的出现， 传统的单纯基于像元的分类技术面临诸多挑战(Van Der Sande 等，2003)。高分辨率遥感影像光谱特征明显且地物纹理、形状等信息丰富(Thomas 等，2003)，仅依靠光谱特征的目标表达或地物分类的结果可靠性较低(Blaschke amp; Hay，2001；Benz 等，2004)。早在 1996 年，Lobo 将面向对象与基于像元的影像分类方法进行对比， 实验表明前者结果更易被理解图斑完整性更好。对地观测数据的分析手段已经由过去的基于多光谱像元向基于多尺度面向对象转变， 这更加符合人们解译影像的视角(Hay 等，2005)。

面向对象分类需要解决的首要问题是均质图像斑块的提取， 因而影像分割的效果直接决定了分类准确度。图像分割方法有直方图阈值法、特征空间聚类法、区域提取法、边缘检测法(Cheng 等，2001) 以及模糊分割(林瑶 amp; 田捷，2002)等智能方法。自从 Baatz Schauml;pe(2000)提出以光谱、纹理、形状、尺度等特征为基础的区域合并分割技术并被引入 eCognition， 面向对象图像分析方法越来越受到人们的欢迎。Jyothi 等(2008)(Object based image aualysis OBIA)同时面临着优势、机遇、弱点与挑战， 该软件存在诸如参数设置较复杂等问题。通过分割得到影像目标在空间上的离散基元(即图像斑块)后， 如何通过归并规则和映射关系实现对地物目标的分类识别是研究难点(周成虎 amp; 骆剑承， 2009)。粗糙集作为知识获取的智能手段已应用于模式识别、图像处理和数据分析等诸多研究领域，在特征选择(潘励等， 2002)、波段选择(孙立新 amp; 高文， 2003)、分类(张东波 amp; 王耀南，2008)等遥感影像处理方面也取得了较好效果。

2遥感影像对象的提取

为了获取具有较高同质度的图像斑块，先获得相位一致梯度图像再用基于前景标记的分水岭变换分割全色影像。因为，相位一致算子相比于 Sobel 和 Canny 等空域边缘特征检测算子， 不需要先使用低通滤波去除噪声，对图像局部亮度和对比度的变化不敏感，能够对线形物体产生高定位精度的单线响应(肖鹏峰，2007)。本文采用的相位一致计算式为(Kovesi，1999)：

式中，W(x)为展频因子，通过扩展频带可以提高相位一致的响应度；An(x) 为第 n 个傅里叶分量的幅值；T为噪声估计，只有当相位偏离度大于 T 时才能计算相位一致； ⎣ ⎦表示取正值时为其本身， 其他的则为0； ε为避免分母为零而引入的一个常量； Delta;Phi;n(x)为相位偏离函数；PC为相位一致值。引入 log Gabor 小波函数(Field，1987)计算局部相位信息， 因为它可以在偶对称滤波器保持零直流分量的情况下构造任意带宽的滤波器， 可使任何局部特征的相位一致得以保持。在笛卡尔坐标系中， log Gabor 函数的频率响应为：

式中， omega;0为滤波器的中心频率， k/omega;0为滤波器形状比值常量。

为了提高分水岭变换的分割效果，首先采用扩展最小变换(Soille，2003)标记那些有意义的局部最小区域；然后用强制最小技术修改相位一致梯度图像，使得局部最小区域仅出现在所标记的位置，而其他无关的局部最小区域则相对地进行“上推”；最后利用浸入式的分水岭变换方法(Vincent amp; Soille，1991)对重建后的相位一致梯度图像进行分割。

3 基于粗糙集的面向对象分类

3.1 粗糙集的基本概念

定义 1(决策表)： 一个知识表达系统由四元组表示： S=(U，R，V， f )； 其中， U 为非空有限对象集合(论域)； R 为对象属性集合； V 为属性值的集合， Va 为属性 a 的值域； f：Utimes;R→V 为信息函数。若 R=Ccup;D， C 为条件属性集， D 为决策属性集， 则该知识表达系统为决策表。

定义 2(不可分辨关系)： 设 Asube;R， 对于forall;aisin;A，若 Xiisin;U 与 Xjisin;U 有相同的属性值，则有：

IND(A) =

{(Xi， X j ) | (Xi， X j ) isin;U times;U，forall; isin;a A f， a (xi ) = fa (xj )}

定义 3(近似集)： 对于每个子集 Xsube;U 与 IND(A)，X 的上近似与下近似可分别定义如下：

Aminus; (X ) =cup; {Y Yi | i isin;U | IND(A Y)， i cap; X ne;phi;}

Aminus;(X ) =cup; {Y Yi | i isin;U | IND(A Y)， i sube; X}

3.2基于粗糙集的面向对象分类框架

利用分水岭变换取得图像斑块后， 由粗糙集构造出决策规则以实现基于知识的粗糙聚类。分类方法的框架如图1， 其中实线方框表示相关数据或操作步骤， 箭头指示各项内容之间的相互关系。具体过程描述如下：

步骤 1 将分割后的遥感影像视为一个知识系统， 提取图像斑块的纹理特征作为条件属性C ，并根据研究区域的特点选取各类地物的样本对象构成决策表。

步骤 2 利用连续区间属性离散化方法对信息决策表的条件属性C 进行离散化。

图 1 分类方法框架

步骤 3 计算条件属性 C 与决策属性 D 的等价集、下近似集以及依赖度。

步骤4 计算forall; isin;c C 的重要度， 并检查属性c约简后的决策表一致性， 进一步核实属性的约简可行性。即从决策表中将属性集C 的某个属性 c 删除， 每次删除后立即检查决策表的一致性， 若不出现步骤 5 消去重复的规则以及决策规则中的冗余属性值。重复步骤 3、4 直至没有决策冗余和属性冗余。

步骤6 规则获取。逐个计算条件属性 c 的等价集与决策属性 forall; isin;d D 的等价集, 计算二者之间规则可信度, 若可信度为 1 则形成规则。

步骤 7 规则化简。对所获取的规则进行化简, 得到最终的分类规则。

步骤 8 基于平均光谱属性, 对影像对象进行最邻近距离分类, 得到初步分类结果。

步骤 9 在初步分类基础上, 将除样本之外其他图像斑块的纹理属性经离散化后与分类规则进行匹配, 得到最后的影像分类结果。

4连续区间属性值离散化方法

4.1离散化原理描述

设有限对象集合 U=(x1,x2,⋯ ,xn), 条件属性集 C=(c1,c2,⋯ ,cm)。那么, 对于forall;cisin;C, 遥感影像中各对象所对应的值域 Vc 可表示为:

式中，vcmin (xi ) 与 vcmax (xi ) 构成第 i 个对象在属性 c 的区间值， n 为对象个数。

对于forall;cisin;C， 断点集合 {d0c， d1c，d2c， dkc， dkc 1} 定义了 Vc 上的一个离散状态 Pc = {[d0c， d1c )，

[d1c， d2c)， ，[dkc， dkc 1)} 。对于forall; sube;k R ， lc = d0c lt;

d1c lt; lt; dkc lt; dkc 1 = rc ， 且 Vc = [d0c， d1c ) cup;[d1c，

d2c ) cup; cup;[dkc， dk 1c ) ， 其中lc 、 rc 分别表示值域Vc 的最小值与最大值。因此， P Pc 定义了一个新

的 决 策 表 S p = (U R V， ， p，f p ) ，对于forall;xisin;U ，

i =

剩余内容已隐藏，支付完成后下载完整资料

资料编号：[279406]，资料为PDF文档或Word文档，PDF文档可免费转换为Word

原文和译文剩余内容已隐藏，您需要先支付 30元 才能查看原文和译文全部内容！立即支付

以上是毕业论文外文翻译，课题毕业论文、任务书、文献综述、开题报告、程序设计、图纸设计等资料可联系客服协助查找。

您可能感兴趣的文章