基于像斑空间关系的遥感图像分类

引用本文

李亮, 舒宁, 龚龑, 王凯. 基于像斑空间关系的遥感图像分类[J]. 国土资源遥感, 2013,25(1): 77-81
LI Liang, SHU Ning, GONG Yan, WANG Kai. Remote sensing classification method based on image segment spatial relationship[J]. REMOTE SENSING FOR LAND & RESOURCES,2013,25(1): 77-81 复制到剪切板

Permissions

《国土资源遥感》编辑部

基于像斑空间关系的遥感图像分类

李亮, 舒宁, 龚龑, 王凯

武汉大学遥感信息工程学院,武汉 430079

第一作者简介: 李亮(1987-), 男, 博士研究生, 主要从事遥感图像智能化解译研究。E-mail:liliang1987wuda@163.com。

收稿日期: 2012-03-31

基金: 国家自然科学基金项目(编号: 41101412)和中央高校基本科研业务费专项基金项目(编号: 3101009,20102130201000139)共同资助

摘要

为充分挖掘遥感图像本身包含的空间关系信息,弥补基于光谱信息的传统图像分类方法的不足,提高分类精度,提出了一种基于像斑空间关系的遥感图像分类方法。通过图像分割获取像斑,利用最大似然法获取初始分类结果,引入马尔科夫随机场对像斑的空间关系予以描述,通过地物的类别邻接矩阵定量地描述各地物类别之间的空间关系,从而对图像的分类结果进行修正,最后采用条件迭代的方法获取最终的图像分类结果,精度较好。实验结果表明,该方法应用于高分辨率遥感图像可取得较好的分类效果。

关键词: 马尔科夫随机场; 地物类别邻接矩阵; 条件迭代; 像斑; 图像分类

中图分类号:TP751.1 文献标志码:A 文章编号:1001-070X(2013)01-0077-05 doi: 10.6046/gtzyyg.2013.01.14

Remote sensing classification method based on image segment spatial relationship

LI Liang, SHU Ning, GONG Yan, WANG Kai

School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China

Abstract

An image classification method based on the spatial relationship of image segment is proposed with the purpose of excavating the spatial relationship between image segments and compensating for deficiencies of the traditional image classification method based on spectral information. Image segmentation is used to get image segments for original image classification employing maximum likelihood (ML) method. Then the spatial relationship of image segments is described by Markov random field (MRF). Quantitative spatial relationship can be obtained by class adjacency matrix (CAM) so as to revise the result of classification. After that the iterated conditional mode (ICM) algorism for classification is presented, which can yield results with higher accuracy. Experimental results show that this method has been functioning well in classification experiments with high resolution remote sensing images.

Keyword: Markov random field(MRF); class adjacency matrix (CAM); iterated conditional mode(ICM); image segment; image classification

Show Figures

0 引言

遥感图像分类是遥感领域的研究热点之一。按照是否需要训练样本, 遥感图像分类方法可分为监督分类法和非监督分类法这2大类。前者分类结果较准确^[1], 但需要合适的训练样本, 而高质量的训练样本一般较难获取; 后者则不需要训练样本, 自动化程度较高, 但分类精度难以保证。两者大多数都是单纯基于光谱信息的传统分类方法。

充分利用遥感图像本身的空间关系信息, 弥补基于光谱信息的传统图像分类方法的不足, 提高分类精度, 已成为很多研究者的共识。赵红蕊等^[2]提出了简单加入空间关系的实用图像分类方法, 构造了2个空间关系波段, 实现了空间约束; 蔡晓斌等^[3]提出了基于图斑空间关系的遥感图像分类方法, 加入了空间对象的空间推理知识; 张路等^[4]提出了顾及上下文的遥感图像模糊聚类方法, 引入了空间隶属度的概念; 乔程等^[5]提出了一种空间邻接支持下的遥感图像分类方法。上述方法均考虑了像素或像斑的空间关系, 取得了较好的分类效果; 但是在空间关系定量描述方面仍显得不足, 未能充分挖掘图像包含的空间关系信息。

本文试图充分挖掘图像本身包含的空间关系信息, 利用马尔科夫随机场(Markov random field, MRF)来描述像斑的空间关系, 以期提高图像分类的精度。类别邻接矩阵(class adjacency matrix, CAM)可定量地描述各地物类别之间的空间关系, 利用该矩阵可对图像分类结果进行修正。本文研究了基于像斑^{[6, 7]}空间关系的遥感图像分类方法, 并在高分辨率遥感图像分类实验中取得了较好的效果。

1 研究方法

假设图像中地物类别的集合为ω ={ω ₁, ω ₂, …, ω _m}, m为地物类别的数目。令X=(X_k, k=1, 2, …, n)表示所有像斑特征向量的集合, 其中X_k为第k个像斑的特征向量, n为像斑数目。将图像中所有像斑的分类结果看作一次随机试验C_j=C_j(k), k=1, 2, …n, 其中C_j(k)∈ ω , 为第j次随机试验结果中第k个像斑的分类结果。所有随机试验结果C_j的集合构成了样本空间C=(C_j, 1≤ j≤ L), 其中L=mⁿ。依据后验概率最大原则获取的最佳分类结果为

C_t=arg $\begin{matrix} \max_{C_{j} \in C} \end{matrix}$ p(C_j|X)=arg $\begin{matrix} \max_{C_{j} \in C} \end{matrix}$ [p(X|C_j)p(C_j)/p(X)], (1)

式中: C_t为最佳分类结果; p(X|C_j)为在随机试验结果C_j的条件下, 特征向量为X的概率; p(C_j)为随机试验结果C_j出现的概率; p(X)为常量。式(1)也可写为

C_t=arg $\begin{matrix} \max_{C_{j} \in C} \end{matrix}$ [p(X|C_j)p(C_j)]。 (2)

依据像斑类别客观分布规律, 本文假设像斑的特征向量仅依赖于该像斑对应的类别, 可得到

p(X|C_j)= $\begin{matrix} \overset{n}{\underset{k = 1}{П}} \end{matrix}$ p[X_k|C_j(k)], (3)

式中p[X_k|C_j(k)]表示在第k个像斑类别为C_j(k)的条件下, 对应特征向量为X_k的概率。将式(3)代入式(2), 可以得到

C_t=arg $\begin{matrix} \max_{C_{j} \in C} \end{matrix}$ {p(C_j) $\begin{matrix} \overset{n}{\underset{k = 1}{Π}} \end{matrix}$ p[X_k|C_j(k)]}。 (4)

1.1 空间关系模型

令G={R, E}是个图; R={R₁, R₂, …, R_s}, 表示节点R_i的集合(i=1, 2, 3, …, s); E是连接结点的边的集合。假定σ 是G的近邻系统, σ ={σ (R₁), σ (R₂), …, σ (R_s)}, 这里σ (R_i)表示结点R_i的近邻, R_i∉σ (R_i); 若R_j∈ σ (R_i), 则R_i∈ σ (R_j)。令I={I₁, I₂, …, I_s}是定义在R上的一组随机变量, 则称I为一个随机场, I_i是与R_i有关的随机变量。如果对于所有I, 有p(I)> 0且p(I_i|I_q, q≠ i)=p[I_i|I_q, q∈ σ (R_i)], 则称I为在G上相对于近邻系统σ 的MRF。MRF能很好地描述随机变量在空间上的依赖关系。将像斑看作结点, 像斑所属类别看作随机变量, 像斑的类别受周围邻近像斑类别的影响, 具有随机性。假设像斑类别仅受直接相邻像斑类别的影响, 则像斑类别的空间依赖关系可用MRF来描述, 即

p(C_j)= $\begin{matrix} \overset{n}{\underset{k = 1}{Π}} \end{matrix}$ p{C_j(k)|[C_j(i), i≠ k]}= $\begin{matrix} \overset{n}{\underset{k = 1}{Π}} \end{matrix}$ p{C_j(k)|[C_j(i), i∈ σ (R_k)]}= $\begin{matrix} \overset{n}{\underset{k = 1}{Π}} \end{matrix} \begin{matrix} \frac{1}{Z} \end{matrix}$ exp{-U[C_j(k)|C_j(i), i∈ σ (R_k)]}, (5)

式中: p(C_j)为第j次分类结果的概率; p[C_j(k)]为第k个像斑在第j次分类结果中地物类别为C_j(k)的概率; Z为归一化常数; U(.)为吉布斯能量函数, 计算公式为

U[C_j(k)|C_j(i), i∈ σ (R_k)]= $\begin{matrix} \sum_{i \in σ (R_{k})} \end{matrix}$ β δ [C_j(k), C_j(i)], (6)

式中: β 为常量, 用来调节空间关系所占权重; δ 函数的计算方法为

δ [C_j(k), C_j(i)]=-p[C_j(k)|C_j(i)], (7)

式中p[C_j(K)|C_j(i)]表示在C_j(i)出现的条件下, 出现C_j(k)的概率。

式(4)最后可以化简为

C_t=arg $\begin{matrix} \max_{C_{j} \in C} \end{matrix} \begin{matrix} \overset{n}{\underset{k = 1}{Π}} \end{matrix}$ { $\begin{matrix} \frac{1}{Z} \end{matrix}$ exp[ $\begin{matrix} \sum_{i \in σ (R_{k})} \end{matrix}$ β p(C_j(k)|C_j(i))]p[X_k|C_j(k)]}, (8)

对式(8)取对数并且略去常量, 得到

C_t=arg $\begin{matrix} \max_{C_{j} \in C} \end{matrix} \begin{matrix} \overset{n}{\sum_{k = 1}} \end{matrix}$ {ln p[X_k|C_j(k)]+ $\begin{matrix} \sum_{i \in σ (R_{k})} \end{matrix}$ β p[C_j(k)|C_j(i)]}。 (9)

1.2 类别邻接矩阵(CAM)

CAM可描述各地类间的相邻关系。若原始图像包含的地类数为m, 地物类别集合为ω ={ω ₁, ω ₂, …, ω _m}, 则CAM为m× m的矩阵T, 即

$T = \begin{matrix} [\begin{matrix} p (ω_{1} | ω_{1}) & \dots & p (ω_{j} | ω_{1}) & \dots & p (ω_{m} | ω_{1}) \\ \dots & \dots & \dots \\ p (ω_{1} | ω_{i}) & \dots & p (ω_{j} | ω_{i}) & \dots & p (ω_{m} | ω_{i}) \\ \dots & \dots & \dots \\ p (ω_{1} | ω_{m}) & \dots & p (ω_{j} | ω_{m}) & \dots & p (ω_{m} | ω_{m}) \end{matrix}] \end{matrix}$ , (10)

式中p(ω _j|ω _i)为矩阵中第i行第j列的元素, 即在当前地物类别为ω _i的条件下, 邻接地物类别为ω _j的概率。

图像分割后得到若干像斑, 每个像斑分别对应着一种地物类别, 遍历所有像斑, 可以得到矩阵T, 计算公式为

p(ω _j|ω _i)=f(ω _i, ω _j)/f(ω _i), (11)

式中: f(ω _i, ω _j)为像斑及其邻接像斑类别依次为ω _i和ω _j的像斑对数; f(ω _i)为所有类别为ω _i的像斑的邻接像斑总数。

获取矩阵T需要已知所有像斑的地物类别, 而本文中基于像斑空间关系的分类方法需要已知矩阵T, 因此需要采用迭代的方法获取矩阵T。本文采用基于光谱信息的最大似然分类法对所有像斑进行初始分类, 在此基础上获取矩阵T的初始值。

1.3 模型求解

在遥感图像上选取m类训练像斑样本, 假设地物类别特征向量符合高斯分布, 利用最大似然法估计各类别对应的高斯分布的参数— — 均值向量和协方差矩阵。依据各个像斑对应类别的高斯分布, 就可以解出p[X_k|C_j(k)]。

由于式(9)的直接求解十分复杂, 本文采用Besag^[8]提出的ICM(iterated conditional mode)算法进行求解。该方法是一种条件迭代算法, 可能会陷入局部最优, 而无法获取全局最优的结果; 然而该算法计算简单, 执行效率高, 在初始值较为准确的情况下, 仍可获得较为精确的估计结果。

本文采用条件迭代算法可兼顾精度和效率。依据所有像斑的特征向量及除当前像斑之外的所有像斑的类别, 来更新当前像斑的类别。算法具体实现步骤为: ①选取适当像斑样本, 采用最大似然法进行图像分类, 获取各个像斑的初始地物类别; ②遍历所有像斑, 通过统计得到类别邻接矩阵T; ③逐个鉴别像斑, 并且更新像斑的类别, 使得式(9)的值最大; ④迭代执行步骤②和③, 直到满足收敛条件为止。

2 算法流程

首先需要通过图像分割获取像斑; 然后选取训练像斑样本, 利用最大似然法进行图像初始分类; 再遍历所有像斑, 获取类别邻接矩阵; 基于类别邻接矩阵更新各个像斑的分类结果, 采用条件迭代的方法进行分类, 直到满足迭代终止条件(图1)。

	Figure Option View Download New Window
	图1 图像分类流程图Fig.1 Flow chart for image classification

3 实验及分析

实验数据采用武汉南湖地区2005年获取的QuickBird图像, 包含蓝、绿、红、近红外4个波段, 按R(3), G(2), B(1)组合方案合成模拟真彩色图像, 经过重采样后, 图像分辨率为2 m(图2(a))。采用文献[9]介绍的快速分水岭变换(watershed transform, WT)算法进行图像分割^[10], α 和coef参数值分别为0.5和0.8, 伪标记剔除阈值为80, 图像分割结果共获取819个像斑(图2(b))。

	Figure Option View Download New Window
	图2 原始数据及图像分割结果Fig.2 Orignal data and image segmentation results

2005年武汉城市化建设速度较快, 大量湖泊被改造为裸地和居民地。图像中主要包含5类地物: 湖泊、裸地、居民地、农田和道路。由于地物光谱的复杂性, 即使同一类地物也会有着不同的光谱特征(例如湖泊的光谱特征就会受到泥沙及叶绿素含量的影响)。同一类地物中具有不同光谱特征的地物称为子类, 为了提高分类精度, 样本的选取尽可能覆盖了地物包含的所有子类。本文依据叶绿素和含沙量的不同将湖泊分为5个子类, 居民地分为白色屋顶和黄色屋顶2个子类, 道路分为水泥路和柏油路2个子类。

采用快速分水岭变换算法对图像分割后获取了若干像斑, 以这些像斑为基础进行图像分类。为了评价基于以上像斑的最大似然分类法(ML)和本文方法的精度, 依据目视解译和实地调查相结合的方法制作了标准的图像分类图(图3)。

	Figure Option View Download New Window
	图3 图像分类结果Fig.3 Results of image classification

最大似然法和本文方法的用户总体精度分别为70.0%和83.0%, 其分类结果混淆矩阵见表1。

表1 最大似然法和本文方法分类结果混淆矩阵 Tab.1 Confusion matrix of classification by ML and method in the article(像元)

将上述2种方法对5种地物的分类精度进行比较^[11], 并绘制对比图(图4)。

	Figure Option View Download New Window
	图4 分类精度比较Fig.4 Comparison of classification accuracy

分析表1及图4可以看出, 本文方法中5种地物类别的用户精度和制图精度均得到明显提高, 整体分类的Kappa系数增加了0.167。其中, 居民地和道路的分类精度得到较大改善。这是由于图像中居民地和道路反射率较高, 有着较大的灰度值, 光谱特征相似, 基于光谱特征的分类器无法正确区分二者, 大量道路被误分为居民地; 而在空间上, 居民地和居民地相邻的概率大于居民地和道路相邻的概率, 因此利用本文方法引入空间关系后较好地改善了道路和居民地的分类结果。

本文算法经过若干次迭代之后, 得到收敛的地物类别邻接矩阵(表2)。

表2 类别邻接矩阵 Tab.2 Class adjacency matrix

从表2对角线上元素可以看出, 5种地物类别与同类型地物相邻的概率大于与其他类型地物相邻的概率。农田与农田相邻的概率同农田与湖泊相邻的概率相差不大, 这是因为南湖地区湖泊和农田交织在一起, 农田附近出现农田和湖泊的机会均等。

4 结论

1)本文提出了一种基于像斑空间关系的遥感图像分类方法。通过最大似然法获取初始分类结果, 统计得到地物类别邻接矩阵; 依据该矩阵对图像分类结果进行修正; 再迭代此过程, 直到图像分类结果达到稳定。

2)该方法充分挖掘了像斑之间的空间关系, 提高了分类精度, 实验证明该方法是有效的。

3)图像分割的最优尺度问题以及像斑纹理特征的引入, 将是下一阶段研究的重点。

The authors have declared that no competing interests exist.

参考文献

View Option

[1]	张乐飞, 黄昕, 张良培. 高分辨率遥感图像的支持向量机分类方法[J]. 武汉大学学报: 信息科学版, 2012, 37(3): 314-317. Zhang L F, Huang X, Zhang L P. Classification of high spatial resolution imagery using support tensor machine[J]. Geomatics and Information Science of Wuhan University, 2012, 37(3): 314-317. [本文引用:1] [CJCR: 0.797]
[2]	赵红蕊, 阎广建, 邓小炼, 等. 一种简单加入空间关系的实用图像分类方法[J]. 遥感学报, 2003, 7(5): 358-363. Zhao H R, Yan G J, Deng X L, et al. A classification method based on spatial information[J]. Journal of Remote Sensing, 2003, 7(5): 358-363. [本文引用:1] [CJCR: 1.077]
[3]	蔡晓斌, 陈晓玲, 王涛, 等. 基于图斑空间关系的遥感专家分类方法研究[J]. 武汉大学学报: 信息科学版, 2006, 31(4): 321-324. Cai X B, Chen X L, Wang T, et al. Remote sensing export classification method based on patch spatial relationship[J]. Geomatics and Information Science of Wuhan University, 2006, 31(4): 321-324. [本文引用:1] [CJCR: 0.797]
[4]	张路, 廖明生. 一种顾及上下文的遥感图像模糊聚类[J]. 遥感学报, 2006, 10(1): 58-65. Zhang L, Liao M S. Contextual fuzzy clustering of remote sensing imagery[J]. Journal of Remote Sensing, 2006, 10(1): 58-65. [本文引用:1] [CJCR: 1.077]
[5]	乔程, 沈占锋, 吴宁, 等. 空间邻接支持下的遥感图像分类[J]. 遥感学报, 2011, 15(1): 88-99. Qiao C, Shen Z F, Wu N, et al. Remote sensing image classification method supported by spatial adjacency[J]. Journal of Remote Sensing, 2011, 15(1): 88-99. [本文引用:1] [CJCR: 1.077]
[6]	Yu Q, Gong P. Object-based detailed vegetation classification with airborne high spatial resolution remote sensing imagery[J]. Photogrammetric Engineering and Remote Sensing, 2006, 72(7): 799-811. [本文引用:1] [JCR: 2.071]
[7]	杜凤兰, 田庆之, 夏学齐, 等. 面向对象的地物分类法分析与评价[J]. 遥感技术与应用, 2004, 19(1): 20-23. Du F L, Tian Q Z, Xia X Q, et al. Object-oriented image classification analysis and evaluation[J]. Remote Sensing Technology and Application, 2004, 19(1): 20-23. [本文引用:1] [CJCR: 1.047]
[8]	Besag J. On the statistical analysis of dirty pictures[J]. Royal J Statist Soc B, 1986, 48: 259-302. [本文引用:1] [JCR: 5.721]
[9]	巫兆聪, 胡忠文, 欧阳群东. 一种区域自适应的遥感图像分水岭分割算法[J]. 武汉大学学报: 信息科学版, 2011, 36(3): 293-296. Wu Z C, Hu Z W, Ouyang Q D. A regional adaptive segmentation algorithm for remote sensing image[J]. Geomatics and Information Science of Wuhan University, 2011, 36(3): 293-296. [本文引用:1] [CJCR: 0.797]
[10]	Meyer F. Color image segmentation[C]//IEEE international conference on image processing and it’s applications, Maastricht, Netherland s: IEEE, 1992: 303-306. [本文引用:1]
[11]	洪志刚, 丛楠, 阎利, 等. 不同时相遥感图像的土地利用分类精度分析[J]. 测绘科学, 2012, 37(1): 112-114. Hong Z G, Cong L, Yan L, et al. Analysis on classified precision of land use using remote sensing images in different temporal[J]. Science of Surveying and Mapping, 2012, 37(1): 112-114. [本文引用:1] [CJCR: 0.603]

2012

0.0

0.797

张乐飞, 黄昕, 张良培. 高分辨率遥感图像的支持向量机分类方法[J]. 武汉大学学报: 信息科学版, 2012, 37(3): 314-317.

Zhang L

, Huang

, Zhang L

Classification of high spatial resolution imagery using support tensor machine[J]. Geomatics and Information Science of Wuhan University, 2012, 37(3): 314-317.

We propose a support tensor machine for remote sensing image classification.The training samples are represented as 3-order tensors with local neighbor information.Then the mathematical model and solution of support tensor machine are discussed in detail.A range of experiments demonstrate that the effectiveness of the proposed method can deliver a high classification rate with a small number of training samples.

1 State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing, Wuhan University,129 Luoyu Road,Wuhan 430079,China

针对高分辨率遥感数据分类多特征、小样本的特点,将训练样本像素邻域的数据立方以三阶张量表征,并提出了利用支持张量机对训练样本进行监督分类的模型和解法。实验结果表明,此方法能够利用少量的训练样本实现更优的分类精度。

... 前者分类结果较准确^[1],但需要合适的训练样本,而高质量的训练样本一般较难获取 ...

2003

0.0

1.077

赵红蕊, 阎广建, 邓小炼, 等. 一种简单加入空间关系的实用图像分类方法[J]. 遥感学报, 2003, 7(5): 358-363.

Zhao H

, Yan G

, Deng X

, et al. A classification method based on spatial information[J]. Journal of Remote Sensing, 2003, 7(5): 358-363.

Remote sensing image classification is one of the basic components in remote sensing image processing, which is the key stage in remote sensing applications. There are two ways to get a higher precision: one is to make full use of spectral information, and another is to add other new information The later is called multi-information processing. Spatial information is one of the most important information in these kinds of multi-information. Since lack of information is always the key problem in remote sensing image processing, we propose a spatial information based method to cope with this problem. What we used is not the spatial data but the spatial characteristics. The main advantage of our method is that it can work without any other data sources. We take advantage of spatial autocorrelation law, which is the first law of geography. The law is widely used in geo-analyses. It denotes that the closer in spatial, the similar the objects or phenomenon is. Two spatial bands are constructed based on spatial autocorrelation law and a spatial classification result is produced. These two additional bands containing spatial location information are used in classification. One real class may be divided into many classes that are not similar in space. But every class has similar spectral and spatial characteristics. Based on the classification results, further processing should be done. First, we replace every pixel value of every spectral band with the average spectral value of its class. Then a traditional classification method, such as ISODATA method, can be used. With these steps, we get the last classification results. The MAS and ASTER images are used to prove our method. The classification results show that some miss classified ground covers in the general spectrum based classification method are thrown off by our method. In other words, our method is resistant to the problem of having apparent same spectrum but different ground cover. Another distinct advantage of our method is about the removal of salt and pepper noise from images. It may be also helpful in other research area, such as landscape research. Our method also proves the importance of spatial information in remote sensing image classification. It gives an easy way to add spatial constraint to classification process. The new classification method reveals its simpleness and practicality.

遥感图像分类是遥感图像处理的一项基本内容，也是遥感应用中关键的一步。为了提高分类的精度，一方面是对光谱信息的合理利用；另一方面，可以加入新的信息源，即进行多源数据处理，并加入地学知识，尤其是对空间信息的利用是至关重要的。但是由于地学知识的复杂性及空间信息利用的难度以及数据源的限制，尚无公认的实用方法。该文提出了一种简单加入空间关系的分类方法，在没有其它数据源的情况下，利用空间关系特性，在分类中构造两个空间关系波段，实现空间约束，部分消除仅依赖光谱数据分类而引起的同物异谱和同谱异物造成的分类错误。简单实用，同时也验证了空间关系在分类中的重要性。

... 赵红蕊等^[2]提出了简单加入空间关系的实用图像分类方法,构造了2个空间关系波段,实现了空间约束 ...

2006

0.0

0.797

蔡晓斌, 陈晓玲, 王涛, 等. 基于图斑空间关系的遥感专家分类方法研究[J]. 武汉大学学报: 信息科学版, 2006, 31(4): 321-324.

Cai X

, Chen X

, Wang

, et al. Remote sensing export classification method based on patch spatial relationship[J]. Geomatics and Information Science of Wuhan University, 2006, 31(4): 321-324.

The homogeneous spectral units' patches,were taken as the classification unit,and the neighboring-searching algorithm was employed to obtain the relationship of different spatial features,which results in higher classification accuracy of wetland,grassland and farmland.The extractions of water body were also discussed and the comparisons were made(among) different extraction methods.

1 State Key Laboratory for Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,129 Luoyu Road,Wuhan 430079,China; 2 State Key Laboratory of Water Resource and Hydropower Engineering Science,Wuhan University,8 South Donghu Road,Wuhan 430072,China

以光谱均质单元———图斑作为空间对象粒度的基本单元,采用图斑邻域搜索算法,获取不同类别图斑间的空间关系。在ERDAS专家分类系统中,根据图斑相邻关系以及DEM信息对初始分类结果进行了修正,提高了湿地、草地和农业用地的分类精度。

... 蔡晓斌等^[3]提出了基于图斑空间关系的遥感图像分类方法,加入了空间对象的空间推理知识 ...

2006

0.0

1.077

张路, 廖明生. 一种顾及上下文的遥感图像模糊聚类[J]. 遥感学报, 2006, 10(1): 58-65.

Zhang

, Liao M

Contextual fuzzy clustering of remote sensing imagery[J]. Journal of Remote Sensing, 2006, 10(1): 58-65.

Fuzzy clustering is an important method in unsupervised classification. In application of traditional fuzzy clustering algorithm to unsupervised classification of remote sensing imagery, pixels are assumed to be independent of each other and their fuzzy memberships are determined individually, so that context information, i.e. statistical dependencies among neighboring pixels, are not taken into account. Aiming at this problem, an improved fuzzy clustering algorithm considering context information is put forward by incorporating the concept of spatial fuzzy membership under MRF framework. In this way, accuracy and reliability of clustering can be improved upon traditional ones.\r\n To evaluate the quality of clustering results, a validation index considering both intra-cluster compactness and inter-cluster separation is introduced, further more it is employed to find out naturally optimal cluster numbers and promote objectivity of clustering results. Finally an experiment on real remote sensing imagery is carried out to demonstrate the effectiveness of our proposed scheme.

模糊聚类是非监督分类中的一类重要方法。传统的模糊聚类方法应用于遥感影像的非监督分类时，均未考虑到邻域像元间的统计依赖关系即上下文信息。针对这一缺陷，在Markov随机场模型框架下，引入了空间隶属度概念，提出了一种顾及上下文信息的模糊聚类算法，有效地提高了聚类精度和抗噪声能力。针对需要预先指定聚类个数的问题，采用了一种兼顾类别内部紧密程度和类别之间分离程度的评价指标，用以检验聚类结果的有效性。从而找出最优的聚类个数，在一定程度上提高了聚类结果的客观性。最后通过实验验证了本文算法的有效性。

... 张路等^[4]提出了顾及上下文的遥感图像模糊聚类方法,引入了空间隶属度的概念 ...

2011

0.0

1.077

乔程, 沈占锋, 吴宁, 等. 空间邻接支持下的遥感图像分类[J]. 遥感学报, 2011, 15(1): 88-99.

Qiao

, Shen Z

, Wu

, et al. Remote sensing image classification method supported by spatial adjacency[J]. Journal of Remote Sensing, 2011, 15(1): 88-99.

The image classification is a key step for remote sensing data transforming into practical information andknowledge, which has always been the core problem in the remote sensing field. The limitations in traditional spectralclassification method otherwise promotes the theory development on the spatial-spectral coupled information cognition ofremote sensing, which focuses more on the spatial relationship. However, the current classification revision methods haveconfigured the spatial forms and relationship while, going further, but there still exist some deficiencies in spatial distributiontheorem about quantitative description, objects’ actual distribution, and so on. Thus, the paper proposes a spatial-adjacencysupportedclassification revision method inclusive of reference object extraction, target object pixels searching and referenceadjacent objects distinguishing which detailed steps are: (1) marking the objects out and getting their distribution rangepickingup the other objects in the range, (2) selecting them as the target object, picking out the unavailable target object in the range andselecting them as a certain objectwhich also provides a convenient and effective way for stepwise and accurate extraction ofother objects subsequently. We also carried out an experiment on offshore area classification revision, and the result proved tobe more accurate and reasonable.

传统光谱分类法的局限性促使了遥感“图谱耦合”认知理论的发展, 使其更加注重了空间信息的应用。然而, 已有的分类方法虽也融入了空间形态、空间关系的应用, 在精度上有一定的提高, 但在空间规律定量描述、地物实际分布边界跟踪等方面仍存在不足。本文发展了一种空间邻接支持下的遥感影像分类方法: 通过基准地物的精确提取进而搜索与其邻接的目标地物, 对邻接范围内的地类混淆以及非邻接范围内的目标类误分一并进行修正,并以近海地物分类为例进行试验, 获得了更为精确、合理的分类结果, 也为后续逐步精确地提取各地物提供了一种便捷有效的途径。

... 乔程等^[5]提出了一种空间邻接支持下的遥感图像分类方法 ...

2006

2.071

0.0

... 本文研究了基于像斑^[6,7]空间关系的遥感图像分类方法,并在高分辨率遥感图像分类实验中取得了较好的效果 ...

2004

0.0

1.047

杜凤兰, 田庆之, 夏学齐, 等. 面向对象的地物分类法分析与评价[J]. 遥感技术与应用, 2004, 19(1): 20-23.

Du F

, Tian Q

, Xia X

, et al. Object-oriented image classification analysis and evaluation[J]. Remote Sensing Technology and Application, 2004, 19(1): 20-23.

High-resolution remote sensing data have a lot of characteristics compared to the other remotesensing data. Object-oriented image classification can make the most of their advantages. Using thismethod, we classify IKONOS image data of Nanjing to realize accurate classification, evaluate theprecision of the classification results, and discuss the potential and restriction of this method in imageclassification so that providing new methods in the high-resolution remote sensing data classification.

(International Institute For Earth System Science,Nanjing University,Nanjing210093,China)

摘要:结合南京市区IKONOS高分辨率遥感数据,研究利用面向对象分类法实现地物精确分类的方法和技巧,评估分类结果的精度,探讨面向对象分类法在地物分类应用中的潜力和限制,为高分辨率遥感在地物分类中的有效应用提供新的技术途径。

... 本文研究了基于像斑^[6,7]空间关系的遥感图像分类方法,并在高分辨率遥感图像分类实验中取得了较好的效果 ...

1986

5.721

0.0

... 由于式(9)的直接求解十分复杂,本文采用Besag^[8]提出的ICM(iterated conditional mode)算法进行求解 ...

2011

0.0

0.797

巫兆聪, 胡忠文, 欧阳群东. 一种区域自适应的遥感图像分水岭分割算法[J]. 武汉大学学报: 信息科学版, 2011, 36(3): 293-296.

Wu Z

, Hu Z

, Ouyang Q

A regional adaptive segmentation algorithm for remote sensing image[J]. Geomatics and Information Science of Wuhan University, 2011, 36(3): 293-296.

Marker-based watershed segmentation can effectively overcome the problem of over-segmentation and performs well in most cases.However it may not be suitable for remote sensing images due to the diversity of ground objects in local and regional scale.Regions with different characteristics can not be segmented with the same criteria arbitrarily.We propose regional adaptive watershed segmentation algorithm.First,a regional adaptive threshold image is constructed using the blurred gradient image combined with an evaluated statistic threshold.Then markers are constructed by threshold segmentation of the gradient image.Segmentation result is finally generated with an marker-based watershed segmentation.The algorithm is proved to be effective through a series of experiments including the comparison with the multi-resolution segmentation algorithm used in eCognition.

1 School of Remote Sensing and Information Engineering,Wuhan University,129 Luoyu Road,Wuhan 430079,China

提出了一种区域自适应的标记分水岭分割方法。该方法利用高斯低通滤波和概率统计相结合的方法,对梯度影像进行区域自适应阈值分割,提取分割标记,然后采用Meyer算法进行标记分水岭分割。实验结果表明,该算法能够有效解决遥感影像不同区域的分割问题,达到比较理想的水平。

1992

0.0

... 采用文献[9]介绍的快速分水岭变换(watershed transform,WT)算法进行图像分割^[10],#cod#x003b1 ...

2012

0.0

0.603

... 将上述2种方法对5种地物的分类精度进行比较^[11],并绘制对比图(图4) ...