|
发布时间: 2019-04-16 |
遥感图像处理 |
|
|
收稿日期: 2018-08-27; 修回日期: 2018-10-07
基金项目: 国家自然科学基金项目(41301479);辽宁省高校创新人才支持计划项目(LR2016061);辽宁省教育厅科学技术研究一般项目(LJCL009)
第一作者简介:
李玉, 1963年生, 男, 博士, 教授, 博士生导师。主要研究方向为遥感数据处理理论与应用基础研究, 包括空间统计学、随机几何、模糊数学在遥感数据建模与分析方面的应用。E-mail:liyu@lntu.edu.cn;
石雪, 女, 博士研究生, 主要研究方向为影像信息提取理论与方法。E-mail:374636252@qq.com; 赵泉华, 女, 博士, 教授, 主要研究方向为遥感数据处理理论与应用基础研究。E-mail:276611089@qq.com.
中图法分类号: TP301.6
文献标识码: A
文章编号: 1006-8961(2019)04-0630-09
|
摘要
目的 高光谱图像波段数目巨大,导致在解译及分类过程中出现“维数灾难”的现象。针对该问题,在K-means聚类算法基础上,考虑各个波段对不同聚类的重要程度,同时顾及类间信息,提出一种基于熵加权K-means全局信息聚类的高光谱图像分类算法。方法 首先,引入波段权重,用来刻画各个波段对不同聚类的重要程度,并定义熵信息测度表达该权重。其次,为避免局部最优聚类,引入类间距离测度实现全局最优聚类。最后,将上述两类测度引入K-means聚类目标函数,通过最小化目标函数得到最优分类结果。结果 为了验证提出的高光谱图像分类方法的有效性,对Salinas高光谱图像和Pavia University高光谱图像标准图中的地物类别根据其光谱反射率差异程度进行合并,将合并后的标准图作为新的标准分类图。分别采用本文算法和传统K-means算法对Salinas高光谱图像和Pavia University高光谱图像进行实验,并定性、定量地评价和分析了实验结果。对于图像中合并后的地物类别,光谱反射率差异程度大,从视觉上看,本文算法较传统K-means算法有更好的分类结果;从分类精度看,本文算法的总精度分别为92.20%和82.96%,K-means算法的总精度分别为83.39%和67.06%,较K-means算法增长8.81%和15.9%。结论 提出一种基于熵加权K-means全局信息聚类的高光谱图像分类算法,实验结果表明,本文算法对高光谱图像中具有不同光谱反射率差异程度的各类地物目标均能取得很好的分类结果。
关键词
波段加权; 信息熵; 类间信息; K-means; 高光谱图像; 分类
Abstract
Objective Hyperspectral remote sensing has become a promising research field and is applied to various aspects. Thus, hyperspectral image classification has become the key part of hyperspectral image processing. The important trait of hyperspectral images is the excessive number of bands, which results in the phenomenon of "the curse of the dimension" in their interpretation and classification. Utilizing this band information fully in the classification of hyperspectral images is difficult. K-means algorithm is the most classical clustering algorithm, which is widely used for image classification. The general idea of K-means algorithm is to treat every feature as equally important. However, when the K-means algorithm is used for the classification of hyperspectral images, every band is regarded as a feature, which leads to the difficulty in feature utilization and poor classification results. To solve this problem, the idea of feature weighting is introduced. Therefore, this study proposes a hyperspectral image classification algorithm based on entropy weighted K-means by considering global information. Method The proposed hyperspectral image classification method is based on the K-means clustering algorithm and considers them to indicate the importance of every band to different clusters and the inter-cluster information. Feature weighting is used to distinguish the importance of every band to different clusters, as described by the band weight. In statistics, entropy represents the degree of uncertainty of information. Thus, entropy information measurement is defined to express the weight distribution. In hyperspectral image classification, the distance between classes greatly influences the clustering results. The distance measurement of inter-cluster information is introduced to realize the global optimal clustering to avoid the local optimal clustering and obtain more accurate results. These two types of measurements are introduced into the K-means clustering objective function, and the optimal classification results are obtained by minimizing the objective function. Result Classification experiments are conducted using the proposed algorithm and K-means algorithm on Salinas and Pavia University hyperspectral images, respectively, to verify the proposed hyperspectral image classification method effectively. The ground objects in the standard images of Salinas and Pavia University are merged on the basis of the difference in the degree of spectral reflectance, and the combined standard images are considered the standard classification information. The classification results demonstrate that the proposed algorithm can effectively obtain better results than K-means algorithm. The overall accuracy and Kappa coefficient are calculated from a confusion matrix and compared with K-means algorithm to evaluate the proposed algorithm quantitatively. The overall accuracy of the algorithm is 92.20% and 82.96%, indicating 8.81% and 15.9% improvement compared with the K-means algorithm. The accuracy values demonstrate that the proposed algorithm can achieve more precise classification results than K-means algorithm. Conclusion This study proposes a hyperspectral image classification method that connects the traditional K-means algorithm with the idea of feature weighting and the inter-cluster information. Experimental results show that this approach is promising and effective and can achieve excellent classification results for all types of ground objects in hyperspectral images with large spectral reflectance differences. In future research, the similarity between features and the spatial information must be improved.
Key words
feature weighting; entropy information; inter-cluster information; K-means; hyperspectral image; classification
0 引言
目前,高光谱遥感[1]已成为一个极具前景的研究领域,广泛应用于植被调查[2]、农作物捡测[3]和海洋环境监测[4]等各个方面。高光谱图像分类[5]是高光谱图像处理的关键部分,分类结果的精度对后续的实际应用起着决定性作用。高光谱图像的最大特点是波段数目巨大,如应用较多的AVIRIS高光谱图像有224个波段,甚至有些高光谱图像的波段数可达上千个。如何充分合理地运用这些波段信息是高光谱图像分类的难点。
目前,高光谱图像分类方法众多,其中最为有效的方法是基于统计的分类方法[6-9]。Shabna等人[10]提出结合层次集图像分割和主成分分析的高光谱图像分类方法。该方法先对图像域进行分块,提取颜色、强度等不同特征;再对图像进行主成分变换降维,选取第一主成分进行后续操作;然后提取边缘,最后合并图像块完成分类。虽然在效率和精度上有一定优势,但在降维操作中由于只选用第一主成分,所以无法充分运用各波段信息。曹建农等人[11]提出基于光谱角(SA)与马尔可夫网(DMN)的高光谱图像分割方法。该方法将SA与DMN相结合,用光谱角余弦来度量DMN,并以DMN信息对高光谱图像进行分割。虽然充分运用各波段信息,并将空间特征与光谱特征相结合,但是同等看待各波段特征,易产生过分割现象。在统计方法中,基于聚类的分类方法是实现图像分类最为常见的方式。Selim等人[12]提出的
本文根据高光谱的特点,利用特征加权的思想,提出熵加权
1 算法描述
1.1 模型建立
设高光谱图像
在基于
$ {J_1}\left( {\mathit{\boldsymbol{W}},\mathit{\boldsymbol{Z}}} \right) = \sum\limits_{j = 1}^n {\sum\limits_{l = 1}^k {{\omega _{jl}}\sum\limits_{i = 1}^m {{{\left( {{z_{li}} - {x_{ji}}} \right)}^2}} } } $ | (1) |
式中,
$ \sum\limits_{l = 1}^k {{\omega _{jl}}} = 1 $ | (2) |
式中,
$ \sum\limits_{i = 1}^m {{\lambda _{il}}} = 1 $ | (3) |
目标函数可定义为
$ {J_2}\left( {\mathit{\boldsymbol{W}},\mathit{\boldsymbol{Z}},\mathit{\boldsymbol{ \boldsymbol{\varLambda} }}} \right) = \sum\limits_{j = 1}^n {\sum\limits_{l = 1}^k {{\omega _{jl}}\sum\limits_{i = 1}^m {\lambda _{il}^\beta {{\left( {{z_{li}} - {x_{ji}}} \right)}^2}} } } $ | (4) |
式中,
$ \begin{array}{*{20}{c}} {{J_3}\left( {\mathit{\boldsymbol{W}},\mathit{\boldsymbol{Z}},\mathit{\boldsymbol{ \boldsymbol{\varLambda} }}} \right) = \sum\limits_{j = 1}^n {\sum\limits_{l = 1}^k {{\omega _{jl}}\sum\limits_{i = 1}^m {{\lambda _{il}}{{\left( {{z_{li}} - {x_{ji}}} \right)}^2}} } } + }\\ {\sum\limits_{l = 1}^k {{\gamma _l}} \sum\limits_{i = 1}^m {{\lambda _{il}}\ln {\lambda _{il}}} } \end{array} $ | (5) |
式中,等式右边第1项是加权类内非相似性测度总和,第2项是正则化项,表示权重大小的不确定程度。
但是目标函数式(5)仅考虑类内距离,在高光谱图像分类中,类间距离对聚类结果亦有很大影响。为了使聚类结果更加准确,本文在目标函数式(5)的基础上引入类间距离,得到熵加权
$ \begin{array}{l} J\left( {\mathit{\boldsymbol{W}},\mathit{\boldsymbol{Z}},\mathit{\boldsymbol{ \boldsymbol{\varLambda} }}} \right) = \sum\limits_{j = 1}^n {\sum\limits_{l = 1}^k {{\omega _{jl}}\sum\limits_{i = 1}^m {{\lambda _{il}}{{\left( {{z_{li}} - {x_{ji}}} \right)}^2}} } } + \\ \sum\limits_{l = 1}^k {{\gamma _l}} \sum\limits_{i = 1}^m {{\lambda _{il}}\ln {\lambda _{il}}} - \eta \sum\limits_{j = 1}^n {\sum\limits_{l = 1}^k {{\omega _{jl}}\sum\limits_{i = 1}^m {{\lambda _{il}}{{\left( {{z_{li}} - {z_i}} \right)}^2}} } } \end{array} $ | (6) |
式中,等式右边第3项是加权类间非相似性测度总和,
$ \mathit{\boldsymbol{\bar z}} = \frac{1}{k}\sum\limits_{l = 1}^k {{\mathit{\boldsymbol{z}}_l}} $ | (7) |
为了使每个聚类中各波段的权重值差异较大,即
$ {\gamma _l} = \max \left( {{D_{ls}}} \right)/4 $ | (8) |
式中
$ {D_{ls}} = \sum\limits_{j = 1}^n {{\omega _{jl}}\left[ {{{\left( {{z_{ls}} - {x_{js}}} \right)}^2} - \eta {{\left( {{z_{ls}} - {z_s}} \right)}^2}} \right]} $ | (9) |
式中,
$ \eta = \frac{{\sum\limits_{l = 1}^k {\sum\limits_{i = 1}^m {{{\left( {{z_{li}} - {z_i}} \right)}^2}} } }}{{\sum\limits_{j = 1}^n {\sum\limits_{l = 1}^k {{\omega _{jl}}\sum\limits_{i = 1}^m {{\lambda _{il}}{{\left( {{z_{li}} - {x_{ji}}} \right)}^2}} } } }} $ | (10) |
式中,分母表示类间分离度,分子表示类内紧密度。
1.2 模型求解
为了获得最优分割,局部优化
$ \left( {\mathit{\boldsymbol{\hat W}},\mathit{\boldsymbol{\hat Z}},\mathit{\boldsymbol{ \boldsymbol{\hat \varLambda} }}} \right) = \arg \min \left\{ {\mathit{\boldsymbol{W}},\mathit{\boldsymbol{Z}},\mathit{\boldsymbol{ \boldsymbol{\varLambda} }}} \right\} $ | (11) |
模型求解过程为:
1) 求解
$ {\omega _{jl}} = \left\{ \begin{array}{l} 1\;\;\;\;\;\sum\limits_{i = 1}^m {{\lambda _{jl}}\left[ {{{\left( {{z_{li}} - {x_{ji}}} \right)}^2} - \eta {{\left( {{z_{li}} - {z_i}} \right)}^2}} \right]} \le \\ \;\;\;\;\;\;\;\;\;\;\;\sum\limits_{i = 1}^m {{\lambda _{il}}\left[ {{{\left( {{z_{ri}} - {x_{ji}}} \right)}^2} - \eta {{\left( {{z_{ri}} - {z_i}} \right)}^2}} \right]} \\ 0\;\;\;\;\;其他 \end{array} \right. $ | (12) |
2) 求解
$ \frac{{\partial J}}{{\partial {z_{li}}}} = \sum\limits_{j = 1}^n {{\omega _{jl}}{\lambda _{jl}}\left[ {2\left( {{z_{li}} - {x_{ji}}} \right) - 2\eta \left( {{z_{li}} - {z_i}} \right)} \right]} = 0 $ | (13) |
求得
$ {z_{li}} = \frac{{\sum\limits_{j = 1}^n {{\omega _{jl}}\left( {{x_{ji}} - \eta {z_i}} \right)} }}{{\sum\limits_{j = 1}^n {{\omega _{jl}}\left( {1 - \eta } \right)} }} $ | (14) |
3) 求解
$ \begin{array}{*{20}{c}} {L = \sum\limits_{j = 1}^n {\sum\limits_{l = 1}^k {{\omega _{jl}}\sum\limits_{i = 1}^m {{\lambda _{il}}\left[ {{{\left( {{z_{li}} - {x_{ji}}} \right)}^2} - \eta {{\left( {{z_{li}} - {z_i}} \right)}^2}} \right]} } } + }\\ {\sum\limits_{l = 1}^k {{\gamma _l}} \sum\limits_{i = 1}^m {{\lambda _{il}}\ln {\lambda _{il}}} + \sum\limits_{l = 1}^k {{\mu _l}\left( {\sum\limits_{i = 1}^m {{\lambda _{il}}} - 1} \right)} } \end{array} $ | (15) |
对式(15)中的
$ {\lambda _{il}} = \frac{{\exp \left( { - \frac{{{D_{li}}}}{{{\gamma _l}}}} \right)}}{{\sum\limits_{s = 1}^m {\exp \left( { - \frac{{{D_{li}}}}{{{\gamma _l}}}} \right)} }} $ | (16) |
综上所述,本文算法的具体实现流程如下:
1) 给定聚类数目
2) 根据式(12)计算隶属度矩阵
3) 利用式(8)(9)(10)计算参数
4) 利用式(14)计算聚类中心矩阵
5) 根据式(16)计算权重矩阵
6) 如果
2 实验结果和讨论
为了验证本文算法的可行性和有效性,在CPU为Intel(R) Core(TM) i7-4790,3.60 GHz, 内存为4 GB的PC机上,使用MATLAB R2014b对高光谱图像进行了分类算法的实验。实验选用了两幅高光谱遥感图像,如图 1所示,其中图 1(a)为美国加利福尼亚州Salinas的AVIRIS数据的假彩色图像,图 1(b)为意大利北部Pavia University的ROSIS数据的假彩色图像。Salinas高光谱图像的光谱测量范围为400~2 500 nm,空间分辨率为3.7 m,共224个波段,去除20个水吸收波段,共有204个波段参与实验,原始图像大小为512×217像素,用波段160、95、50分别表示红、绿、蓝3个波段形成的假彩色图像作为显示图像。Pavia University高光谱图像的空间分辨率为1.3 m,保留103个波段进行实验,原始图像大小为610×340像素,用波段90、60、20分别表示红、绿、蓝3个波段形成的假彩色图像作为显示图像。Salinas和Pavia University图像对应的标准分类图仅仅对其部分区域标识了类属,如,Salinas和Pavia University图像分别将全部111 104和207 400个像素中的54 129和42 776个像素标记为16和9个地物类。与其他以此为测试数据的算法相同,本文算法也只对两幅图像中的标记部分进行了分类,而未分类部分统称为“背景”并在分类结果图中显示为黑色。此外,由于本文方法依据光谱反射率一致性实现地物分类,并将分类标准设置得更为宽泛,所以对原始标准分类图中光谱反射曲线差异程度较小的目标进行了合并,在已知的标准分类图基础上构建了新的标准分类图。
Salinas高光谱图像中包括16类地物目标,分别为杂草—西兰花1、杂草—西兰花2、休耕地、休耕犁、光滑休耕地、农作物残茬、芹菜、未训练葡萄园1、生长葡萄园、枯萎期玉米、4星期莴苣、5星期莴苣、6星期莴苣、7星期莴苣、未训练葡萄园2、垂直篱壁式葡萄园。这些地物目标绝大部分是不同种类农作物的不同时期。对于某些不同种类农作物或者同一种类不同时期的农作物,从农作物地表覆盖程度、叶子浓密程度以及叶绿素含量角度来说,它们的表现基本相同,因此光谱反射率也基本相同,光谱反射曲线差别不大。所以将所有地物目标从叶子浓密程度以及叶绿素含量角度重新归类,得到文中Salinas高光谱图像的标准分类。在此情况下,将16类地物归并为7类,具体情况如下:将杂草—西兰花1、杂草—西兰花2归并为一类,记为第C1类;将休耕地、光滑休耕地归并为一类,记为第C2类;将4星期莴苣、5星期莴苣、生长葡萄园、枯萎期玉米归并为一类,记为第C3类;将未训练葡萄园1、未训练葡萄园2、垂直篱壁式葡萄园、6星期莴苣、7星期莴苣归并为一类,记为第C4类;将休耕犁、农作物残茬、芹菜分别独自归为一类,分别记为第C5、C6和C7类。
图 2为Salinas高光谱图像的分类结果,其中,图 2(a)为Salinas高光谱数据标准分类图,图 2(b)为Salinas高光谱图像
从视觉上来看,
为了定量描述Salinas高光谱图像的分类结果,通过对
表 1
Salinas图像分类结果(用户精度、产品精度、总精度和Kappa系数)
Table 1
User accuracy, product accuracy, overall accuracy and Kappa coefficient for classification of Salinas image
方法 | 精度指标 | 地物类别 | ||||||
C1 | C2 | C3 | C4 | C5 | C6 | C7 | ||
本文算法 | 用户精度/% | 100 | 87.97 | 81.83 | 93.31 | 93.39 | 99.43 | 76.99 |
产品精度/% | 94.39 | 81.31 | 90.93 | 92.01 | 96.92 | 99.12 | 99.30 | |
总精度/% | 92.20 | |||||||
Kappa | 0.90 | |||||||
用户精度/% | 59.88 | 89.91 | 92.42 | 99.33 | 93.44 | 99.97 | 0.06 | |
产品精度/% | 99.72 | 81.61 | 89.43 | 86.64 | 97.06 | 95.55 | 0.06 | |
总精度/% | 83.39 | |||||||
Kappa | 0.78 |
从表 1可以看出,本文算法的精度基本都在80 %以上,精度分布相对稳定,其中只有C7类的用户精度未达到80 %,其余大多数都在90 %以上,总精度达到92 %以上,Kappa系数达到89 %以上。而
Pavia University高光谱图像中包括9类地物目标,分别为柏油、草地、砾石、树木、涂覆金属板、裸土、沥青、砖、阴影。因为这些地物目标中有些类别的构成材料或者成分基本相同,导致有些目标的光谱反射曲线差别不大,所以将所有地物目标从构成材料及组成成分角度重新归类,得到文中Pavia University高光谱图像的标准分类。在此情况下,将原来的9类地物目标归并为6类。具体情况如下:将柏油和沥青归并为一类,记为C1类;将砾石和砖归并为一类,记为C2类;将草地和裸土归并为一类,记为C3类;将树木、涂覆金属板、阴影分别独自设置为一类,分别记为C4、C5和C6类。
图 3为Pavia University高光谱图像的分类结果,其中,图 3(a)为Pavia University高光谱数据标准分类图,图 3(b)为Pavia University高光谱数据
从视觉上来看,
为了定量描述Pavia University高光谱图像的分类结果,通过对
表 2
Pavia University图像分类结果(用户精度、产品精度、总精度和Kappa系数)
Table 2
User accuracy, product accuracy, overall accuracy and Kappa coefficient for classification of Pavia University image
方法 | 精度指标 | 地物类别 | |||||
C1 | C2 | C3 | C4 | C5 | C6 | ||
本文算法 | 用户精度/% | 93.47 | 61.75 | 98.80 | 48.06 | 87.69 | 100 |
产品精度/% | 87.36 | 91.70 | 76.43 | 94.09 | 97.47 | 99.58 | |
总精度/% | 82.96 | ||||||
Kappa | 0.75 | ||||||
用户精度/% | 76.63 | 68.36 | 98.98 | 23.25 | 94.48 | 99.89 | |
产品精度/% | 94.88 | 58.97 | 54.11 | 99.08 | 68.77 | 100 | |
总精度/% | 67.06 | ||||||
Kappa | 0.56 |
从表 2可以看出,本文算法和
3 结论
本文提出一种熵加权
参考文献
-
[1] Bioucas-Dias J M, Plaza A, Camps-Valls G, et al. Hyperspectral remote sensing data analysis and future challenges[J]. IEEE Geoscience and Remote Sensing Magazine, 2013, 1(2): 6–36. [DOI:10.1109/MGRS.2013.2244672]
-
[2] Yu J W, Cheng Z Q, Zhang J S, et al. An Approach to Distinguishing Between Species of Trees and Crops Based on Hyperspectral Information[J]. Spectroscopy and Spectral Analysis, 2018, 38(12): 3890–3896. [虞佳维, 程志庆, 张劲松, 等. 高光谱信息的农林植被种类区分[J]. 光谱学与光谱分析, 2018, 38(12): 3890–3896. ] [DOI:10.3964/j.jssn.1000-0593(2018)12-3890-07]
-
[3] Zhang H, Zhao X M, Guo X, et al. Research progress on application of rice growth monitoring based on canopy hyperspectral information[J]. Jiangsu Agricultural Sciences, 2018, 46(12): 1–9. [张晗, 赵小敏, 郭熙, 等. 基于冠层高光谱信息的水稻生长监测应用研究进展[J]. 江苏农业科学, 2018, 46(12): 1–9. ] [DOI:10.15889/j.issn.1002-1302.2018.12.001]
-
[4] Zhang Q. Abnormal detection and recognition of ship oil spill targets in hyperspectral images[D]. Dalian: Dalian Maritime University, 2018. [张强.高光谱影像船舶溢油目标异常检测与识别[D].大连: 大连海事大学, 2018.] http://cdmd.cnki.com.cn/Article/CDMD-10151-1018078414.htm
-
[5] Du P J, Xia J S, Xue Z H, et al. Review of hyperspectral remote sensing image classification[J]. Journal of Remote Sensing, 2016, 20(2): 236–256. [杜培军, 夏俊士, 薛朝辉, 等. 高光谱遥感影像分类研究进展[J]. 遥感学报, 2016, 20(2): 236–256. ] [DOI:10.11834/jrs.20165022]
-
[6] Li J, Bioucas-Dias J M, Plaza A. Supervised hyperspectral image segmentation using active learning[C]//Proceedings of the 2nd Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing. Reykjavik, Iceland: IEEE, 2010: 1-4.[DOI: 10.1109/WHISPERS.2010.5594844]
-
[7] Bao R, Xue Z H, Zhang X Y, et al. Classification merged with clustering and context for hyperspectral imagery[J]. Geographics and Information Science of Wuhan University, 2017, 42(7): 890–896. [鲍蕊, 薛朝辉, 张像源, 等. 综合聚类和上下文特征的高光谱影像分类[J]. 武汉大学学报:信息科学版, 2017, 42(7): 890–896. ] [DOI:10.13203/j.whugis20150043]
-
[8] Ran Q, Yu H Y, Gao L R, et al. Superpixel and subspace projection-based support vector machines for hyperspectral image classification[J]. Journal of Image and Graphics, 2018, 23(1): 95–105. [冉琼, 于浩洋, 高连如, 等. 结合超像元和子空间投影支持向量机的高光谱图像分类[J]. 中国图象图形学报, 2018, 23(1): 95–105. ] [DOI:10.11834/jig.170201]
-
[9] Yang K M, Liu F, Sun Y Y, et al. Classification algorithm of hyperspectral imagery by harmonic analysis and spectral angle mapping[J]. Journal of Image and Graphics, 2015, 20(6): 836–844. [杨可明, 刘飞, 孙阳阳, 等. 谐波分析光谱角制图高光谱影像分类[J]. 中国图象图形学报, 2015, 20(6): 836–844. ] [DOI:10.11834/jig.20150614]
-
[10] Shabna A, Ganesan R. HSEG and PCA for hyper-spectral image classification[C]//Proceedings of 2014 International Conference on Control, Instrumentation, Communication and Computational Technologies. Kanyakumari, India: IEEE, 2014: 42-47.[DOI: 10.1109/ICCICCT.2014.6992927]
-
[11] Cao J N, Guan Z Q, Li D R. Study on the approaches for segmentation of hyper-spectral images based on DMN[J]. Journal of Remote Sensing, 2005, 9(5): 596–603. [曹建农, 关泽群, 李德仁. 基于DMN的高光谱图像分割方法研究[J]. 遥感学报, 2005, 9(5): 596–603. ] [DOI:10.11834/jrs.20050587]
-
[12] Selim S Z, Ismail M A. K-means-type algorithms:a generalized convergence theorem and characterization of local optimality[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1984, PAMI-6(1): 81–87. [DOI:10.1109/TPAMI.1984.4767478]
-
[13] Modha D S, Spangler W S. Feature weighting in K-means clustering[J]. Machine Learning, 2003, 52(3): 217–237. [DOI:10.1023/A:1024016609528]
-
[14] Chan E Y, Ching W K, Ng M K, et al. An optimization algorithm for clustering using weighted dissimilarity measures[J]. Pattern Recognition, 2004, 37(5): 943–952. [DOI:10.1016/j.patcog.2003.11.003]
-
[15] Huang J Z, Ng M K, Rong H Q, et al. Automated variable weighting in k-means type clustering[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27(5): 657–668. [DOI:10.1109/TPAMI.2005.95]
-
[16] Huang J Z, Xu J, Ng M, et al. Weighting method for feature selection in K-means[M]//Liu H, Motoda H. Computational Methods of Feature Selection. Boca Raton, FL: CRC Press, 2008: 193-209.
-
[17] Jing L P, Ng M K, Huang J Z. An entropy weighting K-means algorithm for subspace clustering of high-dimensional sparse data[J]. IEEE Transactions on Knowledge and Data Engineering, 2007, 19(8): 1026–1041. [DOI:10.1109/TKDE.2007.1048]
-
[18] Dairbekov N S, Paternain G P. Entropy production in thermostats Ⅱ[J]. Journal of Statistical Physics, 2007, 127(5): 887–914. [DOI:10.1007/s10955-007-9301-5]
-
[19] Feng D, Feng S T. The World of Entropy[M]. Beijing: Science Press, 2005: 27-239. [ 冯端, 冯少彤. 溯源探幽:熵的世界[M]. 北京: 科学出版社, 2005: 27-239.]