遥感科学与应用技术

基于Adaboost的高分遥感影像自动变化检测方法

  • 陈伟锋 , 1, 2, 3 ,
  • 毛政元 , 1, 2, 3, * ,
  • 徐伟铭 1, 2, 3 ,
  • 许锐 1, 2, 3, 4
展开
  • 1. 福州大学福建省空间信息工程研究中心,福州 350002
  • 2. 福州大学空间数据挖掘与信息共享教育部重点实验室,福州 350002
  • 3. 福州大学地理空间信息技术国家地方联合工程研究中心,福州 350002
  • 4. 福建工程学院信息科学与工程学院,福州 350118
*通讯作者:毛政元(1964-),男,博士,教授,博士生导师,主要从事时空序列分析、城市变化检测、信息化管理与信息服务研究。E-mail:

作者简介:陈伟锋(1992-),男,硕士生,主要从事遥感影像信息提取、数字图像处理研究。E-mail:

收稿日期: 2018-07-30

  网络出版日期: 2018-12-20

基金资助

国家自然科学基金项目(41701491);福建省自然科学基金面上项目(2018J01619)

Automatic Change Detection Approach for High-Resolution Remotely Sensed Images Based on Adaboost Algorithm

  • CHEN Weifeng , 1, 2, 3 ,
  • MAO Zhengyuan , 1, 2, 3, * ,
  • XU Weiming 1, 2, 3 ,
  • XU Rui 1, 2, 3, 4
Expand
  • 1. Provincial Spatial Information Engineering Research Center, Fuzhou University, Fuzhou 350002, China
  • 2. Key Laboratory of Spatial Data Mining and Information Sharing of Ministry of Education, Fuzhou University, Fuzhou 350002, China
  • 3. National Engineering Research Centre of Geospatial Space Information Technology, Fuzhou University, Fuzhou 350002, China
  • 4. School of Information Science and Engineering, Fujian University of Technology, Fuzhou 350188, China
*Corresponding author: MAO Zhengyuan, E-mail:

Received date: 2018-07-30

  Online published: 2018-12-20

Supported by

National Natural Science Foundation of China, No.41701491;Project of Science and Technology of Fujian Province, No.2018J01619.

Copyright

《地球信息科学学报》编辑部 所有

摘要

基于监督分类的高分辨率遥感影像变化检测需要大量人工标注,且单个监督分类器难以适应高分影像中复杂多样的地表变化信息提取,检测结果中“椒盐噪声”严重、变化图斑破碎。因此,本文提出一种基于Adaboost集成算法、自动标注训练样本的变化检测方法。首先利用非监督分类方法完成变化初检,接着在初检结果中进行“非等距”区间采样自动获取均匀分布的训练样本;然后以Adaboost算法为集成框架,选择决策树桩、Logistic回归和kNN作为弱分类器,构建一种混合分类器集成系统,充分挖掘和利用高分影像中的空间信息以提升分类精度和分类器泛化能力,最后利用SLIC分割算法和空间邻域信息对像元级检测结果进行空间约束滤波,进一步提升变化检测精度。为验证本文方法的有效性,选取SPOT-5和WorldView-2影像为实验数据,结果表明本文方法能有效降低训练样本人工标注成本、提高变化检测精度。

本文引用格式

陈伟锋 , 毛政元 , 徐伟铭 , 许锐 . 基于Adaboost的高分遥感影像自动变化检测方法[J]. 地球信息科学学报, 2018 , 20(12) : 1756 -1767 . DOI: 10.12082/dqxxkx.2018.180353

Abstract

Human annotation is a massive labor cost for the training sample selection process when applying any kind of supervised learning algorithm for change detection based on high-resolution remotely sensed satellite images. It is limited and unreasonable to use just one single sort of classifier generated from a supervised algorithm to extract change information of variety from the time-series images both in completeness and accuracy, let alone the inevitable salt-and-pepper noise and tiny patches falsely detected which turn out to be ubiquitous in and out of geographical entities. To tackle with problems mentioned above, a change detection approach based on a new automatic training sample annotation strategy and an improved Adaboost ensemble learning algorithm was proposed. At first, the unsupervised change detection algorithm CVA was applied to generate a low-level change detection result as referencing labels for further annotation, then the low-level result was divided into several parts with different intervals to ensure the automatic acquisition of the evenly distributed training samples with confidence. Furthermore, decision stump, logistic regression and kNN were employed as the weak classifiers to construct a hybrid multi-classifiers ensemble system with the help of the improved Adaboost algorithm, which would effectively promote the classification accuracy and generalization capacity of weak classifiers by sufficiently mining and making use of the spatial information with potential values. Finally, the SLIC segmentation algorithm was implemented in the difference image, and the segmentation border information was combined with spatial contextual information to build up a dual-filter for spatial constraint aiming at decreasing the omission rate and the false alarm rate of the detection results. To verify the validity of the proposed method, we conducted experiments using two datasets of multispectral images collected by SPOT-5 and WorldView-2. Experimental results indicated that the proposed method would significantly lower the labor costs of training sample annotation and demonstrated superiority compared with four other methods in accuracy.

1 引言

随着遥感数据获取技术的进步,相关研究与实践中越来越多地使用高分辨率遥感影像(下文简称“高分影像”)提取地表覆盖与土地利用变化信息。高分影像的光谱异质性明显大于同类中低分辨率遥感数据,配准误差、“同物异谱”和“异物同谱”现象对信息提取的干扰更为严重,传统针对中低分辨率影像提出的变化检测方法正面临新的挑战。
按照数据处理过程,现有变化检测方法可分为直接比较法、分类后比较法与直接分类法3种类型[1]。后者综合前两种方法的思想,直接构造双时相影像的特征差异影像或特征融合影像,然后利用分类算法提取变化信息。根据是否需要预先标记数据,变化检测又可分为非监督变化检测法(下文简称非监督法)和监督变化检测法(下文简称监督法)。前者简单易行、无需过多人工干预,但存在变化阈值难以自适应确定的问题;后者检测精度通常优于前者,但训练样本的获取需要大量人工干预,且单一算法泛化能力弱、不便移植。如何集成二者的相对优势并同时避免各自的不足,一直是遥感领域高度重视的研究课题,目前尚无理想的解决方案。
针对监督分类需要大量人工标注样本,自动化程度低的问题,文献[2]运用主动学习方法选择信息量较大的训练样本交互由人工标注,提升样本选择的自动化程度,但该法仍需人工设置初始样本并全程参与样本选择;文献[3]将过去人工解译的地物类别知识迁移至新影像,建立新的特征与地物关系进行自动变化检测,但历史专题数据往往不易获取;文献[4]、[5]基于像元变化可能性与变化强度成正比原则,采用一种基于变化强度的样本自动选择方法,将CVA强度图中变化强度前α的像元设置为变化样本,变化强度后α的像元设置为未变化样本,该法可完全实现变化检测训练样本自动选择,但是所选样本难以满足均匀分布的特性;文献[6]、[7]选择等距变化区间内的等量像元作为训练样本,保证自动选择的样本均匀分布,但因为选择了过多不确定性较大的样本,所选样本集中存在大量错误标注像元。
由于集成学习较单一分类器具备更强的泛化能力,因此它更适用于提取高分影像中复杂多变的地表信息,近年来许多学者利用Adaboost算法提取高分影像土地利用/覆盖信息[10-12,16],均取得了较高的分类精度。Adaboost是一种经典的集成学习算法,它采用boosting[9]策略集中关注被已训练的分类器误分类的数据,加权组合多个弱分类器输出强分类结果。Discrete Adaboost[13,14,15]是Adaboost算法的一种经典集成策略,它将每个弱分类器的输出结果分布在实数域中,使分类器表现出更强泛化能力,并能输出二分类结果[13],因此非常适用于变化检测工作。
本文在直接分类法变化检测的数据处理流程中集成监督法与非监督法的相对优势,并利用SLIC超像素分割和空间邻域信息保留地理对象的边缘和内部特征,提出一种高分影像变化检测解决方案。

2 研究方法

本研究首先利用CVA算法和阈值自适应确定方法完成初级变化检测,进而通过“非等距”区间采样的策略自动选择监督分类算法的训练样本;然 后以Adaboost算法为集成框架,选择决策树桩、 Logistic回归作为弱分类器进行同质集成,同时异质集成kNN算法构建混合分类器集成系统,以提升变化检测精度和分类器的泛化能力;最后利用SLIC算法分割差异影像获取边缘特征较完好的超像素块,结合空间邻域信息对检测结果进行空间约束,降低误检率和漏检率。具体流程如图1所示。
Fig. 1 Flowchart of the proposed change detection method

图1 本文变化检测方法流程

2.1 基于CVA的变化检测训练样本自动选择方法

本文综合现有的研究结果,提出一种新的基于CVA的样本自动选择方法。该法结合CVA法与“非等距”采样策略自动选择样本,既可满足训练样本的均匀分布,又能减少错误样本标注。
该样本选择方法先采用文献[8]的自适应阈值确定方法计算CVA强度图的变化阈值T。据此将影像划分成变化像元和未变化像元2部分,然后通过式(1)和(2)自动选择训练样本,式中CtsUts分别是变化样本和未变化样本的集合,CmeanUmean以及CSUS分别代表变化像元和未变化像元的均值及标准差,VminVmax代表像元的最小值和最大值,R为变化像元数量和未变化像元数量之比。
变化样本集由C1,C2,…,Cmm个区间内的像元构成,集合的右端点为最大像元值Vmax,每个区间的长度比上一个区间增加α1×CS/m,区间间隔采样的目的是保证所选样本均匀分布;为满足“变化强度越大越有可能是变化样本”准则,通过采样区间长度递增的形式提高标注样本的准确性,越接近Vmax的区间长度越大,所包含的像元越多,越远离Vmax的区间长度越小,所包含的像元越少。同理,未变化样本集由U1,U2,…,Unn个区间内的像元构成,左端点为最小像元值Vmin,区间增长步长为α2×US/n,越接近Vmin的区间包含的像元数越多,越远离Vmin的区间所包含的像元数越少。
C ts = α 1 0,1 m = fix C mean + CS × V max - T CS C 1 = T + CS , T + 1 × CS × 1 + α 1 m C 2 = T + 2 × CS , T + 2 × CS × 1 + α 1 m C m = T + m × CS , V max (1)
U ts = α 2 = α 1 × R n = fix U mean + US × T - V min US U 1 = T - US , T - 1 × US × 1 - α 2 n U 2 = T - 2 × US , T - 2 × US × 1 - α 2 n U n = V min , T - n × US (2)

2.2 基于Adaboost的混合分类器集成系统

决策树是Adaboost同质集成中广泛使用的弱分类器,文献[16]采用Adaboost集成CART决策树进行道路变化检测,有效抑制了道路变化检测结果中容易出现的边缘破碎和道路断裂现象。文献[17]基于Harr-like特征和GLCM纹理特征,运用Adaboost集成策略进行城市地形变化检测,结果优于BP神经网络。文献[12]基于Adabbost异质集成C4.5决策树、SVM和ANN进行土地利用信息提取,检测结果总体精度优于单分类器,但提升幅度不显著,个别地类精度退化。文献[18]指出,简单的弱分类器被Adaboost集成后效果更好,文献[12]由于选用了SVM、ANN等复杂度较高的强分类器,集成后的算法运行效率低,并存在过拟合现象。此外,现有基于Adaboost的变化检测研究主要面向中低分影像,并且未充分考虑分类器的异质集成以及基分类器的复杂度。区别于传统Discrete Adaboost的直接二类输出策略,本文提出一种改进的Adaboost算法,该方法顾及传统Adaboost的分类模糊度,将输出结果调整为正样本、负样本和待确定样本,同时将简单分类器进行同/异质集成并分层输出分类结果,单层Adaboost同质集成策略过程如下:
(1) 将训练集(N)中的每个样本赋予相同的权重 w i = 1 N , i 1,2 , , N ,构成训练样本权重向量W。训练样本的标签为yi, i 1,2 , , N
(2)迭代训练生成M个弱分类器,对于每个弱分类器 C j , j 1,2 , , M ,重复下述计算:
① 基于加权训练样本学习生成弱分类器 C j ,其分类标签向量为 D ji , i 1,2 , , N ,计算该弱分类器在加权训练数据集上的分类错误率 ε j ,其中 ε j = P D ji y i = w i ,且 ε j 0,1
② 计算弱分类器 C j 的权重 α j = 1 2 ln 1 - ε j ε j , α j R
③ 更新训练样本权重向量W并进行权重的重归一化处理。基于boosting策略的思想,若某个样本被弱分类器 C j 正确分类,则将其权重降低为 w i j + 1 = w i j × e - α j Sum ( W ) ,反之,若样本被错误分类,则将其权重增加为 w i j + 1 = w i j × e α j Sum ( W ) ,增加对错分样本的关注度。
(3)对测试样本集进行分类,对于每个测试样本,输出M个分类器对其的独立分类结果并相加构成结果累计值,然后通过 sign 函数确定该测试样本类别,计算公式为 sign j = 1 M α j × D j 。本文按下述方式处理分类结果累计值:当某个样本的结果累计值大于1时,将其判定为变化像元,累计值小于-1时判定为未变化像元,累计值落于[-1, 1]时判定为待确定像元,这样处理的目的是初步确定分类模糊度较小的样本标签,同时挑选出分类模糊度较大的样本进行后续分类处理。
本文选用逻辑斯蒂回归(LR)、决策树桩(DS)和kNN作为弱分类器,通过Adaboost集成策略,聚合同质/异质分类器构建混合集成系统。普通的线性回归需要分类数据遵守标准正态分布,而逻辑斯蒂回归根据数据特征拟合分类边界,通常用于处理二分类任务,不需要严格的限制约束,且模型简单,计算代价较低。决策树桩是集成学习最常用的一种弱分类器,它具有计算复杂度低,对中间值缺失不敏感,输出结果易于理解的特点。本文采用的训练样本选择方法可保证样本分布的均匀性,而kNN算法根据特征空间中距离待分类点最近的k个点的类别来确定自身类别,可确保集成后的混合分类器简单鲁棒。三者集成的过程如下:
(1)选择逻辑斯蒂回归为弱分类器,通过上述Discrete Adaboost集成策略迭代训练生成M个弱分类器(Ada-LR)并对影像中的每个像元进行分类,确定部分像元的标签,筛选出分类模糊度较大的待确定像元转入步骤(2)处理。考虑到算法的运行效率,采用随机梯度上升法更新回归系数进行优化。
(2)选择决策树桩为弱分类器进行Adaboost集成(Ada-DS),该步骤只对步骤(1)中无法确定的像元进行分类,处理方式类似步骤(1)(训练M个决策树桩),通过该层分类后只剩下小部分待确定像元,将它们转至步骤(3)处理。
(3)对步骤(2)剩余的待确定像元进行如下处理:获取剩余像元在步骤(1)中的分类结果累计值(Ada-LR),通过 sign 函数将累计值落于[0, 1]内的像元设置为变化像元,落于[-1, 0)内的设置为未变化像元,组合构成这些剩余像元的第一组标签向量A;同理利用步骤(2)的分类器(Ada-DS)获取剩余像元的第二组标签向量B;利用kNN算法对剩余像元进行分类,构成第三组标签向量C;最后通过多数投票法对三组标签向量进行投票确定剩余像元的最终类别。由于训练样本集的数据量较大,将k值设置在[10, 20]范围内,通过多次实验确定2个数据集的最佳k值分别为11和17。考虑到算法的复杂度,采用k-d树对kNN算法进行优化。

2.3 空间约束处理

基于上步的检测结果,本文采用SLIC超像素分割算法和空间邻域信息约束优化检测结果,以降低误检率和漏检率。
SLIC[19,20]是一种针对K-means的改进算法,该算法仅需输入一个参数K,用于指定生成的超像元块数。设原始影像的像元个数为N,则影像分割后形成的每个超像元块内部大约有N/K个像元,边长约为 S = N K 。算法首先每隔S个像元选取一个聚类中心作为超像元块的种子点,为了避免将影像边界点或噪声点设置为聚类中心,算法在3×3的邻域内计算梯度值最小的像元点,将其作为聚类中心;然后通过距离D确定每个像素的最近聚类中心,聚类中心的搜索邻域范围为2S×2S,当迭代搜索完成后即形成边缘保持度较高、分割形状不规则的超像元块。D采用式(3)计算,式中 N c 代表紧致度系数,一般默认设为10。对于CIELab色彩空间中的图像,像元五维特征表示形式为 P i = [ l i a i b i x i j i ] T ,其中 l a b 分别代表图像的颜色信息, d c 为像元之间的颜色距离,xy则代表图像的位置信息, d s 为像元之间的欧氏距离。
D = d c N c 2 + d s S 2 (3)
d c = l j - l i 2 + a j - a i 2 + b j - b i 2 (4)
d s = x j - x i 2 + y j - y i 2 (5)
空间约束处理过程如下:
(1)首先对双时相影像的差值影像进行SLIC分割,以避免单一时相影像各自分割所产生的复合边界失真。然后将分割超像元块边界套合至混合分类器集成系统的变化检测结果,计算每个超像元块中变化像元占块内像元总数的比值,当该比值大于预设阈值P1时,则将块内的变化像元确认为变化类别,当比值小于P1时,则将这些被混合系统判定为变化类别的像元转换成未变化类别。通过此步骤可有效抑制基于像元的检测方法所产生的“椒盐噪声”现象,减少变化检测误检率。
(2)遍历经步骤(1)处理的检测影像中的每一个像元,统计每个像元空间八邻域的像元标签,若变化像元数量大于等于6,则将中心像元设置为变化类别,产生最终的变化检测结果图。通过空间邻域约束可有效减少基于像元的检测结果中地理实体内部破碎现象,使检测结果更加完整,降低变化检测漏检率。

3 实验流程与结果分析

3.1 数据源

为了验证本文方法的有效性,分别以SPOT-5(包含红、绿、蓝、近红外和全色波段,全色波段空间分辨率2.5 m)和WorldView Ⅱ(包含红、绿、蓝、近红外和全色波段,全色波段空间分辨率0.5 m)双时相多光谱遥感影像为数据源,选取两组数据区域进行实验,如图2所示,两组数据均经过辐射定标、大气校正、影像融合(采用G-S融合)、几何校正等预处理。第一组数据集[21]为SPOT-5遥感影像,成像时间分别为2006年12月和2007年12月,影像大小为512像元×512像元,影像对应的地理区域位于广东省清远市,主要变化为水体变成裸地,以及裸地变为植被。第二组数据集为WorldView Ⅱ遥感影像,成像时间分别为2012年11月和2016年10月,影像大小为1800像元×1300像元,影像对应的地理区域位于福建省福州市,主要变化是植被、裸土和建设用地三者之间的转换。
Fig. 2 Original images for change detection

图2 变化检测的原始影像

3.2 评价指标

由于变化检测结果图中变化像元与未变化像元的数量差异较大,故本文选用Kappa系数、精确度(Presicison)、误检率和漏检率4个指标进行精度评价。参考遥感影像分类结果精度评价方法,构造误差矩阵计算精度指标,如表1所示。
Tab. 1 Error matrix of change detection

表1 变化误差矩阵

实际变化 实际未变化 总和
检测变化 TP FP P
检测未变化 FN TN N
总和 P′ N′ T
(1)精确度(Presicison)
Precision = TP TP + FP (6)
该指标反映分类器检测成变化的像元中实际变化像元的比例。
(2)Kappa系数
该指标是指一种对遥感图像的分类精度和误差矩阵进行评价的多元离散方法,用于衡量分类精度。
Kappa = Po - Pc 1 - Pc (7)
Po = TP + TN TP + TN + FP + FN (8)
Pc = TN + FN × TN + FP + TP + FP × TP + FN TP + TN + FP + FN × TP + TN + FP + FN
(9)
(3)漏检率(FNR)
FNR = FN TP + FN (10)
该指标反映实际变化检测成未变化的像元占实际变化像元的比例。
(4)误检率(FPR)
FPR = FP TP + FP (11)
该指标反映实际未变化检测成变化的像元占检测成变化的比例。

3.3 实验过程与结果分析

按照本文方法的技术路线,首先求取预处理后的双时相影像的差值影像,如图3所示,然后分别提取差值影像每个波段的光谱特征、GLCM纹理特征(包括对比度、不相似性、熵、同质性、均值、角二阶矩、方差)和形态学特征(包括开运算、闭运算和开-闭运算)构建差异特征集,其中数据集1的纹理特征方向设置为0°,扫描窗口大小为3像元×3像元,灰度量化等级为16;数据集2纹理特征的扫描窗口为5像元×5像元,其它参数设置相同。2个数据集的形态学结构算子均设置为圆形,扫描窗口为3像元×3像元。
Fig. 3 Difference images

图3 差值影像

差异特征集提取完毕后,按照2.1节所述方法进行训练样本自动选择,由于Logistic回归模型简单,计算代价较低,故被用于最佳参数训练。图4分别表示针对两组数据集在采用3种训练样本自动选择方法和不同的参数设置下,利用Logistic回归分类器进行变化初检所获得的变化与未变化误检像元数量之和。本文方法和“等间距”采样法的参数α1范围均为[0.1,0.9],变化强度采样法的参数α范围为[0.01,0.09],为图表展示方便,将α的横坐标扩大10倍。“等间距”采样法的计算公式类似式(1)、(2),该法的α1α2mn参数均与本文方法相同,变化样本采样区间C1,C2,…,Cm-1和未变化样本采样区间U1,U2,…,Un-1的左端点与本文方法相同,调整两类样本集采样区间的右端点使它们的长度各自保持一致,如式(12)-(13)所示。
C 1 = T + CS , T + CS × 1 + α 1 m ( 12 ) C 2 = T + 2 × CS , T + CS × 2 + α 1 m , , C m - 1 (13)
Fig. 4 The number of error pixels obtained with different α1

图4 不同α1下样本选择分类精度

未变化样本集同法处理。每个变化样本采样区间Ci的长度(区间右端点减左端点的绝对值)均为
C 1 = C 2 = , , = C m - 1 = α 1 × CS / m ( 14 )
未变化样本采样区间Cj的长度为 α 2 × US / n 。实验结果表明,本文方法2个数据集的最佳α1值分别为0.3和0.7,“等间距”采样法的最佳α1值分别为0.4和0.2,变化强度采样法的最佳α1值分别为0.2和0.6。将2个数据集在不同的样本选择方法最优参数时的训练样本分布状况进行可视化展示,如图5所示,其中红色像元代表变化样本,绿色像元代表未变化样本,本文方法和“等间距”采样法所提取训练样本均匀分布于整景影像,而变化强度采样法由于只抽取了变化强度两极的像元,训练样本呈现出“扎堆结块”现象,并且由于变化像元过度集中而导致样本冗余。观察图4中不同方法所对应的精度评价结果可知,“区间采样法”明显优于变化强度采样法,而在“区间采样法”中,本文提出的“非等距采样法”总体优于“等距采样法”,说明本文方法能有效减少选取过多的CVA错检样本,提升自动选择的训练样本质量。
Fig. 5 Visualization images of different sample selection strategies with their optimal α1

图5 最佳α1下不同样本自动选择方法可视化结果

基于自动选择的训练样本,利用2.2节所述方案进行变化检测,考虑算法的运行效率以及过拟合现象的影响,本文将弱分类器个数设置在[0,100]范围内,通过多组实验获取最佳弱分类器数量。由于Adaboost集成策略中每次迭代所得的分类器权重对分类结果累计值的影响较小,导致弱分类器数目相差较小时检测结果差异不大,因此每组实验的弱分类器个数设置间隔为10。如图6所示,两组数据集的弱分类器个数分别在60和70时可达到最佳变化检测精度。
Fig. 6 The number of error pixels obtained with different numbers of weak classifiers

图6 不同弱分类器集成数量的变化检测精度

为减轻“椒盐现象”的影响,降低误检率和漏检率,采用SLIC分割和空间邻域信息对像元级检测结果进行双重约束,获取最终的变化检测结果。 2个数据集的超像元预期分割块数分别设置为4500和40 000,其它系数设置为算法默认值;P1设置均为0.25。SLIC分割结果如图7所示,由于综合考虑了影像的光谱特征和空间特征,分割结果与地物的实际边缘吻合度较高。
Fig. 7 The results of SLIC segmentation

图7 SLIC分割结果

本文选用的对比方法包括CVA、Logistic回归、ID3决策树、kNN、Adaboost与Logistic回归同质集成(Adaboost-LR)、Adaboost与决策树桩同质集成(Adaboost-DS)、本文构建的混合分类器集成系统(HCS)、经过空间约束的混合分类器集成系统(HCS-SC),算法检测结果和精度评价如图8图9以及表2表3所示,变化检测标准图通过目视解译并与Google Earth Pro相近时相的同区域影像对比后进行人工标注所得,如图8(i)和图9(i)所示。
Tab. 2 The change detection accuracy assessment result of dataset 1

表2 数据集1变化检测结果精度评价

Precision Kappa TPR FPR
CVA 0.5038 0.5720 0.2314 0.4962
LR 0.6685 0.7036 0.2058 0.3315
ID3 0.6313 0.6589 0.2510 0.3687
kNN 0.7207 0.7188 0.2412 0.2793
Ada-LR 0.7090 0.7530 0.1534 0.2910
Ada-DS 0.7376 0.7279 0.2431 0.2624
HCS 0.8371 0.8472 0.1197 0.1629
HCS-SC 0.8756 0.8691 0.1190 0.1244
Tab. 3 The change detection accuracy assessment result of dataset 2

表3 数据集2变化检测结果精度评价

Precision Kappa TPR FPR
CVA 0.5725 0.5428 0.3481 0.4275
LR 0.6224 0.6024 0.2968 0.3776
ID3 0.7046 0.5809 0.4260 0.2954
kNN 0.6923 0.5841 0.4108 0.3077
Ada-LR 0.7611 0.7050 0.2711 0.2389
Ada-DS 0.8044 0.6794 0.3505 0.1956
HCS 0.9013 0.8320 0.1889 0.0987
HCS-SC 0.9197 0.8383 0.1942 0.0803
Fig. 8 The change detection results and the reference image of dataset 1

图8 数据集1检测结果及标准影像

Fig. 9 The change detection results and the reference image of dataset 2

图9 数据集2检测结果及标准参考影像

变化检测结果表明:CVA针对两组数据的检测效果较差,检测结果中存在大量错检像元,数据集1主要集中于影像北侧,数据集2分散于影像各处; 3种监督分类算法针对2个数据集的分类结果均明显优于CVA,但仍存在不同的程度的漏检和误检现象;而通过Adaboost策略集成后的决策树桩和Logistic回归分类器检测效果更佳,4个类检测指标全部优于单分类器针对同一数据集的检测结果。从检测图8(e)可看出,Ada-LR的结果中存在许多“椒盐现象”的误检像元,但地理实体内部较为完整,而图8(f)即Ada-DS的检测结果中误检现象受到较好的抑制,但是对于中上部分和东北角侧人造裸地变化的检测不完整,因此Ada-LR的漏检率(0.1534)低于Ada-DS(0.2431),而后者的误检率(0.2624)低于前者(0.2910),类似地,数据集2图9(e)(Ada-LR)的检测结果比图9(f)(Ada-DS)更为完整(漏检率分别为0.2711和0.3505),但前者比后者产生出更多的误检碎斑(误检率分别为0.2389和0.1956),这说明通过Adaboost策略集成后的Logistic回归和决策树桩仍然表现出一定程度的性能互补性,因此不管是否经过空间约束处理,本文提出的2种方法的 4个检测指标结果均优于Adaboost的同质单分类器集成。数据集2中虽然本文方法对于影像中光谱特征较强的区域(屋顶颜色更替、阴影遮挡等)出现部分错检,但对于特征较弱的区域的检测效果较好(如影像西北角、北侧、东侧),不管是裸地到建筑的变化还是裸地到大型开阔人造区域的变化均能检测出较完整的边缘信息和内部信息,并且对于“椒盐噪声”的抑制效果较好,误检率(0.0987和0.0803)均低于比较方法中的最低值(0.1956),本文方法总体效果最佳。
此外,经空间约束处理与不经空间约束处理的检测结果对比,前者漏检率与后者持平或略高于后者(数据集1(0.1197和0.1190)、数据集2(0.1889和0.1942)),这是由于经过SLIC分割和空间邻域信息约束处理后,原始检测结果中的大量“椒盐噪声”被滤除,但同时也不可避免地滤除掉一部分被正确检测的细小地物以及地理实体内部的少量不连续像元,但总体而言在保持漏检率基本持平的前提下,空间约束处理能明显降低误检率(数据集1(0.1629和0.1224)、数据集2(0.0987和0.0803))。

4 结论

(1)本文提出的二类变化检测训练样本自动选择策略的自动化程度与正确率均优于现有同类研究成果。如何将该方法拓展到多类变化检测样本自动选择是后续研究努力的方向。
(2)集成学习可显著提升针对高分影像的变化检测精度。就Adaboost集成策略而言,同种分类器的同质集成优于单分类器监督学习算法,而多种分类器的同异质混合互补集成又优于同种分类器的同质集成。如何合理选择异质弱分类器并针对其特定的组合构建高效、鲁棒的集成策略是后续研究的重点。
(3)本文提出的“SLIC分割+邻域空间信息约束”可有效抑制像元层次的影像信息提取结果中存在的“椒盐噪声”,降低变化检测的误检率和漏检率。

The authors have declared that no competing interests exist.

[1]
Singh A.Digital change detection techniques using remotely-sensed data[J]. International Journal of Remote Sensing, 1989,10(6):989-1003.

DOI

[2]
Gu Y J, Jin Z, Chiu S C.Active learning combining uncertainty and diversity for multi-class image classification[J]. IET Computer Vision, 2015,9(3):400-407In computer vision and pattern recognition applications, there are usually a vast number of unlabelled data whereas the labelled data are very limited. Active learning is a kind of method that selects the most representative or informative examples for labelling and training; thus, the best prediction accuracy can be achieved. A novel active learning algorithm is proposed here based on one-versus-one strategy support vector machine (SVM) to solve multi-class image classification. A new uncertainty measure is proposed based on some binary SVM classifiers and some of the most uncertain examples are selected from SVM output. To ensure that the selected examples are diverse from each other, Gaussian kernel is adopted to measure the similarity between any two examples. From the previous selected examples, a batch of diverse and uncertain examples are selected by the dynamic programming method for labelling. The experimental results on two datasets demonstrate the effectiveness of the proposed algorithm.

DOI

[3]
吴田军,骆剑承,夏列钢,等.迁移学习支持下的遥感影像对象级分类样本自动选择方法[J].测绘学报,2014,43(9):908-916.lt;p>面向遥感大范围应用的目标,自动化程度仍是遥感影像分类面临的重要问题,样本的人工选择难以适应当前土地覆盖信息自动化提取的实际应用需求。为了构建一套基于先验知识的遥感影像全自动分类流程,本文将空间信息挖掘技术引入到遥感信息提取过程中,提出了一种面向遥感影像对象级分类的样本自动选择方法。该方法通过变化检测将不变地物标示在新的目标影像上,并将过去解译的地物类别知识迁移至新的影像上,建立新的特征与地物关系,从而完成历史专题数据辅助下目标影像的自动化的对象级分类。实验结果表明,在已有历史专题层的图斑知识指导下,该方法能有效地自动选择适用于新影像分类的可靠样本,获得较好的信息提取效果,提高了对象级分类的效率。</p>

DOI

[ Wu T J, Luo J C, Xia L G, et al.An automatic sample collection method for object-oriented classification of remotely sensed imageries baed on transfer learning[J]. Acta Geodaetica et Cartographica Sinica, 2014,43(9):908-916. ]

[4]
Huo C, Zhou Z, Lu H., et al.Fast object-level change detection for VHR images[J]. IEEE Geoscience and Remote Sensing Letters, 2010,7(1):118-122.A novel approach is presented for change detection of very high resolution images, which is accomplished by fast object-level change feature extraction and progressive change feature classification. Object-level change feature is helpful for improving the discriminability between the changed class and the unchanged class. Progressive change feature classification helps improve the accuracy and the degree of automation, which is implemented by dynamically adjusting the training samples and gradually tuning the separating hyperplane. Experiments demonstrate the effectiveness of the proposed approach.

DOI

[5]
张鑫龙,陈秀万,李飞,等.高分辨率遥感影像的深度学习变化检测方法[J].测绘学报,2017,46(8):999-1008.为提升高分辨率遥感影像的变化检测精度,提出一种利用深度学习的变化检测方法。在预处理的基础上,利用顾及邻域信息的改进变化矢量分析算法和灰度共生矩阵算法获取影像间光谱和纹理变化,并通过设置自适应采样区间提取最可能的变化和未变化区域样本。构建并训练包含标签层的高斯伯努利深度限制玻尔兹曼机模型,以提取变化和未变化区域深层特征,从而有效辨别变化区域。通过WorldView-3与Pléiades-1影像的试验表明本文方法在变化检测精度方面优于对比方法。

DOI

[ Zhang X L, Chen J W, Li F, Et al. Change detection method for high resolution remote sensing images using deep learning[J]. Acta Geodaetica et Cartographica Sinica, 2017,46(8):999-1008. ]

[6]
Tan K, Jin X, Antonio P, et al.Automatic change detection in high-resolution remote sensing images by using a multiple classifier system and spectral-spatial features[J]. IEEE Journal of Selected Topics in Applied Earth Observations And Remote Sensing, 2016,9(8):3439-3451.Change detection (CD) is an active research topic in remote sensing applications including urban studies, disaster assessment, and deforestation monitoring. In this paper, we propose an automatic method for CD in high-resolution remote sensing images that uses a novel strategy for the selection of training samples and an ensemble of multiple classifiers. As for the selection of training samples, our proposed method uses two groups of thresholds instead of just one threshold to enhance the quality of the selected training samples by allowing for their selection in an intelligent manner. In order to achieve high CD accuracy, spatial information such as texture and morphological profiles are utilized in conjunction with spectral information. Our multiple classifier system (MCS) exploits the extreme learning machine (ELM), multinomial logistic regression (MLR), and K-nearest neighbor (KNN) classifiers. To validate our newly proposed approach, we conduct experiments using multispectral images collected by ZY-3. The proposed method provides state-of-the-art CD accuracies as compared with other approaches widely used in the literature for CD purposes.

DOI

[7]
Cao G, Wang B, H. Xavier, et al. A new difference image creation method based on deep neural networks for change detection in remote-sensing images[J]. International Journal of Remote Sensing, 2017,38(23):7161-7175.Abstract In this article, we propose a novel difference image (DI) creation method for unsupervised change detection in multi-temporal multi-spectral remote-sensing images based on deep learning theory. First, we apply deep belief network to learn local and high-level features from the local neighbour of a given pixel in an unsupervised manner. Second, a back propagation algorithm is improved to build a DI based on selected training samples, which can highlight the difference on changed regions and suppress the false changes on unchanged regions. Finally, we get the change trajectory map using simple clustering analysis. The proposed scheme is tested on three remote-sensing data sets. Qualitative and quantitative evaluations show its superior performance compared to the traditional pixel-level and texture-level-based approaches.

DOI

[8]
魏立飞,钟燕飞,张良培,等.遥感影像融合的自适应变化检测[J].遥感学报,2010,14(6):1196-1211.提出一种基于影像融合和自适应阈值选择的遥感影像变化检测方法.首先利用经过改进的融合技术对原始数据的差值影像和比值影像进行处理,构造融合影像,在该融合影像的基础上进行自适应迭代运算得到初步变化阈值范围,然后通过分析阈值范围两侧影像像元的离散程度,求解最终的阈值范围,从而得到更优变化阈值,提取变化区域.实验结果表明,本文方法的检测精度优于传统的变化检测方法,同时具有一定的稳定性和智能性.

[ Wei L F, Zhong Y F, Zhang L P, et al.Adaptive change method of remote sensing image fusion[J]. Journal of Remote Sensing, 2010,14(6):1196-1211. ]

[9]
Freund Y, Schapire R.A decision-theoretic generalization of on-line learning and an application to boosting[J]. Journal of Computer and System Sciences, 1997,55(1):119-139.

DOI

[10]
龚健雅,姚璜,沈欣.利用AdaBoost算法进行高分辨率影像的面向对象分类[J].武汉大学学报·信息科学版,2010,35(12):1440-1443.利用面向对象的遥感影像处理技术,通过影像分割构建影像对象,选取若干分割后的影像对象作为训练样本。利用提取出的训练样本并利用多种特征(光谱、纹理、空间),使用AdaBoost算法对影像对象进行分类。QuickBird影像分类实验证明了此方法的有效性。

[ Gong J Y, Yao H, Shen X.Object-oriented classification of high spatial-resolution remote sensing imagery based on AdaBoost[J]. Geomatics and Information Science of Wuhan University, 2010,35(12):1440-1443. ]

[11]
Woo D M, Do V D.Post-Classification change detection of high resolution satellite images using AdaBoost classifier[J]. Information Technology & Computer Science, 2015,117:34-38.

[12]
Chen Y, Dou P, Yang X.Improving land use/cover classification with a multiple classifier system using AdaBoost integration technique[J]. Remote Sensing, 2017,9(10):1055-1074.

DOI

[13]
Schapire R E, Singer Y.Improved boosting algorithms using confidence-rated predictions[J]. Machine Learning, 1999,37(3):297-336.

DOI

[14]
Wu S Q, H. Nagahashi. Analysis of generalization ability for different AdaBoost variants based on classification and regression trees[J]. Journal of Electrical and Computer Engineering, 2015,2015:1-17.

[15]
Zhang C, Ma Y.Ensemble machine learning: methods and applications[M]. New York: Springer-Verlag, 2012,88.

[16]
Mancini A, Frontoni E, Zingaretti P.Road change detection from multi-spectral aerial data[C]// International Conference on Pattern Recognition, 2010:448-451.

[17]
Le H L, Woo D M.Combination of textural features for the improvement of terrain classification and change detection[J]. International Journal of Software Engineering And Its Applications, 2015,9(5):145-154.

DOI

[18]
Peter H,李锐,李鹏,等.机器学习实战[M].北京:人民邮电出版社,2013.

[ Peter H., Li R, Li P, et al.Machine learning in action[M]. Beijing: Posts And Telecom Press, 2013. ]

[19]
Barreto T L M, Rosa R A S, Wimmer C, et al. Classification of detected changes from multitemporal high-res Xband SAR images: intensity and texture descriptors from superpixels[J]. IEEE Journal of Selected Topics In Applied Earth Observations and Remote Sensing, 2016,9(12):5436-5448.Remote sensing has been widely employed for monitoring land cover and usage by change detection techniques. In this paper, we cope with the early detection of the first signs of deforestation, which is the gateway for illegal activities, such as unauthorized urban sprawl and grazing use. In recent years, object-based approaches have emerged as a more suitable alternative than pixel-based methods for change detection in remote sensing images. Even though several classifiers have been tested, there was little effort in selecting appropriated features for the classification of detected changes. After a deep analysis of the existing segmentation, feature extraction, and classification approaches, we propose an object-based methodology that consists of: 1) segmenting multitemporal Xband high-resolution synthetic aperture radar (SAR) images into superpixels employing the simple linear iterative clustering algorithm; 2) extracting features using the object correlation images framework and with the gray-level cooccurrence matrix; and 3) classifying areas into unchanged, deforestation, and other changes by means of a multilayer perceptron supervised learning technique. Experiments were performed using high-resolution SAR images obtained by the airborne sensor OrbiSAR-2 from BRADAR in challenging scenarios of the Brazilian Atlantic Forest, including a wide variety of vegetation, rivers, sea coasts, urban, harvest and open areas, and humidity changes. We perform an extensive experimental analysis of the results, comparing the proposed method with a state-of-the-art approach. The results demonstrate that our method yields an improvement of over 10% in the accuracy while detecting changes and classifying deforested areas.

DOI

[20]
Toro C, Martín C G, Pedrero Á G, et al.Superpixel-based roughness measure for multispectral satellite image segmentation[J]. Remote Sensing, 2015,7(11):14620-14645.The new generation of artificial satellites is providing a huge amount of Earth observation images whose exploitation can report invaluable benefits, both economical and environmental. However, only a small fraction of this data volume has been analyzed, mainly due to the large human resources needed for that task. In this sense, the development of unsupervised methodologies for the analysis of these images is a priority. In this work, a new unsupervised segmentation algorithm for satellite images is proposed. This algorithm is based on the rough-set theory, and it is inspired by a previous segmentation algorithm defined in the RGB color domain. The main contributions of the new algorithm are: (i) extending the original algorithm to four spectral bands; (ii) the concept of the superpixel is used in order to define the neighborhood similarity of a pixel adapted to the local characteristics of each image; (iii) and two new region merged strategies are proposed and evaluated in order to establish the final number of regions in the segmented image. The experimental results show that the proposed approach improves the results provided by the original method when both are applied to satellite images with different spectral and spatial resolutions.

DOI

[21]
Zhou L, Cao G, Li Y, et al.Change detection based on conditional random field with region connection constraints in high-resolution remote sensing images[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2016,9(8):3478-3488.In this paper, a novel change detection method based on conditional random field (CRF) with region connection constraints in multitemporal high-resolution remote sensing images is proposed. The change detection problem is formulated as a labeling issue to discriminate the changed class from the unchanged class in the difference image. In the CRF model, the unary potential is described by using the memberships of unsupervised fuzzy C-means clustering algorithm. The pairwise potential adopts a boundary constraint based on Euclidean distance. In addition, region iteration potential defined on a set of pixels is incorporated into CRF model to suppress the oversmooth performance. A chief advantage of our approach is to be able to achieve correct change map and avoid training a large number of model parameters. Experimental results demonstrate that the proposed method improves the change detection accuracy, is more robust against noise than other state-of-the-art approaches, and preserves boundary information.

DOI

文章导航

/