融合深度空间特征的TSVM自动遥感变化检测方法  

Automatic Remote Sensing Change Detection with TSVM Fusing Depth Space Features

作　　者：谢志伟 李文刚 孙立双[2] 苏国庆 XIE Zhiwei;LI Wengang;SUN Lishuang;SU Guoqing(Key Laboratory of Virtual Geographic Environment,Ministry of Education,Nanjing Normal University,Nanjing 210023,China;School of Transportation and Geomatics Engineering,Shenyang Jianzhu University,Shenyang 110168,China;State Key Laboratory of Geographical Environment Evolution,Nanjing 210023,China;Jiangsu Collaborative Innovation Center for Development and Application of Geographic Information Resources,Nanjing 210023,China)

机构地区：[1]南京师范大学虚拟地理环境教育部重点实验室,南京210023 [2]沈阳建筑大学交通与测绘工程学院,沈阳110168 [3]地理环境演化国家重点实验室培养基地,南京210023 [4]江苏省地理信息资源开发与应用协同创新中心,南京210023

出　　处：《遥感信息》2025年第1期10-18,共9页Remote Sensing Information

基　　金：国家自然科学基金(42101353);教育部人文社会科学研究一般项目(21YJC790129);辽宁省教育厅基本科研项目(LJKMZ20220946、LJKMZ20222128)。

摘　　要：为了解决直推式支持向量机(transductive support vector machines,TSVM)在样本选择自动化程度低和特征学习充分性不足的问题,提出了一种融合深度空间特征与传统影像对象特征的TSVM自动高分遥感影像变化检测方法。首先,采用基于分形网络演化算法的叠置分割获取多时相高分遥感影像的影像对象,通过卷积神经网络提取遥感影像的深度空间特征,并与灰度、指数和纹理等传统影像对象特征联合构建特征空间;然后,利用卡方变换计算多维特征的加权特征差异度,采用最大期望算法和贝叶斯最小错误判别规则得到二值分割结果,依据变化概率自动将分割结果中准确率较高的部分标记为训练样本;最后,采用标记训练样本获得TSVM的多维特征空间二值分割超平面,进而完成自动变化检测。选择武汉市的两组高分数据集作为实验数据。实验结果表明,该方法能够实现样本自动选择,并且通过融合深度空间特征可以有效提高特征学习的充分性,平均准确率达到了88.84%,平均漏检率较仅利用传统影像对象特征的TSVM法降低了3.29个百分点,在定性和定量的变化检测有效性评价中均得到了提高。In order to solve the problems of low automation of sample selection and insufficient feature learning adequacy of transductive support vector machines(TSVM),an automatic high-resolution remote sensing image change detection method for TSVM that integrates deep spatial features with traditional image object features is proposed.First,the image objects of multi-temporal high resolution remote sensing image are obtained by superposition segmentation based on fractal network evolution algorithm,and the depth space features of remote sensing image are extracted by convolutional neural network,and the feature space is jointly constructed with the traditional image object features such as grayscale,exponent,and texture.Then,the weighted feature difference degree of multi-dimensional features is calculated by using Chi-square transform,and the maximum expectation algorithm and the Bayesian least-error discrimination rule are used to obtain the binary segmentation results,and the part of the segmentation results with higher accuracy rate is automatically labeled as the training samples based on the change probability.Finally,the labeled training samples are used to obtain the binary segmentation hyperplane of the multidimensional feature space of the TSVM,which in turn completes the automatic change detection.Two sets of high-resolution data sets from Wuhan are selected as experimental data,and the experimental results show that the method in this paper can realize automatic sample selection,and the adequacy of feature learning can be effectively improved by fusing the deep spatial features,with an average accuracy of 88.84%,and an average leakage rate of 3.29%lower than that of the TSVM method which only utilizes the features of the traditional image objects,and the validity of the change detection is improved both in qualitative and quantitative evaluation.

关 键 词：叠置分割 样本自动选择 直推式支持向量机 变化检测 卷积神经网络 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]