基于Xception模型的遥感影像场景变化检测  

Scene Change Detection of Remote Sensing Image Based on Xception Model

作　　者：李长城 何海清[1] 章李乐 涂明 LI Changcheng;HE Haiqing;ZHANG Lile;TU Ming(Faculty of Geomatics,East China University of Technology,330000,Nanchang,PRC;Nanchang Water Conservancy Planning Service Center,330000,Nanchang,PRC)

机构地区：[1]东华理工大学测绘工程学院,南昌330000 [2]南昌市水利规划服务中心,南昌330000

出　　处：《江西科学》2022年第3期568-573,共6页Jiangxi Science

基　　金：国家自然科学基金项目(41861062);抚州市青年科技领军人才计划项目(2020ED65);江西省水利厅科技项目(202123TGKT12)。

摘　　要：由于遥感场景固有的语义特征,深度学习在场景分类中具有突出的语义特征提取能力,近年来得到了广泛的研究。然而,现有的基于深度学习的方法大多没有精细描绘场景几何形状,而是利用固定大小的滑动窗口来对影像分类,这些窗口可能由多种场景类型的混合像元组成,导致场景识别的可分离性低,细节粗糙,分类精度低。针对这一问题,将简单线性迭代聚类(Simple Linear Iterative Clustering,SLIC)与迁移学习相结合,实现场景斑块的精确分类,进一步提高场景变化检测的准确性。首先基于SLIC算法,考虑强度分布的相关性,将遥感影像分割成多尺度斑块;其次利用Xception模型结合迁移学习网络从多尺度斑块中提取深度语义特征,并基于softmax分类器计算概率,从而判别场景斑块的类别;最后,基于场景分类的结果提取地物的变化信息。实验结果表明,该方法能够适应不同尺度的场景识别,达到利用窗口网格检测方法的最高水平,提高了场景变化检测的精度和效率。Due to the inherent semantic features of remote sensing scenes,deep learning has outstanding semantic feature extraction ability in scene classification and has been widely studied in recent years.However,most of the existing methods based on deep learning do not describe the scene geometry in detail,but classify the image by using sliding windows of fixed size.These windows may be composed of mixed pixels of various scene types,resulting in low separability of scene recognition,rough details and low classification accuracy.To solve this problem,simple linear iterative clustering(SLIC)is combined with transfer learning to achieve accurate classification of scene patches,which further improves the accuracy of scene change detection.Firstly,based on SLIC algorithm,the remote sensing image was segmented into multi-scale patches considering the correlation of intensity distribution.Secondly,deep semantic features are extracted from multi-scale patches using Xception model combined with transfer learning network,and probability was calculated based on softmax classifier to distinguish the category of scene patches.Finally,the change information of ground objects is extracted based on the result of scene classification.Experimental results show that this method can adapt to scene recognition of different scales and achieve the highest level of using window grid detection method.At the same time it improve the accuracy and efficiency of scene change detection.

关 键 词：场景分类 变化检测 简单线性迭代聚类 迁移学习 Xception 

分 类 号：TP237[自动化与计算机技术—检测技术与自动化装置]