摘要
我国城市化发展迅速,地表利用信息处于不断变动中,及时掌握这些变化信息十分必要。但实际作业中,变化信息获取方式仍然以高人力成本方法,如实地外业调查或者目视遥感影像判读为主,生产效率低。因此,本文提出了一种基于语义分割的深度学习变化检测方法。首先,利用编码-解码深度卷积网络,实现遥感影像地物的自动分类;然后,利用Mean-Shift方法分割前后期影像,融合其光谱、纹理和语义信息等特征,对比前后期影像的特征差异,提取出变化置信度图;最后通过EM算法分割变化与未变化类生成二值变化图,得出变化区域范围。该方法为自动化实现地物变化监测提供了有效的解决办案,实验证明,该方法相比人工以及传统分类模型具有更好的检测精度,有效降低了内外业工作量。
China′s urbanization is developing rapidly,and land-use information is constantly changing.It is necessary to detect these changes in time.However,in current practice,the methods to detect change information are mostly field surveys or manual interpretation,so very labor intensive and costly,and the production efficiency is low.Therefore,this paper proposes a change detection method which is based on deep learning semantic segmentation.Firstly,the encoding-decoding deep convolutional network is used to realize the automatic classification of the ground features in the remote sensing images.Secondly,the Mean-Shift method is used to segment the image,and fuse image features of segmented objects,such as spectrum,texture and semantics.Then,it compares the difference of previous and later image features and extracts the change confidence map.Finally,the EM algorithm is used to segment the change classes and the unchanged classes to generate a binary change map,and to obtain the change area boundary.This method provides an effective solution for detecting the land change automatically.It not only has better detection accuracy than the manual and traditional classification models,but also reduces both field and indoor workload.
作者
吴海平
温礼
邓凯
陈璐瑶
李小凯
WU Haiping;WEN Li;DENG Kai;CHEN Luyao;LI Xiaokai(Institute of China Land Surveying and Planning,Beijing 100035,China;Handleray Technology Co.,Ltd.,Wuhan 430073,China)
出处
《测绘与空间地理信息》
2021年第7期102-106,共5页
Geomatics & Spatial Information Technology
基金
自然资源部专项——全国土地利用变更调查监测与核查(ZX180901)资助。
关键词
深度卷积网络
语义分割
影像特征
EM算法
deep convolutional network
semantic segmentation
image features
EM algorithm
