改进特征金字塔网络的遥感影像崩滑体提取  

Collapse and landslide extraction from remote sensing image based on improved feature pyramid network

作　　者：高琛 冯德俊 胡金林 王杰茜 GAO Chen;FENG Dejun;HU Jinlin;WANG Jiexi(Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学地球科学与环境工程学院,成都611756

出　　处：《测绘科学》2021年第11期32-38,46,共8页Science of Surveying and Mapping

基　　金：科技基础资源调查专项(2019FY202504)。

摘　　要：针对大多数网络存在精度低,特征冗余,计算量大,训练时间长等问题,提出密集连接特征金字塔网络(DCFPN),将特征提取网络得到的特征图通过一组并行深度可分离空洞卷积进一步计算其全局语义信息,并搭建解码上采样网络,加入连接组合特征层的结构,对遥感影像进行语义分割实现崩滑体提取,较好地解决了参数量过多,计算时间较长和精度较低等问题。通过特征金字塔网络(FPN)和DCFPN在崩滑体数据集上的大量对比实验表明,DCFPN在崩滑体语义分割方面有更高的精度并且计算量更少,训练时间更短,能够更好地为应急抢险工作。In view of the problems of low precision,redundant feature,large computation and long training time in most networks,dense connection feature pyramid network(DCFPN)is proposed.The whole semantic information of the feature map obtained from feature extraction network is further calculated by a set of parallel depth separable dilated convolution,and the sampling network is built on decoding to join the structure of the combined feature layer.The semantic segmentation of remote sensing image can realize the Collapse-landslip body extraction,which can solve the problems of too many parameters,long calculation time and low accuracy.A large number of comparative experiments between FPN and DCFPN on the collapse-landslip body data set show that DCFPN has higher accuracy in semantic segmentation of collapse-landslip body,with less calculation,shorter training time,and can better serve the emergency rescue work.

关 键 词：崩滑体 密集连接特征金字塔网络 深度可分离卷积 空洞卷积 

分 类 号：P237[天文地球—摄影测量与遥感]