多尺度轻量化CNN在SAR图像地物分类中的应用  

Application of multi-scale and lightweight CNN in SAR image-based surface feature classification

作　　者：孙盛[1] 蒙芝敏 胡忠文[2] 余旭[3] SUN Sheng;MENG Zhimin;HU Zhongwen;YU Xu(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China;Key Laboratory for Geo-Environmental Monitoring of Great Bay Area,Ministry of Natural Rresources,Shenzhen University,Shenzhen 518000,China;School of Civil and Transportation Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广州510006 [2]深圳大学自然资源部大湾区地理环境监测重点实验室,深圳518000 [3]广东工业大学土木与交通工程学院,广州510006

出　　处：《自然资源遥感》2023年第1期27-34,共8页Remote Sensing for Natural Resources

基　　金：国家自然科学基金项目“纠缠态的超级纠缠目击者特性及其推广研究”(编号:61672007);自然资源部大湾区地理环境监测重点实验室开放基金项目“基于多极化星载合成孔径雷达图像的粤港澳大湾区海岸线动态监测”(编号:2019002);广东省国际合作领域项目(编号:2019A050509009)共同资助。

摘　　要：结合粤港澳大湾区的亚热带气候特点,采用TerraSAR-X雷达遥感卫星对实验区域进行了图像采集;针对雷达卫星观测场景中地物目标尺度变化不一的问题,提出了一个应用于地物分类的卷积神经网络模型(ENet convolution spatial pyramid pooling,ENet-CSPP)。利用了普通卷积比空洞卷积更好保持领域信息的特点,提出了多尺度特征融合模块——卷积空域金字塔池化模块;针对SAR遥感图像数据集训练样本偏少的问题,提出了将多尺度特征融合模块和轻量化卷积神经网络结合起来的方法;ENet-CSPP网络的编码器部分由改进后的ENet网络和卷积空域金字塔池化模块构成,解码器部分实现深、浅层特征的融合后输出地物分类图像。在GDUT-Nansha数据集上进行了定量对比实验,ENet-CSPP模型在像素精度、平均像素精度和平均交并比3个性能指标上都要优于其他模型,表明多尺度轻量化的模型有效提高了地物分类的精度。Targeting the subtropical climate characteristics of the Guangdong-Hong Kong-Macao Greater Bay Area,this study acquired the images of the experimental area from the TerraSAR-X Radar remote sensing satellite.Given the varying scale of the surface feature targets in the Radar satellite observation scenes,this study proposed an ENet convolution spatial pyramid pooling module(ENet-CSPP)model for surface feature classification.Since ordinary convolution can more effectively maintain domain information than atrous convolution,this study proposed a multi-scale feature fusion module based on convolution spatial pyramid pooling.Since there were a few training samples in the SAR remote sensing image dataset,this study combined the multi-scale feature fusion module with the lightweight convolutional neural network.The encoder of the ENet-CSPP network consisted of an improved ENet network and the convolution spatial pyramid pooling module.The decoder output surface feature classification images after the fusion of deep and shallow features.The quantitative comparison experiments were conducted on the GDUT-Nansha dataset.The ENet-CSPP model outperformed other models in three performance indices,namely pixel accuracy,average pixel accuracy,and mean intersection over union.This result indicates that the multi-scale lightweight model effectively improved the accuracy of surface feature classification.

关 键 词：合成孔径雷达 地物分类 卷积神经网络 轻量化网络 

分 类 号：P236[天文地球—摄影测量与遥感]