基于图卷积网络的海洋测绘遥感影像场景分类研究  

Research on Scene Classification of Remote Sensing Imagery for Marine Mapping Based on Graph Convolutional Networks

作　　者：王志彬 刁永洲 Wang Zhibin;Diao Yongzhou(Guangzhou Maritime Surveying and Mapping Center,Guangzhou 510200,China)

机构地区：[1]广州海事测绘中心,广州510200

出　　处：《科技通报》2023年第11期15-18,23,共5页Bulletin of Science and Technology

摘　　要：海洋测绘遥感影像噪声难以去除,在遥感影像场景分类时,存在去噪效果不明显、增强效果不显著、分类准确率低等问题。针对这些问题本文提出基于图卷积网络的海洋测绘遥感影像场景分类方法。采用小波阈值图像去噪算法对遥感影像场景进行去噪,再通过NSST(nonsubsampled shearlet transform)方法得到去噪后影像的低频子带滤波与高频子带滤波,并对其展开增强处理,将预处理之后的遥感影像输进图卷积网络模型中,运用该模型中的差异化单元、分类器学习单元与遥感影像特征差异化单元实现最终分类。实验结果表明:与对照方法相比,该方法对遥感影像的去噪效果更好,增强效果更明显,场景分类准确率更高。说明该方法能够有效提升海洋测绘遥感影像的场景分类性能。It is difficult to remove the noise from the remote sensing image of marine surveying and mapping.In the classification of remote sensing image scenes,there are some problems such as insignificant noise removal effect,insignificant enhancement effect,and low classification accuracy.Aiming at these problems,a method of scene classification of marine remote sensing image based on graph convolution network is proposed.The wavelet threshold image denoising algorithm is used to denoise the remote sensing image scene.Then,the low frequency sub-band filter and high frequency sub-band filter of the denoised image are obtained by NSST(nonsubsampled shearlet transform)method and enhanced.The remote sensing image after preprocessing is input into the graph convolution network model,and the final classification is achieved by using the differentiation unit,classifier learning unit and remote sensing image feature differentiation unit in the model.The experimental results show that compared with the control methods,this method has better denoising effect,more obvious enhancement effect and higher accuracy of scene classification.It shows that the proposed method can effectively improve the scene classification performance of marine mapping remote sensing images.

关 键 词：遥感影像场景分类 小波阈值图像去噪算法 高频子带滤波 图卷积网络模型 

分 类 号：TN391[电子电信—物理电子学]