高分辨率单极化SAR图像慢动船只散射特性稳健高层表征研究

Slowly moving ship classification using synthetic aperture radar image (SAR ship classification) plays a crucial role in oceanographical remote sensing, providing precise ship information for territorial security, maritime law enforcement and environment protection in all day and night, under various weather conditions, and over a relatively large area. SAR ship classification confronts with many obstacles, including the unstable low level image features, lack of quad-pol SAR images with resolution better than 1m, and insufficient labeled ship samples. To solve these problems, a deep learning based robust high level representation of scattering characteristics for slowly moving ship in single polarimetric high resolution SAR image is proposed. In this study, the main research efforts include: 1) establishing the SAR imaging model for slowly moving ship to analyze the cause of scattering characteristics inversion errors and devise the Doppler compensation method based on the refocusing process; 2) inverting the scattering characteristics of ships in the refocused single polarimetric SAR image using the time-frequency hyperimage model and constructing the multi-dimension scattering feature image (MDSFI); 3) establishing the stacked denoising convolution sparse auto-encoder (SDCSAE) model based on the convolution-pooling structure to extend the denoising sparse auto-encoder (DSAE) into multi-dimension feature image domain for ship scattering characteristics coding and high level representation constructing, which will greatly benefit the ship classification capability with limited labeled ship samples. This study provides a new way for robust physical feature extraction in SAR ship classification, which can be widely used in both military and civil field, including national defense, traffic and environment.

高分辨率SAR图像慢动船只分类可为领海安全、海上执法和海洋环境保护等活动提供全天时、全天候、大范围和高精度的船只信息，是海洋遥感的研究热点之一。针对其面临的特征高易变、米级以上全极化数据不足和有标记样本缺乏等问题，本项目拟开展基于深度学习的高分辨率单极化SAR图像慢动船只散射特性高层表征方法研究。建立慢动船只SAR成像模型，分析慢动船只散射特性反演失真成因，设计船只运动多普勒相位补偿和图像重聚焦方法，提高散射特性反演精度。采用时频超图像模型，在单极化重聚焦SAR图像中反演船只低层散射特性，构建多维散射特征图像。建立基于卷积-池化结构的栈式降噪卷积稀疏自编码器模型，将深度学习自编码器扩展到多维特征图像域，提取慢动船只散射特性潜在高层表征，改善有限样本下船只分类性能。本项目提出了一种SAR图像船只稳健物理特征提取的新方法，对推动SAR船只分类技术在国防、交通、环境等领域的应用具有重要意义。

针对高分辨率单极化SAR图像海洋船只分类研究中存在的有标记样本不足，常用特征易受雷达成像条件和图像质量影响具有高易变性的问题，本课题将SAR图像时频超图像模型(TFHI)与深度学习自编码器模型结合，开展了船只散射特性稳健高层表征提取研究。设计了基于均质区域分割和全局迭代审查的船只切片自动提取技术，改善了高分辨率SAR图像船只检测的检测性能，降低了虚警；在对慢动船只成像散焦机理分析的基础上，引入相位梯度自聚焦算法，补偿了散焦图像中船只运动导致的方位向残余非线性多普勒相位，实现了对散焦图像的重聚焦；提出了基于Batch-Triplet卷积自编码器的船只峰值中心二维谱图散射特性提取方法，在有限样本情况下，实现了对船只峰值中心局部散射特性的提取；提出了复数域多尺度卷积自编码器模型，实现了对船只复子孔径图像序列的中间层散射特性提取；提出了基于稀疏自编码器模型的宏特征编码与空间池化技术，实现了对船只峰值中心局部级散射特征的中间层表征学习，设计了多尺度一维卷积网络，实现了对船只TFHI多投影散射特性的融合与船只散射特性高层表征的提取。基于高分辨率TerraSAR和Radarsat-2数据实验结果表明，通过多级深度学习网络抽象得到的船只散射特性高层表征可有效的改善船只分类精度。研究成果对于我国海上交通监测、渔业资源保护和海洋环境保护具有重要意义。