随机海况下的长序列全极化SAR围填海特征分析与精确识别研究

Marine reclamation is an important task in the dynamic monitoring of sea area utilization. The novel GF-3 satellite can provide long-sequence polarimetric synthetic aperture radar (SAR) data, which is beneficial to represent the backscatter characteristics of various types of reclamation and provide an opportunity for the accurate detection of reclamation under random oceanic conditions. As for various types of reclamation under random oceanic conditions, their imaging characteristics in GF-3 SAR images have not been studied, their information are affected by speckle noises, and classic recognition models cannot detect the information change in the same image of different sea areas or in different images of the same sea area. In order to overcome these problems, the research project takes the backscattering characteristics analysis of reclamation under the random oceanic conditions as the breakthrough, and carries out the following work. (1) To obtain polarimetric features and reveal the backscattering characteristics law of various types of reclamation through ISAR imaging simulation experiments under random oceanic conditions. (2) Combined with imaging characteristics, the rapid detection models for coastlines and typical types of reclamation is established, and the region of reclamation can be located rapidly. (3) Weight clustering nonlocal stacked autoencoders based on L1,2 norm is proposed, and applied in the decision-level judgement method for long-sequence SAR marine reclamation detection. Reclamation changes can be detected rapidly and accurately. After that, the demonstrative verification can be implemented in key areas across the country. The research project results provide a theoretical basis for the dynamic monitoring of reclamation by GF-3 satellite.

围填海是国家海域使用动态监测的重要类型，新型GF-3卫星可提供长序列极化SAR数据，利于表征多种围填海类型后向散射特性，为实现随机海况多种围填海精确检测提供契机。针对随机海况下GF-3多种围填海类型成像特性未明确、围填海信息淹没于相干斑噪声、经典识别模型无法在同景遥感影像不同海域或不同影像同一海域中精确检测围填海变化的问题，本课题以随机海况下多种围填海类型后向散射特性分析为突破口，开展如下研究：（1）通过随机海况下ISAR成像模拟仿真实验，获得极化特征，揭示多种围填海类型后向散射特性规律；（2）结合成像特性构建海岸线和典型围填海类型快速检测模型，实现围填海区域快速定位；（3）构建基于L1,2范数的权重聚类非局部堆栈自编码模型，应用基于决策级判定的长序列SAR影像围填海检测方法，快速精确检测围填海变化，并在全国重点区域进行示范验证。本课题研究成果对实现围填海GF-3卫星动态监测提供理论依据。

大规模的围填海是一种严重改变海域自然属性的用海行为，但是因为沿海地区常为多云阴雨天气，高分辨率多光谱遥感数据有效覆盖频率受到极大限制。本项目基于新型国产GF-3号极化SAR卫星遥感影像，开展围填海散射机制分析、海岸线检测方法研究、变化检测方法研究和海上目标检测方法研究，主要分为以下四个方面。一是分析了5种围填海类型，建立了不同围填海方式在高分辨率真彩色影像和SAR影像上的遥感解译标志。二是提出了基于区域距离正则化几何主动轮廓模型，引入距离正则项，将区域面积项系数与SAR图像等效视数建立非线性拟合关系，实现了模型的自适应调整；提出了改进的距离正则化水平集演化方法的海岸线变化检测，引入基于后验概率权重的指数边缘停止函数，同时考虑每次迭代的目标后验概率和指数函数的性质来加速曲线在背景区域的演化，结合遥感图像各个波段的特性，采用遥感图像的局部熵来初始化改进的水平集演化方法；三是提出了基于协同聚类和权重注意力稀疏编码网络的变化检测方法，将注意力模块加入到稀疏自编码网络的每一层中，利用模块分别对网络的正、负权重进行特征压缩，随后将压缩后的特征经过非线性单元归一化并与正、负权重加权；提出了基于权重聚类稀疏自编码网络的联合变化检测方法，通过逐层合并相似的权重使网络学习样本精简的特征，引入 L1/2 范数约束来提高网络特征选择能力并降低过拟合问题影响；对2010－2020年马来西亚马登红树林进行遥感监测，使用土地利用转移矩阵和景观格局指数对马登红树林进行评价。四是设计了LCSE-ResNet的海上石油平台和船只检测方法，在残差网络中加入了高斯拉普拉斯算子和连通域控制子，对遥感影像进行线性拉伸与下采样等预处理，利用海上目标的像素值大于周围环境的特点，提取出疑似目标，再对疑似目标进行分类。本项目研究成果对海南日月岛围填海变化进行了联合无人机影像的现场验证和应用，具有较好的识别精度，从而为GF-3号极化数据的围填海卫星遥感监测提供重要理论依据。