高分辨率极化SAR图像场景分类研究

High-resolution polarimetric SAR (PolSAR) image classification is a leading topic on the frontiers of interpreting SAR images. Compared to low-resolution ones, high resolution remote sensing provides a richer scene and target information (such as the geometric structure of the targets, etc.) for Earth observation, but at the same time raises new problems and challenges for the interpretation of images. This project focuses on the geometric structure feature extraction and scene classification of high-resolution PolSAR images, in order to meet the needs of the development of high-resolution PolSAR sensors and their applications. First, to address the inadequacies of traditional polarimetric scattering features for describing the high-resolution PolSAR images, a geometric feature extraction scheme will be studied based on statistical and morphologic representations, which shall provide an effective description of the geometric structures of images. Second,it is worth noticing that overwhelming quantities of high-resolution PolSAR data are now available and support accurate earth observations and topographic measurements, but traditional statistical learning methods require a large amount of densely-labeled training data to produce an effective terrain classifier. Such labeling is very labor-intensive and it typically requires expert knowledge to compensate phenomena such as speckles. Therefore, a weakly-supervised classification method will be investigated, which can directly learn category models from key-words labeled training data by combining the topic models and tree based hierarchical random fields. It can improve the efficiency of the entire classification process while ensuring the classification accuracy. Finally, in the case of training data is completely missing, an unsupervised PolSAR image classification algorithm is designed to automatically select informative features for each category and determine the number of categories, by integrating information projection theory and cluster sampling algorithm into a unified bayesian framework. The research results of this project will make significant progresses on the geometric structure feature extraction, scene classification model and algorithm, and provide technical support for automatic interpretation and the large-scale application of PolSAR images.

高分辨率极化SAR图像场景分类是SAR图像解译的前沿课题。本项目紧密结合高分辨率极化SAR传感器的发展与应用需求，开展高分辨率极化SAR图像场景分类算法研究。首先，研究基于统计和形态学表达的几何结构特征提取方法，以弥补传统极化特征的不足，为高分辨率极化SAR图像提供更有效的地物几何结构描述。其次，针对高分辨率极化SAR图像数据量大、标注像素级训练数据困难的情况，研究新的高分辨率极化SAR图像场景分类模型和算法。具体来说，一方面，结合隐语义模型和树结构分级随机场模型，利用仅有关键词标注的训练数据，研究弱监督分类算法；另一方面，在训练数据完全缺失的情况下，基于信息投影理论和聚类采样算法，研究能同时完成自动特征选择和自动确定地物类别数目的非监督分类算法。本项目研究可望在高分辨率极化SAR图像几何结构特征提取，场景分类模型与算法上取得显著进展，为高分辨率极化SAR图像自动解译和应用提供技术支撑。

全极化合成孔径雷达(PolSAR)通过不同极化方式的组合对地物目标进行全极化测量，记录了地物目标的完整散射矩阵，实现了对目标特性更为全面的描述，极大地增强了成像雷达对目标信息的获取能力，已成为资源普查、城市规划、环境监测以及灾害评估的重要手段，在国民经济和军事领域中都有着重要应用。场景分类是PolSAR应用研究中的热点问题，研究和挖掘PolSAR影像包含的极化特征信息，发展高可信的地物识别分类方法，可以推动PolSAR遥感技术在各行业领域应用，具有重要的理论意义和应用价值。.在极化SAR 数据资源日益丰富的同时，如何从极化SAR 图像获取更多的本质信息并有效的“表示”出来，进行自动的“学习”和推理，完成极化SAR 图像的分类解译，这是当前极化SAR 图像解译系统的迫切要求。本项目针对高分辨率极化SAR 图像场景分类问题，重点围绕其中所涉及的极化特征提取与表达，监督和非监督分类模型与算法展开研究。在特征提取与表达方面，为了弥补传统极化特征的不足，研究了基于统计和形态学表达的几何结构特征提取方法，提出了“极化-纹理-结构”综合特征描述子和集成投影统计特征描述子。在此基础上，结合集成学习和聚类森林，提出了有效的特征选择与组合方法。在监督分类方面，针对图像在标记样本不足情况下的应用问题，提出了结合多尺度超像素表达，最大熵隐语义模型和分级随机场模型的弱监督分类算法，在保证分类精度的同时提高了分类过程的效率。提出了只有正样本标注情况下的目标提取问题，利用 PU-Learning将只有正样本情况下的目标提取问题转换为标准二分类器学习问题，实现了改进的“两步策略”算法，在PolSAR图像建筑区提取中获得了较好的效果。在非监督分类方面，针对极化SAR数据协方差矩阵的黎曼流形结构，引入黎曼稀疏编码框架，提出了一种基于黎曼稀疏编码和稀疏诱导相似度的非监督分类方法；通过引入混合Wishart分布模型和Cauchy-Schwarz散度以精确度量PolSAR图像的统计相似性，并结合空间上下文信息，进一步提高了分类效果。.本项目研究在高分辨率极化SAR图像的特征提取与表达、分类模型与算法上取得了显著进展，为高分辨率极化SAR图像自动解译和应用提供了有力的技术支撑｡