基于精细化描述和选择集成学习的高分辨率PolSAR图像解译

Aiming at the computer interpretation of high-resolution PolSAR image and trying to simulate the process of visual interpretation, this study focuses on fine characterization of typical targets, effective learning algorithms and the incremental learning interpretation system. It includes: multi-scale fine characterization and features selection of high-resolution PolSAR image; the effective image interpretation algorithms based on generalized selected-ensemble learning; the interpretation system framework with incremental learning ability through knowledge feedback based on case-based reasoning. During the study, the selection strategy of active learning and the ensemble strategy of ensemble learning are the key issues in the high-resolution PolSAR image interpretation; while, the learning mechanism of case-based reasoning is a key scientific issue to construct a high-resolution PolSAR image interpretation system with incremental learning ability. By means of signal processing and artificial intelligence techniques, the study aims to dig the abundant information in the high-resolution PolSAR images and achieve a clear and complete understanding of targets and accurate interpretation of PolSAR images by using the computer. The research results can establish the basis for high-resolution PolSAR image computer interpretation, and provide beneficial theoretical and practical support for promoting the application of high-resolution PolSAR technology and high-resolution earth observation. Meanwhile, the study presents the theoretical value and scientific significance for the development and application of machine learning theory.

针对高分辨率PolSAR图像计算机解译的需求，试图模拟目视解译的过程，围绕目标精细化特征描述、有效学习算法以及可增量学习的解译体系开展相关研究，包括：高分辨率PolSAR图像多尺度精细化特征描述和选择；基于泛化选择集成学习的图像解译算法；基于案例推理构建可实现知识反馈增量学习的解译框架。其中，自主学习的选择策略和集成学习的集成策略是实现高分辨PolSAR图像解译的关键问题；而案例推理的学习机制是构建可增量学习高分辨率PolSAR图像解译系统的关键科学问题。本项目的研究通过信号处理手段和人工智能技术深入挖掘图像中的丰富信息，利用计算机实现对观测目标清晰完整的认识和图像内容的精确判读。研究成果可为高分辨率PolSAR图像计算机解译提供依据，对扩大高分辨率PolSAR技术的应用以及实现高分辨率对地精确观测具有重要的工程实用价值。同时对于机器学习理论的发展和应用拓展，具有现实的理论意义。

针对高分辨率PolSAR图像计算机解译的需求，本项目试图模拟目视解译的过程，围绕目标精细化特征描述、有效学习算法以及可增量学习的解译体系开展相关研究，主要内容包括：高分辨率PolSAR图像多尺度特征提取和精细化描述；基于泛化选择集成学习的图像解译算法；基于案例推理构建可实现知识反馈增量学习的解译框架。项目取得的成果如下：实现了典型目标的散射特性分析和实测数据验证，明晰了其散射机理和散射特征的关系，建立特征描述的数学表达与物理意义的关联；通过像素级、目标级和场景级的多尺度特征提取，实现了高分辨率PolSAR图像的精细化特征描述；针对PolSAR图像分类小样本问题的求解，将集成学习引入到PolSAR图像解译，实现了高分辨PolSAR图像的选择集成高精度解译，同时引入稀疏分解理论，提出了基于上下文稀疏表示的PolSAR图像分类模型；针对高分辨率PolSAR图像的精细复杂性导致的类内可分性增加及类间可分性降低问题，将超像素分割引入到高分辨PolSAR图像处理中，实现了PolSAR图像的超像素分割；构建了基于半监督案例推理的高分辨率PolSAR图像计算机解译框架，将针对某个数据或应用的成功经验以案例的形式存储起来，实现了计算机解译系统的增量学习和经验积累。研究成果发表在IEEE TRGS，IEEE JSTARS和IGARSS等刊物和国际会议上，共15篇，申请国家发明专利4项。研究成果可为高分辨率PolSAR图像计算机解译提供依据，对扩大高分辨率PolSAR技术的应用以及实现高分辨率对地精确观测具有重要的工程实用价值。