基于语义理解的高分辨率遥感图像变化分析方法研究

With the increasing improvements in spatial resolution and temporal resolution, there is more and more semantic information contained in high resolution (HR) remote sensing images, so semantic understanding based HR remote sensing imagery change analysis becomes more and more important. However, the sophisticated imagery scene and various target semantic information cause the change detection for HR remote sensing images even more difficult compared with low/medium resolution images. This program tries to propose new algorithms for HR remote sensing imagery change information fine analysis on the foundation of image semantic understanding. To this end, we first thoroughly investigate the possible ways to define and describe the high-level object semantic information in the complex scene by making full use of the abundant image low-level visual features. Then, we build a hypergraph instead of a pairwise graph to model the various semantic relationships among the image objects. In this way, a semantic relation network is constructed by expressing the semantic relationships quantitatively. Moreover, to take advantages of the large quantity of unlabeled samples, semi-supervised graph-based learning is adopted and labels are iteratively propagated to the unlabeled samples on the graph guided by an energy function with a novel local linear neighborhood label consistency regularization term. Finally, to satisfy the great demand of fine image interpretation, this program plans to study the approaches for change information analysis in consider of both the physical level changes and the semantic level alterations. By combining these fine change detection results with the expert prior knowledge, object trend analysis and prediction methods are studied, which is useful for appreciating the intelligence value. Additionally, experiments are designed to test the effectiveness of the proposed algorithm and a demo system will be implemented by the support of this program.

随着遥感图像空间和时间分辨率的不断提升，图像中蕴含的语义信息越发丰富,基于语义理解的高分辨率遥感图像分析及应用的重要性日渐凸显。本项目针对高分辨率遥感图像视觉特征多样性和语义信息丰富性，从图像语义目标的视觉表征出发，分析目标对象底层视觉特征到高层语义信息的映射途径，重点研究高分辨率遥感图像中典型地物目标对象之间的语义约束关系表征方式；针对高分辨率遥感图像中单个目标语义信息的局限性和多个目标语义信息的复杂性，以场景语义理解为目的，研究基于超图模型的复杂语义约束关系网构建方法，并结合半监督学习方法，完成语义变化信息的主动传递；针对语义表达的模糊性和认知应用的精确性，面向认知理解，通过目标物理层变化信息与语义层次变化信息有机结合，实现对变化信息的精细化解译，并探索典型目标变化信息分析与专家知识推理协同模式，增强对目标动向情况的简单推理能力，提升高分辨率遥感图像变化分析的应用价值。

基于语义理解的高分辨率遥感图像变化分析着眼于高分辨率遥感图像实际应用需求，针对高分辨率遥感图像变化分析中变化的复杂性与歧义性问题，从变化特征表征出发，研究了变化特征之间的语义约束关系表征方法及从底层视觉特征到高层语义信息的映射途径，构建了基于超图模型的复杂语义约束关系网络模型，并结合半监督学习方法，设计了适应于网络模型的语义变化信息的主动传递模式，探索并初步开展了基于变化信息分析与专家知识推理相协同的典型目标精细化变化分析方法研究与试验验证，收集整理了20余个典型时间序列变化分析试验数据集，总结归纳了2类典型场景中目标变化分析专家知识经验，完成了面向语义场景理解的高分辨率遥感图像变化分析演示原型系统开发。课题执行期间，共发表和录用学术论文8篇，申请发明专利6项，其中授权专利3项。