基于稀疏概率图模型的高分辨率遥感影像场景语义理解方法研究

The interpretation way has been changed from the pixel-based processing to the object-based processing for high resolution remote sensing imagery, which leads to the significant improvement of the interpretation accuracy. However, the interpretation level has only located in the ground object level, and it is hard to acquire the interested semantic information in the scenes by the object-based processing methods. It is urgent to develop the theory and algorithms for the scene understanding. Most of the existed methods of scene understanding for high resolution remote sensing imagery are derived from the study of natural image scene understanding. Because they do not consider the variability, changeability and distribution complexity of the ground objects in the high resolution remote sensing scene, there are the interpretation problems, such as the limitation of the learning ability for the low-level features, the redundancy of mid-level features and the lack of the high-level semantic information from the high resolution remote sensing imagery. This project will combine the unified modeling ability of the probabilistic graphical model and the essential representing ability of the sparse representation theory to propose the research main line from sparse learning of the low-level features and topic representation of the mid-level features, to the deep understanding of high-level scene semantics. Based on the main line, the project will build the over-completed dictionary for the multiple low-level features and develop the sparse probabilistic topic models to extract the mid-level features with low redundancy for the scene representation. The theory and methods of the scene understanding based on sparse probabilistic graphical models for high resolution remote sensing imagery will be built to dig out the key objects and the spatial semantic relationship between the objects in the scenes. This study can bridge the semantic gap between the low-level feature and the high-level scene semantic for high resolution remote sensing imagery, and significantly enhance the potential applications of high resolution remote sensing imagery, therefore have important theoretical and applied significance.

高分辨率遥感影像解译方式已基本实现了从面向像素到面向对象的转变，解译精度得到很大的提高。然而其解译层次只到地物类别层，难以获取人们所感兴趣的场景语义信息，亟需发展场景理解的理论与方法。现有高分辨率遥感场景理解方法大都来自于自然图像场景理解领域，而未考虑遥感场景特有的地物多样性、可变性、分布复杂等特点，存在底层特征学习能力不足、中层特征过于冗余、高层语义信息缺乏等问题。本项目拟综合概率图理论的统一建模能力和稀疏表达理论对高维数据的本质表达能力，以“底层特征稀疏学习—中层特征主题表达—高层语义深度理解”为主线，构造底层多特征稀疏字典，提出冗余度低的中层特征稀疏概率主题表达模型，建立基于稀疏概率图模型的高分辨率遥感影像场景语义理解理论与方法，揭示场景内所蕴含的关键目标及目标间的空间语义关系，成果将跨越底层特征与高层语义间的“语义鸿沟”，提升高分辨率遥感影像的应用潜力，具有重要的理论与应用意义。

针对高分辨率遥感场景理解存在底层特征学习能力不足、中层特征过于冗余、高层语义信息缺乏等问题，本项目基于概率图、稀疏表达和深度学习等理论，以“底层多特征学习—中层特征主题表达—高层语义深度理解”为主线，系统开展了高分辨率遥感影像“特征—目标—场景”语义理解的理论与方法研究，取得的代表性成果有：（1）构造了全局空谱特征学习、社会经济特征提取等多特征学习方法，提供了过完备底层特征；（2）提出了基于自适应稀疏深度语义、多特征融合主题模型等中层特征表达方法，实现了区分度高、冗余度低的复杂场景中层特征表达；（3）建立了多尺度卷积神经网络、点线面目标场景一体化理解等高分辨率遥感影像场景语义理解理论与方法，跨越了底层特征与高层语义间的“语义鸿沟”；（4）基于理论成果，研发了高分辨率遥感影像场景语义理解原型系统，应用于极光场景分类、城市土地利用制图、灾害应急等方面。在项目资助下，项目组共发表论文70篇，其中在“Remote Sensing of Environment”、“ISPRS Journal of Photogrammetry and Remote Sensing”、 “IEEE Transactions on Geoscience and Remote Sensing ”等国际高水平期刊上发表/接收论文55篇（其中一区论文13篇，二区论文36篇），在国内学报和国际会议上发表论文15篇；申请发明专利2项；在人才培养方面，项目组成员1人被聘为副教授；2人被聘为副研究员；毕业博士生6人，硕士生10人，获测绘科技进步一等奖、第十五届中国青年科技奖、美国摄影测量与遥感协会（ASPRS）约翰戴维森主席奖一等奖、IEEE 地球科学与遥感协会（GRSS）高光谱视频目标跟踪竞赛冠军等奖励。