基于地理景观层次模型和区域统计建模的高分辨率遥感影像多尺度场景分类

With the appearance of an increasing amount of high spatial resolution sensors, geographic object-based image analysis (GEOBIA) represents the development trend of the information extraction in remotely sensed image of high spatial resolution, but facing the problems of the scene interpretation, semantic gap, and so on. According to the hierarchical structure of ground objects cognition, this is from low-level vision to high-level semantic, which is “pixel-image object-geographic object- geographic scene” hierarchical structure. Based on the framework of geographic landscape hierarchical model of multi-scale regional statistical-modeling, this project will research the high-level semantic automatic interpretation of high resolution remote sensing image. Firstly, based on knowledge of geography and ecological landscape, constructing the geographic landscape hierarchical model and modeling and expressing the geographic object feature based. Second, the segmentation method is adopted, which fusing spectrum, texture and structure feature with the boundary constraint for high resolution remote sensing image, and the optimal scale segmentation results are obtained. Thirdly, based on the GEOBIA framework and geographic object feature based, the multifeature probabilistic latent semantic analysis (MPLSA) method is adopted to recognizing geographic object. Finally, based on the geographic landscape hierarchical model and multi-scale regional MRF model framework, the multi-hierarchical regional merging algorithm guided by geographical landscape semantic information is researched, which obtain multi-scale scene classification results for high resolution remote sensing image. This research can provide a new idea of automatic interpretation of multi-scale high remote sensing image scene and has important theoretical and application value.

随着国内外大量高分辨率传感器的出现，面向地理对象的影像分析方法成为遥感影像信息提取的发展趋势，但面临着场景解译和语义鸿沟等问题。本课题将根据地物认知从底层视觉到高层语义的“像素—影像对象—地理目标—地理场景”的层次结构，在基于地理景观层次模型的多尺度区域统计建模的框架下，研究高分辨率遥感影像高层语义自动解译问题。首先，基于地理学和生态景观学知识，构建地理景观层次模型和地理目标特征库的表达和建模；然后，采用融合光谱、纹理和结构特征的边界约束的高分辨率遥感影像分割算法，获得最优尺度分割结果；在面向对象的框架下及地理目标特征库支持下，采用多特征概率潜在语义分析方法（MPLSA）进行地理目标识别；最后，根据地理景观层次模型，在多尺度区域MRF模型框架下研究地理景观层次语义信息指导的多层次区域合并算法，实现遥感影像多尺度场景分类。项目研究将为遥感影像自动解译提供一条新思路，具有重要的理论和应用价值。

针对基于传统OBIA的高分辨率遥感影像信息提取方法存在语义鸿沟等问题，项目根据地物认知从底层视觉到高层语义的“像素—影像对象—地理目标—地理场景”的层次结构，构建高分辨率遥感影像多尺度场景解译框架。主要研究包括：（1）提出了融合地学指数特征和边界约束的多尺度分割算法，实现了影像对象的最优分割；（2）基于生态学和地理学知识，实现了典型地理场景的地理目标特征库的表达和建模，并构建了相应的地理景观层次模型；（3）基于地理目标特征库，采用稀疏表达、深度学习等机器学习方法实现典型地理目标高精度识别；（4）在地理景观层次模型指导下，实现村庄和城市场景多尺度智能解译。针对上述提出的理论和算法，采用国内外高分辨率遥感影像数据进行实验验证，研究结果表明该项目提出的影像对象分割、地物目标识别和场景解译算法的精度优于传统的高分辨率遥感影像智能解译算法。项目研究成果将丰富高分辨率遥感影像智能解译的理论和方法，并能应用于智慧城市和资源环境监测。按照既定的研究计划，项目圆满地完成了课题设定的研究计划，实现了预期目标。在项目资助下，本研究共发表学术论文7篇，其中SCI论文3篇论文，EI论文4篇；撰写1部专著（即将出版）；在人才培养方面，项目负责人入选云南省“万人计划”青年拔尖专项人才计划，培养了博士生1名、硕士生6名。