基于语义网络与面向对象的海岸带遥感图像中地物成像差异识别机制研究

More details of earth surface appeared in the high resolution remotely sensed imagery. The imaging differences between the parts of ground object result in several separate segments during the later multi-scale analysis. The reasoning and combination of neighbour segments become a challenge of image understanding in coastal remote sensing. The proposed project utilises semantic network and image object oriented approach for spatial imaging context information extraction in high resolution satellite imagery of coastal zone. The scene model is applied for imaging procedure analysis, which provide the imaging differences of the parts of special object and validated in the typical high resolution images. To express spatial knowledge, spatial semantic network model takes the images as the basic analysis object and semantic surrounding, analyzes its semantic feature of the space, attribute, and context information. The handling image is performed multi-scale analysis to yield hierarchic candidates firstly. And the meaningful image objects are identified and the feature extracted from them. The neighborshood among them are calculated by using Voronoi algorithm. Then the object-oriented graph forms which present the whole spatial location and topological relation. A novel spatial structure match algorithm is used to get a spanning tree between the image-object graph and semantic network, which constraint by the paths that have max spatial relationship regularity. The mapping result is used to identify the corresponding segments of the imaging parts of object finally. Hereby the scene imaging model is used to step over the difficult in the interpretation of coastal remote sensing images. The philosophy, methodology and related algorithms of this model are fully described in terms of basic ideas, model definitions, dynamic mechanisms, reasoning methods and implementation abilities. The proposed project can helped to improve cognitive model and machine intelligence in remote sensing.

高分辨率遥感影像中地面细节进一步凸显，地物内部的成像差异在后续的图像分割中呈现为具有光谱差异性的多个图斑，图斑间的组合推理成为认知海岸带地物类型的新挑战。本项申请提出基于语义网络和面向对象的海岸带高分辨率遥感影像中地物成像差异的识别机制和实现途径。通过应用海岸带典型地物的地表场景模型，构建图像对象层次的空间语义网络表达成像差异及其在图像上的光谱特征；待解译的遥感图像通过多尺度分析、图像对象识别、遥感特征提取，建立邻域空间关系，形成内含空间信息的图组织图像对象及其遥感特征；利用空间关系和光谱特征约束的图匹配算法，图像对象邻域关系图与空间语义网络间建立映射，以此识别地物对象的成像差异信息。本项研究试图在米级遥感影像解译中应用场景成像分析认知海岸带地物的成像差异，在遥感信息获取中引入可计算分析的知识表达和利用方式，探索遥感图像新的认知模式与机器理解机制。

我们开展适合高分辨率遥感影像的图像分割、对象特征提取与选择；从场景理解的角度,开展海岸带水体相关地物类型识别等技术研究；进而，实现了地理基于本体论的图像对象解译方法，以及利用图像空间上下文信息的技术途径。.针对海岸带地物尺度分布差异大的特点，发展利用边缘完整性确定各类地物的最优分割的方法，解决了分割对象边界与真实地物边界对应不一致的问题。其优势在于算法通过获得边缘完整度曲线中的极大值来确定对应的最优分割对象，强化了对象整体边界信息对分割结果的影响，代替了人为调节尺度参数来获取地物最优分割尺度。此外，边缘完整度描述的是对象边界与其周围边缘的关系，反映的是对象的整体边缘信息，由此实现确定影像中每一地物最优分割状态的目标。实验表明，该方法对于海岸带这种地物尺寸差异性明显区域具有较强的分割优势，对影像中存在弱边缘的地物也就有较好的分割效果。.结合海岸带的典型水体地物特点，开展了结合空间语义特征的水体类型识别研究。通过分析高分遥感影像中海岸带水体的内在特性建立提取水体信息的规则集，以分割后的图像对象为基础并将空间邻接的水体图斑进行组合形成有潜在意义的水体对象；根据水体地物间存在的关联特征建立位置、拓扑等知识规则，对水体进行细分为沿海海域、围海养殖池与内陆水体。研究表明，该方法摆脱了对对象特征进行统计分类的传统框架，利用先验知识构建加入空间语义信息的知识规则库，地物间的空间关系得到真正地充分利用，有效提高了海岸带地物类型识别精度，也更符合人对遥感影像的理解过程。.提出并实现了地理基于本体论的图像对象解译方法，并且针对上下文信息表达的问题，提出了结合多尺度分割图像对象与概念定义在相邻、内部组成和区域三方面表达上下文信息的方法和解译匹配过程。在海岸带高分辨率遥感图像中，基于本体论的遥感图像对象都优于监督分类。本项研究突出了上下文对提升图像解译上了重要地位，以及本体论在组织语义上的优越性。