Due to the acceleration of urbanization process in modern society, there are many metropolises throughout the world, such as New York, Tokyo, and Shanghai. Such large cities typically have a carefully designed space planning. For example, the residential area should be distant from the industrial area due to the potential environmental pollution issue. In order to cope with public health emergencies, the infrastructure needed for modern urban governance needs to be improved. Notably, establishing a quality visual model to exploit such sophisticated spatial configurations is an important but challenging task. Such a model can facilitate many applications such as urban planning, environmental evaluation, smart transportation, and urban governance. However, the flexible spatial interactions among multiple regions make it difficult to apply a traditional visual model to encode them. In this work, a quality-guided feature selection framework is proposed to obtain a set of high-quality topologies to model the discriminative structures from different land spatial city regions. Given a city region from a metropolitan area, the well-known super pixel algorithm SLIC is used to decompose each land spatial city region image into multiple atomic regions. Based on this, a binary graph is used to model the spatial interactions among these regions. Each binary graph is then decomposed into multiple subgraphs and a topology selection algorithm is proposed to discover subgraphs with highly discriminative topologies. By leveraging these high-quality subgraphs, the image kernel machine is used to convert the high quality subgraphs from each city region’s image into a feature vector. Afterward, a multi-category support vector machine (SVM) is learned to classify each city region’s image into one particular category. Comprehensive experimental results by comparing with many state-of-the-art have shown the competitiveness of this method. Furthermore, the selected high-quality topologies have demonstrated that highly representative spatial interactions are nicely encoded. The results are of great significance to the establishment of the land spatial database management system.

由于现代社会中城市化进程的加速，世界上有许多大都市，例如纽约，东京和上海。这样的大城市通常都有精心设计的空间规划。例如，由于潜在的环境污染问题，住宅区应远离工业区。为了应对突发公共卫生事件，需要改善现代城市治理所需的基础设施。值得注意的是，建立高质量的视觉模型来利用这种复杂的空间配置是一项重要但具有挑战性的任务。这种模型可以促进许多应用，例如城市规划，环境评估，智能交通和城市治理。然而，多个区域之间灵活的空间互动使得难以应用传统的视觉模型对其进行编码。在这项工作中，提出了质量指导的特征选择框架，以获取一组高质量的拓扑，以对来自不同土地空间城市区域的判别结构进行建模。给定一个大都市区域的城市区域，众所周知的超像素算法SLIC用于将每个陆地空间城市区域图像分解为多个原子区域。基于此，使用二元图对这些区域之间的空间相互作用进行建模。然后将每个二元图分解为多个子图，并提出一种拓扑选择算法以发现具有高度区分性的拓扑的子图。通过利用这些高质量的子图，图像内核机用于将每个城市区域的图像中的高质量子图转换为特征向量。之后，学习多类别支持向量机（SVM），将每个城市区域的图像分类为一个特定类别。通过与许多最新技术进行比较的综合实验结果证明了该方法的竞争力。此外，所选的高质量拓扑已证明可以很好地编码具有高度代表性的空间交互。研究结果对建立土地空间数据库管理系统具有重要意义。通过与许多最新技术进行比较的综合实验结果证明了该方法的竞争力。此外，所选的高质量拓扑已证明可以很好地编码具有高度代表性的空间交互。研究结果对建立土地空间数据库管理系统具有重要意义。通过与许多最新技术进行比较的综合实验结果证明了该方法的竞争力。此外，所选的高质量拓扑已证明可以很好地编码具有高度代表性的空间交互。研究结果对建立土地空间数据库管理系统具有重要意义。