城市结构类型是邻域尺度上城市地表结构的特定空间模式，反应了每个城市特有的结构、功能特征。使用场景分类的方法能提取出遥感影像中地物的特征和空间组织模式，获取具有一定自然地理或社会人文经济属性的城市结构类型分布信息，为可持续发展的城市规划和决策提供参考。

本文以北京市平原区为研究区，以图像场景为基本单元，定义城市结构类型的场景分类体系，构建了基于CNN和迁移学习的城市结构类型的场景分类方法。通过分析各城市结构类型的空间分布特征，探究北京市平原区的发展格局。本文的主要工作及成果如下：

1、确定了基于中高分辨率遥感影像城市结构类型场景分类的最优场景单元大小，以及相应的城市结构类型分类体系。通过ESP尺度评价法评估不同分割尺度下影像的内部标准差，确定300m（50×50像元）是SPOT6遥感影像进行城市结构类型划分的最优场景单元大小。在前人研究的基础上，依据城市功能和使用性质的差异，给出了基于场景的北京市平原区城市结构类型分类体系，并通过对各城市结构类型的特征进行定量描述分析其可分性，说明城市结构类型的定义具有一定科学性。

2、开展了基于CNN场景分类和迁移学习的城市结构类型提取方法研究。主要工作包括：（1）选择目前较流行的基于大型数据集ImageNet训练的CNN模型，在研究区的城市结构类型数据集上分别利用特征提取和微调两种迁移模式进行比较实验。实验表明，利用微调将在ImageNet上训练得到的InceptionV3网络模型迁移到城市结构类型场景分类中是最佳策略。2019年城市结构类型测试集总体分类准确率为88.45%，Kappa系为0.87，与城市发展密切相关的高密度建筑区、中密度建筑区、仓储及工业区、设施农业、空地、机场等城市结构类型分类效果都较好，验证了基于CNN和迁移学习的城市结构类型场景分类方法在城市动态监测中的可行性。（2）通过不同场景单元大小分类结果的精度分析，探究了城市结构类型在场景分类时的尺度敏感性。结果表明，高密度建筑区、中密度建筑区、仓储及工业区、设施农业、矿地、机场等城市结构类型对场景单元大小较为敏感，语义依赖一定场景单元大小下的整体特征。（3）选择2014年SPOT5影像进行场景分类方法的可移植性分析。

3、通过对城市结构类型空间聚集性和协同性等空间分布模式研究，揭示出北京市平原区的多中心组团发展模式以及城市环状、放射状的发展格局。利用DBSCAN聚类方法提取出中密度建筑区的聚集区，除主城区外，外围分布了70个相对集中的分布区；新兴城区围绕老城区环状蔓延并呈放射状延伸，城市结构呈现环状分布的空间格局；五环外的郊区地带高密度建筑区与仓储及工业区关联分布，且主要集中在顺义区和大兴区。

Urban structure types (USTs) are specific spatial patterns of urban structure at the neighborhood scale, reflecting the unique structural and functional characteristics of each city. Using the method of scene classification can extract the features and spatial organization patterns of ground objects in remote sensing images, and obtain the distribution information of urban structure types with certain physical geography or socio-cultural and economic attributes, and provide reference for urban planning and decision-making of sustainable development.

This paper takes the plain area of Beijing as the research area, and takes the scene as the basic unit to define the classification system of the urban structure types, and constructs a scene classification method for urban structure types based on CNN and transfer learning. By analyzing the spatial distribution characteristics of each urban structure type, we explore the development pattern of the plain area of Beijing. The main work and results of this paper are as follows:

1. Determine the optimal scene unit size for scene classification based on high or medium resolution remote sensing images, and the corresponding urban structure types classification system. The internal standard deviation of images under different segmentation scales is evaluated by ESP method, and 300m (50 × 50 pixels) is the optimal scene unit for urban structure types classification based on SPOT6 remote sensing image. On the basis of previous studies, according to the differences in urban functions and usage properties, a scene-based classification system of urban structure types in the plain area of Beijing is presented, and the divisibility is analyzed by quantitatively describing the characteristics of each urban structure type, indicating that the definition of urban structure types is scientific.

2. Research on the extraction method of urban structure types based on CNN scene classification and transfer learning. The main work includes: (1) Select the currently popular CNN model trained on the ImageNet dataset, and use feature extraction and fine-tuning on the urban structure types dataset in the study area to conduct comparative experiments. Experiments show that using fine-tuning to transfer the InceptionV3 network model trained on ImageNet to urban structure types scene classification is the best strategy. The overall classification accuracy of the 2019 urban structure types test set is 88.45%, and the Kappa system is 0.87. The classification results of high-density building area, medium-density building area, storage and industrial area, facility agriculture, open space, and airports are good, and these types are closely related to urban development, which verifies the feasibility of the urban structure types scene classification method based on CNN and transfer learning in urban dynamic monitoring. (2) Through the accuracy analysis of the classification results of different scene sizes, the scale sensitivity of urban structure types in scene classification is explored. The results show that urban structure types such as high-density building area, medium-density building area, storage and industrial area, facility agriculture, mine, and airport are more sensitive to the scene size, and the semantics depend on the overall characteristics of a certain scene size. (3) The 2014 SPOT5 remote sensing image was selected for transferability analysis of the scene classification method.

3. By studying the spatial distribution patterns of urban structure types, such as spatial aggregation and co-location, it reveals the multi-center cluster development pattern and the urban  ring-like and a radial diffusion patterns in the plain area of Beijing. DBSCAN clustering method is used to extract the clustering areas of medium-density building area. In addition to the main urban area, 70 relatively concentrated distribution areas are distributed in the urban periphery; The new urban zone spreads around the old urban zone in a ring shape and extends radially, and the urban structure presents a ring distribution; High-density building areas in suburban areas outside the Fifth Ring Road are associated with storage and industrial areas, and are mainly concentrated in Shunyi District and Daxing District.