顾及先验地理信息的城市空间结构类型深度特征表达及识别

Urban spatial-structural types are used to describe the spatial information in neighborhood space, can well and truly reflect complex land cover condition of urban area. In order to solve the problems such as information expressing not enough and weak recognition capability widely existing among traditional neighborhood information extraction methods developed on pixel-wise or object-oriented unit, high-spatial-resolution imagery is taken as the main data source in this project, and to research the key theories and techniques in urban spatial-structural type recognition based on scene analysis unit. Specifically, the main research contents include: (1) VGI data is employed as a constraint condition to dynamically generate spatial-structural scene images; (2) focusing on widely existing challenges of designing models and limited training samples in applying deep convolutional neural network model to a real task, on the basis of transfer learning, fine-tuning and metric learning, to improve the application performance of deep convolutional neural network model in scene image feature representation via developing multiscale, multilayer and multi-model feature fusion methods; (3) to construct urban spatial-structural type recognition technological process and method based on scene analysis unit. The achievement of this project can provide technical support or reference to extract neighborhood space information, and promote the study of new method in analysis and interpretation for high-spatial-resolution remotely sensed imagery, further extend the application of high-spatial-resolution remotely sensed imagery in urban area.

城市空间结构类型表征城市邻域范围空间信息，能够更准确反映城市区域复杂的地表覆盖空间分布。现有面向像元或对象单元的邻域空间信息提取方法在城市空间结构类型识别中存在信息表达不足、识别能力差等问题。项目以高分辨率遥感图像为数据源，研究面向场景分析单元城市空间结构信息识别的相关理论和关键技术。主要研究内容包括：（1）引入VGI数据建立先验地理信息约束下的空间结构场景动态生成方法；（2）针对卷积神经网络模型在实际应用中存在模型设计困难和样本不足问题，基于模型迁移、模型微调和度量学习等策略，通过构建多尺度、多层次、多模型特征融合表达方法来提高深度卷积神经网络模型在遥感图像场景特征表达中的应用效果；（3）构建面向场景分析单元的城市空间结构类型识别技术流程和方法。项目研究成果可为邻域范围空间信息提取提供技术支持或参考，推动高分辨率遥感图像分析与信息解译新方法研究，拓展高分辨率遥感图像在城市研究领域的应用。

城市空间结构类型表征城市邻域空间所呈现的空间信息，能够更为综合的反映城市区域地表覆盖空间分布状况。项目突破传统基于像元或面向对象空间信息表达不足及理解能力差等问题，开展了先验地理信息支持下的城市空间结构类型解译单元定义和生成机制研究，以场景理解为技术手段的空间结构类型场景图像特征学习、特征综合及分类识别方法研究。经过三年时间的探索研究，（1）构建了在VGI或土地利用调查空间矢量数据已知情况下城市空间结构类型场景单元动态生成机制，可以针对线状或面状目标自动生成多尺度场景图像；（2）针对场景理解问题，以深度学习为主要技术手段，发展了基于迁移模型、改进双线性模型、二阶池化模型的深度特征学习及表达方法，显著提高了场景理解的准确度；（3）进一步扩展没有先验地理信息情况下城市空间结构类型场景单元生成模式，提出超像素分割引导下的场景图像生成机制，并在此基础上发展了融合不同层次特征的场景内容表达方法，提升了算法在城市空间结构类型分类识别应用中的适应性和准确性；（4）将提出的算法模型应用于国家土地第三次调查土地利用类型自动识别及图斑检查中，取得了较好的应用效果。