基于深度学习街景影像解译和景感生态学的视域环境定量解读

城市物理环境带给居民丰富而生动的视觉意象，目前许多文献结果表明其宜人性与公共福祉以及健康状况息息相关。景感生态为探究城市物理环境与居民感知信息之间的联系提供了指导依据，通过人本尺度的定量手段解读城市环境中视觉、听觉和味觉等多维度感知信息。秉承景感生态学的基本原理，引入一种结合街景大数据和深度学习的城市环境量化手段，以北京市六环范围为例，将景感视率作为测度对人本视角下的城市环境展开定量解读。在全面把控多维景感要素的同时，旨在实现以人为本的城市物理环境优化设计，从而满足人们对生活品质提升的实际需求。实验结果显示：（1）从视觉感受的宏观表现来看，北京四环路范围内建成环境的"闭合感"较强，而对绿植的感知程度相对偏弱，因而需要开展存量环境设计并优化视域界面结构；（2）以景感视率作为特征值进行聚类得出3类主导空间（绿色空间、灰色空间和蓝色空间），可针对灰色空间着重部署垂直绿化资源，提高城市视觉绿化的可感知性，从而营造舒适宜人的绿色氛围、促进公众身心健康；（3）为景感生态学提供了基于大数据思维的数据集和定量方法补充。综上，以街景影像和景感生态视角对北京市中心城区的视域环境展开定量分析，采用了先进的深度学习框架（Detectron2）并结合经典的机器学习方法（K-Means）对人本视域内多维景感要素的空间分布特征进行解读。借助景感生态规划可以有针对性的改善城市视域界面的感知质量，提升智能管理水平，帮助城市规划设计人员和管理者从人本视角提升城市公共环境品质和风貌。

Quantitative interpretation of visual environment based on street view images analyzed with deep learning and from Lendsenses Ecology perspective

Urban physical environment brings rich and vivid visual image to the residents, and plenty of quantitative research results have shown that the amenity is closely bound up with urban public well-being and health conditions. Landsenses Ecology provides a new guidance for exploring the connection between physical environment and residents'' perception information, and interprets multi-dimensional senses such as vision, auditory and gustation in urban environment through quantitative measures in the human scale. This article follows the basic principles of Landsenses Ecology, introduces a city environment quantification method that combines a street view dataset and deep learning framework. Taking the Sixth Ring Road area of Beijing as an example, the Landsenses View Factors (LVFs) are used as metrics to explain the urban environment in a human-oriented perspective. While comprehensively controlling the multi-dimensional landsenses factors, our goal is to realize the people-oriented optimization design of urban physical environment, so as to meet the actual needs of residents to improve the quality of life. The experimental results show that: (1) from the perspective of visual perception at the macro level, the "closedness" of building environment within the Fourth Ring Road in Beijing is relatively strong, and the perceptibility of green view is relatively weak. It means that environmental designs in stock should be deployed and the components of visual interface should be optimized in the Fourth Ring Road area. (2) Clustering with the LVFs as the feature variables yields three types of dominant spaces (green space, gray space and blue space, separately). The vertical green infrastructures can be concentratedly deployed on the "gray space" to improve the perceptibility of urban green view at eye level, thereby creating a comfortable and pleasant green atmosphere, and promoting physical and mental health of the public. (3) This article provides data and method supplement based on big data thinking for Landsenses Ecology. In summary, this article analyzes the urban space of Beijing''s central area from the perspective of Street view images (SVIs) and Landsenses Ecology, using the state-of-the-art deep learning framework named Detectron2 combined with a conventional machine learning model (K-Means clustering algorithm) to interpret the spatial distribution characteristics of the multiple LVFs in the humanistic perspective. Taking advantage of landsenses ecological planning, the perception quality of urban visual interface and the level of intelligent management can be improved, which helps the urban planners and managers to improve the quality and aesthetic of urban public environment from the human scale.