街景图片识别城市贫困的适用性——基于广州市中心城区的验证

以广州市为例，选取中心4个区6 670个采样点（涵盖121个社区）的百度街景图片，从城市建成环境特征探讨了城市贫困识别的可能。首先，训练基于深度神经网络的街景图片分类模型后，对街景要素进行语义分割，并通过缓冲区分析统计社区尺度的街景指标；其次，经主成分分析法提取出建筑围合感、植被围合感、天空开阔感和道路开阔感4个街景主因子，并验证其与多重贫困指数（IMD）的相关性；最后，通过采用简单随机抽样法选取61个社区，构建街景预测的多元线性回归模型，对剩余60个社区进行贫困预测，验证街景指标测度城市贫困的度量精度。结果发现，案例社区的多重贫困指数（IMD）与建筑围合感呈正相关，与植被围合感、天空开阔感、道路开阔感呈负相关；从整体看来，街景预测结果与传统城市贫困测度的空间规律基本相符，而且结果通常比传统测度的城市贫困程度高。这是因为受测度内容、社区类型、街道属性等方面的影响，街景识别方式比较适用于判断建成环境较差的贫困社区。街景图片预测有利于刻画城市贫困人群真实的生活环境，便于对城市建成区进行及时监测，在一定程度上可以与传统城市贫困测度相互校正、弥补不足。

The Applicability of Street View Images to Identify Urban Poverty in the Central Urban Region of Guangzhou

Taking Guangzhou as an example, Baidu street view images of 6670 sample points in the four central districts of the city were selected as test data to explore the possibility of urban poverty measurement in an urban built environment. Firstly, after training the street view picture-classified model based on FCN-8s deep neural network, the street view features were semantically divided into categories such as buildings, sky, trees, and so on. The street view indicators of communities were calculated by buffer analysis. Four street view main factors including building enclosure, vegetation enclosure, sky openness, and road openness, were extracted by principal component analysis; the correlation between them and multiple poverty index (IMD) was verified. Finally, the multi-linear regression model to estimate the poverty level, which is based on street view indicators, was constructed from 61 communities through a simple random sampling method. The prediction model was used for the remaining 60 communities in order to verify the accuracy of the street view measurement. It was found that IMD is positively correlated with building enclosure, and negatively correlated with vegetation enclosure, sky openness, and road openness. In addition, the measurement results from street view were basically consistent with the spatial patterns of traditional urban poverty measurement, which were usually higher than the poverty level in traditional poverty measurement. This is because the two methods produce different results that are affected by the measurement content, community type, street attributes, and other factors. For example, street view can depict the real living environment of the poor urban people, which is more suitable for poor communities that occupy terribly built environments, such as crowded and dilapidated old neighborhoods, urban villages with serious construction violations, rural homes, and under-construction industrial parks and so on. On the other hand, the traditional IMD measurement method is more suitable for those areas where the external and surrounding built environments are not very bad, such as redeveloped constructions that replaced decaying structures, affordable housing with pleasing environments, and so on. To a certain extent, the two methods can complement each other, and street view can provide timely monitoring of changes in urban built-up areas. It can also help poor communities put forward concrete proposals to effectively improve their living environment, in terms of physical amenities, such as roads, buildings, and green spaces. In this paper, a series of street view indicators related to urban poverty were extracted by analyzing the street view data of communities in the four central districts of Guangzhou. Based on these indicators, a model of urban poverty was successfully constructed. Findings about the characteristics of street view, factors influencing street view measurement, and the scope of application when compared with the traditional urban poverty method are presented. The use of an urban poverty measurement technology with high precision, wide coverage, fast updating, and information-rich streetscapes can benefit and refine urban poverty research and enrich the dimension of urban poverty measurement indicators. Such measurement also has practical significance for the redevelopment of decrepit structures and areas inhabited by poor communities.