基于贝叶斯逻辑回归模型研究百度街景图像微观建成环境因素对街面犯罪的影响

街面犯罪对公众的生活安全构成一定的威胁。以往对于公共盗窃和寻衅滋事等街面犯罪的研究往往停留在社区甚至更宏观的层面,难以向微观尺度深入,它们忽略了通过环境设计预防犯罪（Crime Prevention Through Environmental Design, CPTED）理论中所主张的地址级的建成环境的精确特征。地址级的微观建成环境被广泛认为对各类犯罪的发生有着直接或间接的影响,然而对微观建成环境的度量一直是一个挑战。先前的大多数研究都是通过调查样本来表征建成环境,会受到两方面的限制：① 建成环境特征描述不完整的限制;② 数据在空间覆盖方面具有稀疏性的限制。百度街景图像作为一个新的数据来源,可以被用来提取地址级的微型环境的建成特征,从而使犯罪研究可以聚焦在更微观的尺度中。本研究使用深度学习全卷积图像分割算法从百度街景图像中提取地理位置的环境变量,共选取树木、通车道路、人行道等8种变量来表现研究区微观建成环境的差异。在控制了与街面犯罪有关的其他因素后,采用贝叶斯逻辑回归模型来评估微观建成环境影响因素对公共盗窃和寻衅滋事案件的影响。结果表明加入了微观建成环境物理特征之后的模型表现更好。对比寻衅滋事案件,树木多的隐蔽地方更容易发生公共盗窃案件,通车道路、人行道多的地方更不易发公共盗窃案件,这也说明了更隐秘的地方公共盗窃案多发。总的来说,全卷积深度学习图像分割算法可以有效地提取街景衍生变量,这些变量为微观空间尺度的犯罪分析增加了新的维度。本研究不仅对于犯罪地理文献具有贡献,而且为基于CPTED原则的犯罪预防提供了新的视角。

Using Baidu Street View Images to Assess Impacts of Micro-built Environment Factors on Street Crimes: A Bayesian Logistic Regression Approach

Street crimes pose threats to human life. Previous research tackling street crimes such as public thefts and provocations were often conducted at the neighborhood or even larger levels, thus ignoring the fine characteristics of the built environment at the address level, as is advocated by the theory of Crime Prevention Through Environmental Design (CPTED). It has been well established that the micro built environment at the address level has a direct or indirect impact on the occurrence of various types of crime. However, the measurement of the micro built environment, often called environmental auditing, has been challenging. Most of the previous studies relied on survey samples to characterize the built environment. They suffered from two limitations: incomplete characterization of the built environment and sparsity in spatial coverage. Baidu street view imagery is a new data source that captures built environment at the micro level. It often captures images in four different directions at a given location on the street, and it covers virtually all the drivable streets. This study used a full convolution deep learning image segmentation algorithm to extract the micro characteristics of the built environment from Baidu street view images and applied them to study crimes at the small address level. Eight variables including tree, road, sidewalk et al. were selected to show the differences of micro built environment in the study area. Bayesian logistic regression models were used to evaluate the impact of micro-built environment factors on public thefts and provocations by controlling other relevant criminological factors. Results show that the model performance was improved after adding the micro characteristics. Additionally, we found that places with more trees hence less visibility were more prone to public theft cases, and places with more roads and sidewalks were less prone to public theft cases, when compared with provocative cases. In sum, the convolution deep learning image segmentation algorithm is effective in extracting street-view-derived variables, and these variables added new dimensions to crime analysis at the micro spatial scale. This study not only contributes to the literature in crime geography, but also bring new insight for crime prevention following the principle of CPTED.