摘要： 城市功能区识别对于城市规划和管理具有重要的支撑作用，目前大部分研究主要依赖于影像和兴趣点（POI）分布数据进行识别，但多将区域内不同出行行为的人群混杂在一起，没有考虑不同群体对区域产生的不同影响。结合物以类聚、人以群分的思想构建城市功能区识别模型UFAI，通过学习不同功能区人群出行活动的特征识别相应功能区。基于大样本粗粒度的匿名轨迹数据，刻画并提取个体出行特征，依据个体的出行特征划分人群类型。在此基础上，构建并训练多任务深度学习模型，实现城市功能区识别。选取北京市2 000万匿名用户10个月的手机信令数据作为人群出行轨迹数据，使用UFAI模型进行计算，并与决策树、随机森林、集成学习梯度提升决策树等7种传统分类模型进行对比。实验结果表明，UFAI模型的F1值达到0.95，与对比模型相比提升了0.10~0.29，具有更好的识别性能。

关键词: 城市功能区, 时空数据, 行为轨迹, 城市感知, 深度学习

Abstract: Urban functional areas identification plays an important supporting role in urban planning and management.Most studies mainly rely on images and Point of Interest(POI) distribution data for identification, but mostly mix people with different travel behaviors in the region, without considering the different effects of different groups on the region.Basis that birds of a feather flock together and people flock together, an urban functional areas identification model, UFAI, is constructed to identify the corresponding functional areas by learning the different characteristics of people's travel activities in different functional areas.Based on large sample, coarse-grained anonymous trajectory data, individual travel characteristics are characterized and extracted, and population types are divided according to individual travel characteristics.A multitask deep learning model is constructed and trained to identify urban functional areas.The 10-month mobile signaling data of twenty million anonymous users in Beijing were selected as the crowd travel trajectory data, calculated by the UFAI model, and compared with seven traditional classification models, such as the decision tree, random forest, and integrated learning gradient boosting classifier.The results show that the F1 value of the UFAI model reaches 0.95, which is improved by 0.10~0.29, compared with the comparison model, and has better recognition performance.

Key words: urban functional area, spatio-temporal data, behavior trajectory, urban perception, deep learning

中图分类号: 

• TP18