Mapping urban villages based on point-of-interest data and a deep learning approach

Abstract

In the process of urban development, spatial structure within cities undergoes great changes, where the rural areas are surrounded by newly urban blocks, leading to the widespread of urban villages. Thus, quick and accurate prediction of urban villages is crucial for urban planning, management and sustainability. Recently, point-of-interest (POI) data mining has emerged as a popular topic in urban research. This study aims to propose an urban village prediction model in complex urban landscape patterns by utilizing POI data as a single data source. We firstly calculated word embeddings of POI types as the semantic features of urban villages based on Word2Vec. Afterwards, a BiLSTM-Multiscale-Attention (BMA) model is proposed to predict urban or non-urban villages based on POI word embeddings. Experimental results in several major cities of China, including Beijing, Tianjin, Xi'an, Shijiazhuang, Wuhan, and Guangzhou indicates that the proposed model achieved an average overall accuracy of 84.06 %, outperforming several other data-driven methods. This study demonstrates that POI data can provide accurate spatial distribution information for urban villages. These findings provide new ideas and references for comprehensive understanding of urban villages at a fine scale.