Measuring urban poverty using multi-source data and a random forest algorithm: A case study in Guangzhou

Abstract

Conventional measurements of urban poverty mainly rely on census data or aggregated statistics. However, these data are produced with a relatively long cycle, and they hardly reflect the built environment characteristics that affect the livelihoods of the inhabitants. Open-access social media data can be used as an alternative data source for the study of poverty. They typically provide fine-grained information with a short updating cycle. Therefore, in this study, we developed a new approach to measure urban poverty using multi-source big data. We used social media data and remote sensing images to represent the social conditions and the characteristics of built environments, respectively. These data were used to produce the indicators of material, economic, and living conditions, which are closely related to poverty. They were integrated into a composite index, namely the Multi-source Data Poverty Index (MDPI), based on the random forest (RF) algorithm. A dataset of the General Deprivation Index (GDI) derived from the census data was used as a reference to facilitate the training of RF. A case study was carried out in Guangzhou, China, to evaluate the performance of the proposed MDPI for measuring the community-level urban poverty. The results showed a high consistency between the MDPI and GDI. By analyzing the MDPI results, we found a significantly positive spatial autocorrelation in the community-level poverty condition in Guangzhou. Compared with the GDI approach, the proposed MDPI could be updated more conveniently using big data to provide more timely information of urban poverty.