Perceiving the fine-scale urban poverty using street view images through a vision-language model

Abstract

Urban poverty, as a complex socio-economic phenomenon, has been accorded significant priority in recent decades. However, most studies on measuring urban poverty primarily rely on statistical survey data that is challenging to collect and lacks spatial details. Furthermore, these existing measurements tend to overlook subjective perceptions. As a response, we propose a framework integrates street view images (SVIs) with a large vision-language model, termed CLIP, to accurately measure fine-grained urban poverty. Subsequently, multi-source data poverty index (MDPI) is calculated and used to assess the reliability of poverty perceptions. Finally, the nonlinear effects of street scenes on poverty perceptions are analyzed. This newly-proposed method not only demonstrates a consistent distribution with MDPI but also exhibits better spatial resolution and efficiency than existed ones. This study introduces a CLIP into urban studies, showing potentials for measuring urban poverty, particularly in emerging market countries where there is an abundance of Internet open data but limited statistical records. By providing a comprehensive assessment of urban poverty, this study can serve as a scientific basis for governments to formulate poverty alleviation policies and achieve precise urban management, thereby contributing to the achievement of SDGs.