CGPM: Poverty Mapping Framework Based on Multi-Modal Geographic Knowledge Integration and Macroscopic Social Network Mining

Abstract

Having high-precision and high-resolution poverty map is a prerequisite for monitoring the United Nations Sustainable Development Goals(SDGs) and for designing development strategies with effective poverty reduction policies. Recent deep-learning-related studies have demonstrated the effectiveness of the geographically-fine-grained data composed with satellite images, geolocated article texts and Open-Street-Map in poverty mapping. Unfortunately, there is no presented method which considers the multimodality of data composition or the underlying macroscopic social network among the investigated clusters in socio-geographic space. To alleviate these problems, we propose CGPM, a novelty end-to-end socioeconomic indicator mapping framework featured with the cross-modality knowledge integration of multi-modal features, and the generation of macroscopic social network. Furthermore, considering the deficiency of labeled clusters for model training, we proposed a weak-supervised specialized framework CGPM-WS to overcome this challenge. Extensive experiments on the public multimodality socio-geographic data demonstrate that CGPM and CGPM-WS significantly outperforms the baselines in semi-supervised and weak-supervised tasks respectively of poverty mapping.