23 JUN 2025 (MON) 11:00-12:00

Mapping Asia’s Poverty Dynamics using AI foundation models from remote sensing and other geospatial data

Mr REN Yongqian 

( Supervisor: Prof Shunlin Liang )

Abstract:

Asia is home to over half of the world’s poor, yet traditional household surveys remain costly, time-consuming, and geographically incomplete, which limits the timely tracking of poverty trends essential for effective policy and humanitarian intervention. Satellite imagery and complementary geospatial data offer a promising alternative by providing regular, cost-effective, and broad spatial coverage. However, current studies face significant limitations, including saturation effects in brightly lit urban areas, reduced accuracy in rural and underdeveloped regions, and weak cross-country transferability. To address these challenges, this study develops a comprehensive, continent-wide poverty-mapping framework that integrates diverse spatial datasets, such as multispectral Sentinel-2 imagery, nighttime light intensity, digital elevation models, population distribution, land-use types, air quality, and infrastructural features into a unified Swin Transformer model. Initially self-supervised on millions of satellite images from Sentinel-2, Landsat, and MODIS, the model is then fine-tuned with Demographic and Health Survey wealth index data from multiple Asian countries to improve cross-regional robustness. The framework will produce regularly updated, high-resolution (10 km) poverty maps for major Asian nations from 2015 onward while simultaneously performing temporal analyses and attributing observed poverty changes to specific explanatory factors, such as infrastructure improvements, urban growth, and environmental disturbances. The innovation of this research lies in the first large-scale implementation of a remote sensing foundation model for poverty mapping, which integrates multimodal data into an end-to-end architecture. This provides a scalable, interpretable, and temporally dynamic tool for accurate poverty assessment and monitoring.