AI-Driven Multidimensional Poverty Mapping Using Satellite and Socioeconomic Data

Accurate poverty assessment remains critical for effective policy implementation and resource allocation, yet traditional survey-based approaches are costly and time-intensive, limiting their scalability. This paper introduces a novel AIdriven framework that integrates multi-spectral satellite imagery with socioeconomic indicators to generate high-resolution poverty maps at unprecedented spatial granularity. Our approach employs a multi-modal ResNet-18 architecture with same-scaled initialization, processing Landsat- 8 imagery (30 m resolution) combined with VIIRS nighttime lights and demographic survey data from 23 African countries. We implement a physicsinformed neural network layer that enforces spatial continuity constraints and incorporates the Alkire-Foster Multidimensional Poverty Index formulation. The framework achieves R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.847 for wealth index prediction and 91.3 % accuracy for binary poverty classification, representing a 23.7 % improvement over traditional satellite-only approaches. Validation on Demographic Health Surveys spanning 19,669 villages demonstrates robust generalization across diverse geographical contexts. Our model reduces poverty mapping costs by 89.4 % compared to conventional household surveys while maintaining spatial resolution of 1km<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>. The framework enables real-time poverty monitoring for emergency response, targeted social assistance programs, and sustainable development goal tracking across data-scarce regions.

Перевод пока недоступен