Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Compact Satellite Embeddings and Graph Neural Networks for Large-Scale Poverty Mapping

Markus B. Pettersson, Adel Daoud

TL;DR

This work addresses the shortage of fine-grained poverty maps by leveraging compact AlphaEarth satellite embeddings within graph neural networks to model spatial relations among DHS survey points and nearby settlements. A fuzzy label loss is proposed to account for DHS coordinate displacement, while six modeling variants evaluate the benefits of graph structure, unlabeled GeoNames nodes, and displacement-aware supervision. Results show modest gains from incorporating graph structure, with ego-graph variants delivering the strongest performance, and reveal that fuzzy-label supervision can be brittle given incomplete priors. The approach demonstrates the practicality of continental-scale poverty mapping using lightweight EO embeddings and spatially aware graph methods, paving the way for more targeted interventions in data-scarce regions.

Abstract

Accurate, fine-grained poverty maps remain scarce across much of the Global South. While Demographic and Health Surveys (DHS) provide high-quality socioeconomic data, their spatial coverage is limited and reported coordinates are randomly displaced for privacy, further reducing their quality. We propose a graph-based approach leveraging low-dimensional AlphaEarth satellite embeddings to predict cluster-level wealth indices across Sub-Saharan Africa. By modeling spatial relations between surveyed and unlabeled locations, and by introducing a probabilistic "fuzzy label" loss to account for coordinate displacement, we improve the generalization of wealth predictions beyond existing surveys. Our experiments on 37 DHS datasets (2017-2023) show that incorporating graph structure slightly improves accuracy compared to "image-only" baselines, demonstrating the potential of compact EO embeddings for large-scale socioeconomic mapping.

Leveraging Compact Satellite Embeddings and Graph Neural Networks for Large-Scale Poverty Mapping

TL;DR

This work addresses the shortage of fine-grained poverty maps by leveraging compact AlphaEarth satellite embeddings within graph neural networks to model spatial relations among DHS survey points and nearby settlements. A fuzzy label loss is proposed to account for DHS coordinate displacement, while six modeling variants evaluate the benefits of graph structure, unlabeled GeoNames nodes, and displacement-aware supervision. Results show modest gains from incorporating graph structure, with ego-graph variants delivering the strongest performance, and reveal that fuzzy-label supervision can be brittle given incomplete priors. The approach demonstrates the practicality of continental-scale poverty mapping using lightweight EO embeddings and spatially aware graph methods, paving the way for more targeted interventions in data-scarce regions.

Abstract

Accurate, fine-grained poverty maps remain scarce across much of the Global South. While Demographic and Health Surveys (DHS) provide high-quality socioeconomic data, their spatial coverage is limited and reported coordinates are randomly displaced for privacy, further reducing their quality. We propose a graph-based approach leveraging low-dimensional AlphaEarth satellite embeddings to predict cluster-level wealth indices across Sub-Saharan Africa. By modeling spatial relations between surveyed and unlabeled locations, and by introducing a probabilistic "fuzzy label" loss to account for coordinate displacement, we improve the generalization of wealth predictions beyond existing surveys. Our experiments on 37 DHS datasets (2017-2023) show that incorporating graph structure slightly improves accuracy compared to "image-only" baselines, demonstrating the potential of compact EO embeddings for large-scale socioeconomic mapping.

Paper Structure

This paper contains 15 sections, 3 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Method
  4. (A) Baseline: DHS Points.
  5. (B) DHS graph.
  6. (C) Full graph (DHS + GeoNames).
  7. (D) Ego graphs.
  8. (E) Points With Fuzzy labels.
  9. (F) Graph With Fuzzy labels.
  10. Model and training
  11. Results
  12. Discussion and Outlook
  13. Displacement mechanism
  14. Model details
  15. Cross-Validation Folds

Figures (2)

  • Figure 1: Structures of the six methods evaluated. In this figure, The colored nodes are different survey locations from the DHS while the the white are settlement locations from GeoNames. Both of these have corresponding AlphaEarth embeddings $x_i$. The grayed out GeoNames locations in A, B, D and E signify that they are not included in the model training. For E and F, where the embeddings of the survey clusters are not included, the larger circle highlights the possible displacement area which assigns fuzzy labels to nearby GeoNames nodes.
  • Figure 2: Predicted cluster-level IWI across GeoNames settlements in selected countries. Coverage and spatial density vary by country.