↓ Skip to main content
Dr. Sunayana Ghosh | Computational Geometry & Geospatial ML Consultant
  1. Selected Work & Projects/

Predicting Well-Being from Space: A Solve for Good Fellowship

·4 mins
Social Impact Machine Learning Social Impact Geospatial Python Satellite Data DSSG
Dr. Sunayana Ghosh
Author
Dr. Sunayana Ghosh
PhD in Computational Geometry • 15+ years in Medtech & Climate Tech • Building data solutions that matter
Table of Contents

Solve for Good × World Resources Institute
#

Programme: Data Science for Social Good (DSSG) fellowship Partner: World Resources Institute (WRI) Repository: dssg/solveforgood-wri

The Challenge
#

Traditional measures of economic well-being — census data, household surveys, wealth indices — are expensive to collect, infrequent, and often sparse in exactly the regions where they are most needed. In large parts of India, the most recent ground-truth data on wealth and living standards is years out of date.

Can satellite imagery, open map data, and machine learning substitute for costly in-person surveys?

This was the central question of the Solve for Good fellowship collaboration with the World Resources Institute. The goal: build a well-being data layer that predicts the wealth index of any given area in India using only freely available, continuously updated data sources.

Ground Truth: DHS Wealth Index
#

Wealth index distribution across survey clusters
DHS wealth index values across India — the target variable for ML prediction

The project used Demographic and Health Survey (DHS) data as ground truth. DHS clusters each have a wealth index scored on a 1–5 scale, derived from household-level indicators including:

  • Access to electricity
  • Toilet type (flush vs. pit)
  • Water source
  • Roofing material
  • Household assets

Wealth label distribution across clusters
Distribution of DHS wealth index labels across the study area

A key challenge: DHS applies a GPS privacy offset to all cluster coordinates — up to 5 km in rural areas, 2 km in urban — making precise spatial matching with other data sources non-trivial.

Spatial Data Preparation: Laguerre-Voronoi
#

To match DHS clusters with spatial features from OSM and satellite data, each cluster needed a defined geographic area of influence. Standard buffers or equal-area Voronoi cells ignore the fact that urban and rural clusters represent very different population densities.

The solution was Laguerre-Voronoi tessellation (weighted Voronoi / Power diagrams), implemented in the companion gis-laguerre open-source library. Each DHS cluster is assigned a cell weighted by its urban/rural classification, creating a spatial partition that reflects representativeness.

Weighted Voronoi tessellation clipped to study region
Laguerre-Voronoi cells for DHS clusters — the spatial framework for aggregating OSM and satellite features

Feature Engineering
#

Three data streams were aggregated per Voronoi cell:

OpenStreetMap (OSM) Extracted via OSMnx: amenities, buildings, highways, and tagged infrastructure features. OSM encodes observable proxies of economic activity — density of roads, hospitals, markets, and services all correlate with local wealth.

Night-Time Light (NTL) Satellite Data

  • NASA VIIRS/NPP LunarBRDF-Adjusted Night-Time Light data at 500 m resolution
  • Google Earth Engine monthly composites
  • Averaged per Voronoi cell as a continuous proxy for electrification and economic activity

DHS Indicators Exploratory analysis revealed strong correlations between wealth and:

  • Electricity access
  • Flush toilet prevalence
  • Urban/rural classification

Model & Evaluation
#

Problem formulation: Regression — predict a standardised wealth score [0–1] per cluster from OSM + NTL features.

Model: Decision tree, chosen for interpretability and explainability — essential for a social impact context where stakeholders need to understand and trust predictions.

Evaluation: Leave-One-Out Cross-Validation (LOOCV), which is appropriate for small, spatially structured datasets where standard random splits can introduce optimistic bias.

Decision tree model accuracy
Decision tree prediction accuracy across the evaluation set

Predictions vs. Ground Truth
#

Real wealth values
Actual DHS wealth index values
Model predictions
Model-predicted wealth scores
Actual wealth indexPredicted wealth scores

Spatial analysis of the Araria district in Bihar, India — a focus region chosen for its data availability — revealed localised wealth variation within districts, with prediction errors primarily in the 0.1–0.2 range.

Key Findings
#

  • Wealth correlates most strongly with electricity access and flush toilet use in the DHS data
  • OSM feature density provides meaningful signal, though data completeness varies geographically
  • Night-time light data is a consistent proxy for economic activity at the cluster level
  • Preliminary results showed the model can learn to differentiate wealth levels, providing evidence that higher-quality features would yield a deployable system

Impact & Scope
#

A reliable well-being prediction layer would enable:

  • Policy targeting — directing interventions to underserved areas without waiting for the next survey cycle
  • Impact monitoring — tracking changes in well-being over time using continuously updated satellite data
  • Coverage extension — scaling predictions beyond India to any region where DHS data exists

The project was scoped with deployment across India and eventual extension to other countries.

Related Work#

The spatial data preparation methodology was packaged into the open-source gis-laguerre library and documented in a Towards Data Science article, making the Laguerre-Voronoi approach available to the wider data science community.

Domain: Applied ML · Social impact · Geospatial data science · Satellite imagery · Economic development

Data: DHS (Demographic and Health Surveys) · OpenStreetMap · NASA VIIRS Night-Time Light

Stack: Python · GeoPandas · OSMnx · Google Earth Engine · scikit-learn