AI + Earth observation to measure what changes lives—and why.

Machine‑learning models trained on satellite imagery can predict household wealth with R² values approaching 0.80, but these predictions shrink toward the mean and attenuate estimated causal effects—e.g., a true 5% impact may appear as only 2–3%. The associated paper introduces two post‑hoc corrections (Linear Calibration and Tweedie’s) that debias predictions without needing fresh ground‑truth data, enabling a “one map, many trials” paradigm for reuse across multiple causal evaluations.

Figure

A continent‑scale evaluation compares Chinese and World Bank projects across 9,899 neighbourhoods in 36 African countries between 2002–2013, covering about 88% of the population. Using machine‑learned wealth indices from 6.7 km² satellite mosaics and inverse‑probability weighting, the study finds that both donors raise wealth but China’s projects deliver larger and more consistent gains; sector‑level extremes include World Bank trade & tourism projects adding +12.29 IWI points and China emergency‑response projects adding +15.15 IWI points. World Bank project placement is more predictable from imagery, suggesting Chinese placements depend more on unobserved factors.

Panel 1 Panel 2 Panel 3 Panel 4 Panel 5 Panel 6

The paper introduces Multi‑Scale Representation Concatenation, which turns any single‑scale earth‑observation CATE estimator into a multi‑scale version by concatenating image representations and feeding them into a causal forest. Simulation studies and applications to Peru and Uganda anti‑poverty RCTs show that multi‑scale models capture effect heterogeneity that single‑scale models miss and improve the Rank Average Treatment Effect Ratio (RATE). For Uganda, multi‑scale models increase the RATE‑ratio by 0.95 (s.e. 0.10) compared with 0.41 for raw single‑scale models; in Peru the gains are 0.68 vs 0.00, with optimal heterogeneity detection achieved by combining small (~64‑pixel) and large (~350‑pixel) image contexts.

Figure 1 Figure 2 Figure 3 Figure 4

Many social and environmental phenomena change over time, yet image sequences are under‑utilised in causal inference. This paper compares models for estimating Conditional Average Treatment Effects (CATEs) from sequences of satellite images and finds that richer image‑sequence models (with more parameters) better detect treatment effect heterogeneity. Applied to two RCTs—a poverty intervention in Cusco, Peru and a water‑conservation experiment in Georgia, USA—the methods show how model choice, data source (images vs tabular), and evaluation metric affect detected heterogeneity, and demonstrate how satellite sequences can generalise RCT results to larger geographic areas.

Figure 1 Figure 2 Figure 3 Figure 4

This study formalises how patterns in satellite images can confound causal estimates and develops methods to adjust for such image‑based confounders. In a Nigeria case study, where about 40% of the population lives on less than $2/day despite 3% annual growth, the authors combine DHS wealth data (International Wealth Index), AidData aid locations and 14.25 m‑resolution Landsat imagery. They define treatment as an aid program within 7 km of a survey cluster and show via simulations and neural‑network experiments how image resolution and model specification influence bias and the need for confounder adjustment.

Figure 1 Figure 2 Figure 3

A comprehensive scoping review catalogues Earth‑observation–machine‑learning (EO‑ML) methods used for causal inference and identifies five workflows: (1) outcome imputation, (2) image deconfounding, (3) treatment effect heterogeneity, (4) transportability analysis and (5) image‑informed causal discovery. Development assistance reached $223.7 billion in 2023 and the global extreme‑poverty rate fell to 8.4% in 2019, yet up to 575 million people may still live in extreme poverty by 2030; the review argues that EO‑ML techniques can help address these persistent challenges and provides protocols for data requirements, model selection and evaluation when integrating EO data into causal analyses.

Composite Workflow

This work develops a deep probabilistic model that clusters satellite images by their treatment‑effect distributions, enabling estimation of CATEs directly from images. Simulation results show the model better recovers true effect clusters than TARNet‑based clustering under high noise. Applied to a Ugandan anti‑poverty RCT using Landsat imagery, the model distinguishes high‑response areas (e.g., accessible terrain with good transportation) from low‑response regions (harsh, mountainous areas) and provides posterior predictive maps of cluster probabilities across Uganda; the correlation between raw and orthogonalized cluster probabilities is 0.85, indicating robust image‑derived heterogeneity estimates.

Figure 1 Figure 2 Figure 3