Motivation and questions

A new dataset to study subnational development in Bolivia

Satellite imagery: A modern instrument for measuring local development

Three ways to see Bolivia

Data and methods

The prediction targets: 15 development goals

The outcome variable

Cover of the book Atlas Municipal de los Objetivos de Desarrollo Sostenible en Bolivia 2020, showing a snow-covered Andean plain under a blue sky beside the grid of SDG icons.

62 indicators from official statistics
▶
15 goal indices, each scored 0–100 — higher is better
▶
a 2017 value for each of 339 municipalities
SDG 1Poverty
SDG 2Hunger
SDG 3Health
SDG 4Education
SDG 5Gender
SDG 6Water
SDG 7Energy
SDG 8Jobs
SDG 9Infrastructure
SDG 10Inequality
SDG 11Cities
SDG 13Climate
SDG 15Land
SDG 16Institutions
SDG 17Partnerships

Municipal Atlas of the Sustainable Development Goals in Bolivia (Andersen et al., 2020) · only two of the 62 indicators come from satellites

Nighttime lights: the predictor economists already know

NASA Black Marble composite of the Earth at night, with cities and road networks visible as clusters of light.

Mean · Median

Typical brightness of each pixel across every cloud-free night of 2017.

Masked mean · median

The same, with background glow, fires and aurora removed: stable human light only.

Minimum · Maximum

The dimmest and brightest nights: how steady or episodic the light is.

Two coverage counts

How many nights were observed, and how many were cloud-free: data quality, not light.

Six radiance statistics plus two counts, aggregated to each municipality: all 8 enter the model together.

NASA Black Marble · Suomi NPP VIIRS day/night band, annual composite · radiance in nW/cm²/sr

Example: Nighttime lights and SDG 1 (no poverty)

VIIRS annual composite, population-weighted municipal means · 2017 · ordinary least squares, in-sample

Daytime embeddings: what the new predictors are

Six daytime satellite-image tiles showing agriculture, a city, forest, wetlands, mountains, and dry terrain flow through the AlphaEarth Model into a 64-dimensional satellite embedding represented by an 8-by-8 colored grid and a shortened numeric vector.

1 · A year of images

Each patch of land, observed again and again through 2017.

2 · Compressed by a model

A neural network trained worldwide, with no social, economic, or nighttime-light data.

3 · Into 64 numbers

Places that look alike get similar numbers. No single number has a name.

Intuition: a principal-components index of the landscape, with 64 components instead of one.

AlphaEarth Foundations satellite embeddings · explore how they work, in pictures ↗

From pixels to municipalities

Average every pixel equally or weight each pixel by the people on it.

Prediction and monitoring methods

How each question is answered

Two illustrated sequences. PREDICT: many decision trees merge into one prediction, a dial tunes the model, and a map with one region lifted out is tested on unseen data. MONITOR: a municipality linked to the neighbours it touches, a scatter plot with an upward trend showing global clustering, and a map with a red cluster, a blue cluster and one outlier.

Prediction · machine learning

  • Random forest, one model for each goal index.
  • Tuned by Bayesian search, inside a nested cross-validation.
  • Scored out of sample: random folds, then whole departments held out.

Monitoring · spatial dependence

  • Neighbours are municipalities whose borders touch — queen contiguity.
  • Global Moran’s I — how strong is the clustering process.
  • Local indicators of spatial association — hotspot, coldspot, or outlier.

Can satellites read development?

Part one · Prediction

On the raw average, nightlights win headline goals

Simple average · out of sample · 339 municipalities · daytime embeddings lead on only 8 of 15 — and trail on poverty and energy

Most land is empty, let’s weight each pixel by the people on it

Key methodological choice

Weight pixels by population and daytime embeddings win

Weighted by population · out of sample · daytime embeddings now lead on 14 of 15 goals (8 → 14) — though the lead clears fold noise on only 5 of the 15

Which goals can we map and how well?

Part two · Monitoring

Integrating nighttime lights and daytime embeddings provides a better prediction

Each goal spans the two ways of testing · The three goals above the 0.60 line are the ones we map next

The combined satellite view recovers the geography of no poverty (SDG1)

Red: clusters of low poverty · Blue: poverty traps · Each panel reports its agreement with the actual map and its clustering strength

The combined satellite view recovers the geography of affordable and clean energy (SDG7)

Red: clusters of good access · Blue: traps of poor access · The lights reach r = 0.73 on the level, and half that on the pattern

The combined satellite view recovers the geography of climate action (SDG13)

Red: clusters of strong action · Blue: traps of weak action · Best cluster recovery of the three goals — but both embedding views make the map look more clustered than it is

Satellites can monitor the material geography of development

Concluding remarks

1 · Weight by where people live

Daytime embeddings’ average R² rises from 0.29 to 0.41; they lead on 14 of 15 goals.

2 · What a goal is made of matters

Poverty reads at 0.69; health, gender and institutions barely register.

3 · Combine both views

Lights plus embeddings read poverty at 0.54 to 0.72 out of sample, across the two tests.

4 · Map the clusters

Poverty’s clusters reappear: 80% agreement with the actual map.

Use multiple population-weighted satellite data to better monitor local development between surveys and census.