Skip to content

GeoDevelopment Observatory of Bolivia / Tools

Predicting poverty from satellite embeddings

Machine-learning models predict municipal multidimensional poverty in 2012 and 2024 from 2017 satellite embeddings and night-time lights, with random and spatial cross-validation.

This Colab notebook trains random forests and XGBoost to predict the Multidimensional Poverty Index of the SDG Atlas of Bolivia 2026 for the 2012 and 2024 censuses from the 64 dimensions of Google’s 2017 satellite embeddings, and compares them with a baseline of night-time lights and population. It uses the 333 municipalities whose territory is unchanged since the older map the features were computed on. Under random 5-fold cross-validation, population-weighted embeddings plus lights explain 62% of the variance in 2012 poverty and 49% in 2024, against at most 35% for lights and population alone. When a whole department is held out, the best R² falls to 0.32 and 0.20, and the errors cluster by department; maps of the out-of-fold residuals show where the models fail, for example in Pando in 2012.

Back to the GeoDevelopment Observatory of Bolivia