GeoDevelopment Observatory of Bolivia / Tools
Predicting poverty from satellite embeddings
Random forest and XGBoost predict multidimensional poverty from 64 satellite embedding dimensions, with random and spatial cross-validation.
This Colab notebook trains random forests and XGBoost to predict the Multidimensional Poverty Index of 2012 from the 64 satellite embedding dimensions of 2017, and compares them with a baseline of night-time lights and population. Population-weighted embeddings beat area-weighted ones, and adding lights helps: the best models reach an out-of-fold R² of about 0.55–0.58 with random 5-fold cross-validation. When whole departments are held out, R² falls to about 0.2, a reminder that models struggle to extrapolate to new regions. A map of residuals shows where they fail: Pando is under-predicted. The text discusses the 2017/2012 timing mismatch.