ML and Satellite Imagery for Conflict Prediction
Ensemble ML pipeline (RF, NN, KNN) achieving 92.3% accuracy and 0.93 ROC AUC to predict violent conflict across 50km grid cells continent-wide using ERA5 climate, UCDP events, and NASA VIIRS data
Overview
Predicting conflict in Africa applies logistic regression, k-nearest neighbors, random forest, and neural networks to grid-level features from UCDP, ERA5, Meta Relative Wealth Index, Hansen forest change, and VIIRS nighttime lights. A conservative ensemble balances recall and precision for early-warning–style use cases.
Demo and code
- Interactive site: climate-conflict-ml.vercel.app (data overview, full report PDF, Kepler.gl viewer)
- Repository: anfelipecb/conflict-prediction-ml
Authors
Andrés Felipe Camacho, Pablo Hernández Pedraza, Agustín Eyzaguirre — MSCAPP, University of Chicago.