Abstract

This study employed seven machine-learning algorithms: Random Forest, XGBoost, LightGBM, Support Vector Machine (SVM-RBF), Elastic Net, Multilayer Perceptron (MLP), and hybrid PCA-enhanced models to predict the Livelihood Vulnerability Index (LVI-IPCC) of smallholder farmers in Southern Zambia’s Agro-Ecological Region I. Using grouped cross-validation to prevent spatial bias, the PCA-MLP and Random Forest models emerged as top performers, achieving R² values above 0.95 and RMSE below 0.05. These models effectively captured nonlinear socio-ecological interactions that influence vulnerability. Feature importance analyses identified education, income, water access, and drought exposure as key predictors. The integration of dimensionality reduction (PCA) improved model stability and interpretability. These findings demonstrate that hybrid machine-learning approaches outperform traditional LVI aggregation in predicting household vulnerability, providing scalable, data-driven insights for climate adaptation planning. The results highlight the potential of artificial intelligence to revolutionize vulnerability assessments and inform targeted resilience strategies in regions affected by climate change.

This work is licensed under a Creative Commons Attribution 4.0 License.