From Climate Extremes to Risk: Quantifying Agro-Ecological Climate Vulnerability in Ghana Using Explainable Machine Learning

Climate extremes pose increasing risks to agro-ecological systems in sub-Saharan Africa, yet their drivers, interactions, and spatial variability remain insufficiently quantified. This study develops an integrated framework to assess climate vulnerability across Ghana by combining machine learning, explainable artificial intelligence, and spatio-temporal analysis. A Climate Extremes Index (CEI) was constructed from long-term station data and modelled using Random Forest, with performance evaluated against OLS, GAM, and XGBoost. Model interpretation was conducted using SHAP to identify key drivers and their interactions, while partial dependence analysis and heatmaps were used to examine non-linear responses and temporal patterns. A Climate Risk Score (CRS) was further derived to characterise the evolution of risk across eco-zones. Results show that Random Forest outperforms alternative models, capturing the non-linear and interaction-driven structure of climate extremes. Lagged temperature and humidity emerge as dominant predictors, highlighting the role of short-term climatic memory. Strong interaction effects, particularly between temperature and moisture variables, indicate that extremes are driven by compound processes rather than isolated factors. Marked spatial heterogeneity is observed, with distinct temporal clustering of extremes across Coastal, Forest, Transition, and Savannah zones. Risk patterns reveal persistent differences in magnitude and temporal structure across eco-zones. These findings demonstrate that climate vulnerability in Ghana is governed by non-linear dynamics, antecedent conditions, and spatially differentiated processes. The study provides a scalable, and interpretable framework for climate risk assessment, providing direct relevance for improving early warning systems and supporting targeted location-specific adaptation strategies in data-constrained environments.