The frequency and intensity of extreme weather events pose various types of hazards and vulnerabilities to the railway system. This article aims to challenge the assumption that a single, nationally and internationally aggregated model is sufficient to capture region-specific failure patterns. While useful, these national models are often unable to account for linkages and feedback between heterogeneous regions and associated railway failure mechanisms and processes. This study develops and evaluates regional-national specific machine learning (ML) framework to enhance the classification of climate-related railway failures utilizing a synchronized five-fold cross-validation strategy.

We illustrate the approach through a case study of the Switches and Crossings (S&C) assets across five railway operational regions in Sweden, over a 24-year period, to demonstrate meaningful climate variability. Our methodology provides a consistent procedure that consistently compares national models with region-specific models. Results show that the localized model outperforms the national baseline, with improvements of up to 2.0% in Matthews Correlation Coefficient and 1.2% in F1-score at the 24-hour window preceding failure. Shapley Additive Explanations (SHAP) analysis revealed regionally distinct predictors, such as snow accumulation in the North and weather data coverage issues in the West. The results provide a basis for further process refinement, offering interpretable insights for operation and maintenance planning, risk assessment and supporting climate adaptation actions to achieve a climate-proof and resilient railway system.