Landslides pose a significant global threat, causing extensive loss of life, economic damage and environmental degradation. Despite advancements in landslide susceptibility mapping, existing methods often lack global-scale applicability and fail to incorporate robust optimization strategies for improved predictive accuracy. This study addresses these gaps by developing an optimized framework using support vector regression (SVR) enhanced with meta-heuristic algorithms (grey wolf optimizer [GWO] and bat algorithm) to refine model hyper-parameters. It integrates a globally representative data set of 37,984 landslide and non-landslide locations, ensuring broader applicability and generalizability. The information gain ratio method assessed the relative importance of 12 geo-environmental factors influencing landslide. The results indicated that all models achieved good predictive performance during the testing phase, as evidenced by an area under the receiver operating characteristic curve (AUC) value exceeding 0.8, but the SVR-GWO model exhibited the highest prediction accuracy (AUC = 0.92), making it suitable for large-scale hazard assessment. Plan curvature emerged as the most influential factor, surpassing slope, land use, and rainfall that are dominant at regional or local scales. The five countries with the highest landslide-prone areas were Russia, Canada, USA, China, and Brazil. The results support policymakers and urban planners in developing efficient strategies to minimize landslide risks.