The rapid advancement of autonomous driving and ADAS technologies has increased the demand for high-definition (HD) lane-level maps that accurately preserve rich geometric information while scaling to city-wide coverage. While traditional polyline-based formats are widely used, they struggle to provide continuous representations of key geometric properties such as curvature and heading angle, which are essential for autonomous driving applications. Curve-based representations have been introduced to address these limitations, but existing methods are often restricted to simplified or sparsely connected road networks, limiting their effectiveness in large-scale, real-world environments. This study presents an arc-spline-based lane representation framework that efficiently models complex, large-sized lane-level maps while preserving continuous road geometry. To achieve this, we introduce a novel road network decomposition and merging method that enables structured parameterization without requiring full map-scale optimization. Instead, optimization is localized to cluster connection regions, significantly enhancing computational efficiency. Validation using lanelet maps from the nuScenes dataset demonstrates that our approach maintains an average approximation error of 4.3 cm while preserving detailed lane topology and global tangential continuity, while also achieving a significant reduction in storage requirements compared to conventional polyline formats.