Abstract
<title>Abstract</title> <p>Traditional traffic safety research relies heavily on global linear models that implicitly assume spatial stationarity, a paradigm that fails to account for the profound spatial heterogeneity where crash mechanisms differ fundamentally across regions due to unique combinations of local infrastructure, traffic composition, and climatic conditions. This aggregation approach leads to a significant masking effect where localized risk structures are smoothed out by state-level averages, resulting in persistent homogeneity bias and ineffective safety resource allocation. To resolve this, this study introduces the Manifold Spectral SHAP framework which utilizes Graph-based PCA to conduct non-linear dimensionality reduction and reconstruct the latent structure of nationwide crash data. This process effectively disentangles complex risk drivers into five physically interpretable latent dimensions including lighting constraints, urban traffic conflicts, atmospheric moisture, high-severity geometric risks, and thermodynamic winter mobility hazards. By integrating Gaussian Mixture Modeling, the research achieves precise identification of crash typologies, revealing that crash risks are geographically fragmented rather than uniform. Empirical findings confirm this spatial non-stationarity by contrasting logistics-driven hazards in states like Utah and Pennsylvania with commuter-demand risks in California and Oregon, and by demonstrating how county-level analysis uncovers localized risk structures that state-level averages systematically obscure. Specifically, the identification of the Cluster 3 typology in northern regions such as North Dakota and Minnesota highlights that high-risk environments are defined by the lethal coupling of localized thermodynamic instability such as black ice-induced friction loss, couples with infrastructure deficits to create unique high-risk environments. Ultimately, this research refutes the assumption of a universal safety baseline and establishes a physics-informed typology that provides a robust evidence-based foundation for transitioning from generalized governance strategies to targeted precision-oriented regional safety interventions.</p>