A Road-Segment-Based Rockfall Susceptibility Mapping Approach Integrating Physically Informed Slope-Cutting Features and Comparative Machine Learning Models

Rockfall hazards are frequently observed within mountainous road networks. Significant uncertainties regarding the optimal selection of evaluation units and spatial modeling scales are still being identified in this field. A comprehensive comparative framework for rockfall susceptibility mapping is presented in this study, using Wufeng County as the empirical study area. Five evaluation scenarios were constructed to systematically isolate the independent predictive contributions of the spatial domain, the mapping unit morphology, and the physics-informed engineering proxy. These scenarios included a whole-county macro-scale raster; three multi-scale road buffers with widths of 1 km, 2 km, and 3 km; and an object-oriented vector road evaluation unit (REU) framework. To parameterize localized engineering-induced risks, a physics-informed feature defined as the theoretical slope-cutting height (Hcut) was structurally introduced into the vector-based assessment. Thirteen representative machine learning, deep learning, and statistical algorithms—including Random Forest, LightGBM, and TabNet—were systematically cross-examined under both unconstrained splits and strict Leave-One-Road-Corridor-Out Validation (LORCOV) protocols. The empirical multi-metric sensitivity analysis explicitly decouples the three structural effects. First, isolating the effect of the spatial domain reveals that restricting the validation extent from a broad countywide area to a narrow road corridor purges unperturbed background terrain noise, shifting the focus from easy negatives to geomorphological hard negatives. Second, evaluating the independent effect of the evaluation unit demonstrates that transitioning from continuous raster pixels to homogeneous vector REUs successfully resolves the terrain smoothing effect, precisely characterizing sharp geomechanical gradients adjacent to cut slopes. Third, isolating the effect of adding Hcut proves that this engineering indicator drives the primary descriptive gain, enabling tree-based ensembles to achieve a peak baseline AUC of 0.7763 and maintain a robust spatial validation AUC of 0.6129 under strict geographic block constraints, whereas legacy deep learning architectures exhibit an inductive bias mismatch on small-scale tabular records. Rather than asserting a single optimal paradigm, this coordinated feature–unit matching framework provides transport authorities with a highly calibrated, target-tiered decision matrix to optimize localized public works safety budgets and protect critical linear infrastructure assets.