Leakage-aware comparative machine learning for earthquake susceptibility mapping in the Rif region, Northern Morocco

Assessing earthquake-prone areas in the Rif region remains methodologically challenging because susceptibility patterns are controlled by complex tectonic structure and because seismicity-derived predictors can artificially improve model performance when not handled carefully. In this study, a comparative machine-learning framework was developed for northern Morocco using a USGS earthquake inventory (1980–2025, M ≥ 3.0) and 7 GIS-derived conditioning factors prepared at 30 m resolution. After assembling a labelled dataset of 840 observations (443 earthquake samples and 397 pseudo-absence samples), 5 classifiers were evaluated: Random Forest, Gradient Boosting, Logistic Regression, Support Vector Machine, and Multilayer Perceptron. To limit circularity associated with seismicity density, the analysis combined predictor ablation, temporally constrained validation, calibration assessment, and 4-fold spatial block cross-validation. Among the tested models, Random Forest showed the most stable performance, with the best temporal AUC (0.925), a temporal Brier score of 0.120, and the strongest spatial cross-validation results (AUC = 0.946 ± 0.052, accuracy = 0.916 ± 0.014, Brier score = 0.070 ± 0.017). The results indicate that fault-related variables and training-period seismic clustering exert the strongest control on predicted susceptibility. The final map highlights concentrated high-susceptibility corridors in the central and eastern Rif, consistent with the main seismotectonic structure of the region. Overall, the study emphasizes that robust validation is as important as predictive accuracy when producing regional earthquake susceptibility maps.