A hybrid approach combining physics-based model with extreme value analysis for temporal probability of rainfall-triggered landslide

Abstract

The interplay between climate change–induced extreme rainfall and slope failure mechanisms presents a significant challenge. To address this, a new temporal modeling of landslides that integrates dynamic rainfall patterns with slope failure mechanisms is proposed. The approach features three steps: (1) analysis of the critical continuous rainfall (CCR) level using a physics-based model with Monte Carlo simulation; (2) calculation of the cumulative distribution function of the generalized extreme value distribution; and (3) estimation of the temporal probability map. Then, combined with the landslide spatial probability obtained from one-dimensional convolution neural network (1D-CNN), the landslide hazard probability was estimated for future periods of 5, 10, 20, and 50 years. The CCR and spatial probability maps were validated using the 2018 landslide event in Hiroshima Prefecture, Japan. The CCR map achieves an area under the receiver operating curve (AUC) of 74.8%. Cohesion and friction angle are the most sensitive in the hybrid model. The proportions of temporal probabilities > 0.5 yielded by the non-stationary model (10, 19, 28, and 38%) were greater than those of the stationary model (6, 10, 16, and 24%) for periods of 5, 10, 20, and 50 years, respectively. The 1D-CNN model (AUC = 84.1%) outperformed logistic regression (AUC = 80.1%) and naïve Bayes (AUC = 80.1%) models. The landslide hazard probability obtained from the non-stationary model is more susceptible than that of the stationary model. These results indicate that the proposed approach is a valuable tool for future landslide risk assessment and may be applicable even in areas without a landslide inventory.