Physically based and statistically based rockfall susceptibility along communication routes and in urban areas in Italy

Landslide susceptibility is the likelihood for a particular location to experience a landslide, based on terrain attributes and past landslide occurrences. The recent literature exhibits different approaches for the spatial zonation of landslide susceptibility. At the opposite sides of the spectrum of possible approaches lie physically based and statistically based methods. Physically based approaches calculate slope stability using well-defined equations, specific of the peculiar landslide type; here, we considered rockfalls [1,2]. Statistically based and machine learning approaches establish correlations between several topographic and environmental data and landslide presence [3].

Information on past landslides is useful for both methods, to calibrate model parameters and assess model performance. However, they differ significantly in their input requirements and methodological framework. In this study, we compare two state-of-the art susceptibility zonations, and their predictions at the location of different infrastructure in the whole of Italy, obtained by a physically based method [4] and with a slope unit-based statistical method [5].

To compare the two results, beyond classification performance, one has to figure out ways to cast the output maps of the two models in a similar format. Simulations with the 3D rockfall model produce raster maps, with a trajectory count for each grid cell, while the statistical result is a polygonal map [6]. We compared the two susceptibility zonations on the whole of Italy, first, and then we considered the predictions of the two results restricted to urban areas, railways, and road network.

The main difficulty lays in choosing an aggregation function for each polygonal or linear feature, to homogenize the two results. We performed either an average, for slope unit polygons, and empirical cumulative density functions (ECDFs), for linear features and urban areas. For the latter, we considered functional urban areas, or commuting zones, a standard choice to describe urban boundaries. Once average or ECDF values were obtained, for each polygon/linear segment, and for each version of susceptibility maps, we classified both results with an equal interval scheme. We acknowledge that the choice of aggregation functions and classification schemes are crucial for the final comparison, but we maintain that out choices are simple and objective.

The results indicate that the maps based on the considered models are drastically different. The observed disparities stem from the distinct conceptual frameworks and data dependencies of the two methods. While the physically based method can easily capture the mechanics of rockfall initiation, it requires input potentially limiting its use to data-rich locations. In contrast, the statistically based method is more flexible, and suitable for to regional-scale mapping. However, reconciling the two maps still looks challenging, and these preliminary results suggest complementary use of both methods.

                                                                                

[1] Guzzetti et al., Comp. Geosci. 28 (2002) https://doi.org/10.1016/S0098-3004(02)00025-0

[2] Sarkar et al., Nat. Haz.120 (2024) https://doi.org/10.1007/s11069-024-06821-9

[3] Alvioli et al., Earth-Sci. Rev. 258 (2024) https://doi.org/10.1016/j.earscirev.2024.104927

[4] Alvioli et al., Eng. Geol. 293 (2021) https://doi.org/10.1016/j.enggeo.2021.106301

[5] Loche et al., Earth-Sci. Rev. 232 (2022) https://doi.org/10.1016/j.earscirev.2022.104125

[6] Alvioli et al., Geomorphology (2023) https://doi.org/10.1016/j.geomorph.2023.108652