Assessing the utility of regionalized rock-mass geomechanical properties in rockfall susceptibility modelling in an alpine environment

The main goal of this study was to develop a reliable rockfall susceptibility map for Valchiavenna (275 km(2)), located in the Italian Central Alps, through the introduction of outcrop-scale geomechanical properties (Joint Volumetric Count-Jv, rock mass Weathering Index-Wi and Equivalent Permeability-Keq) as spatially distributed predictors. Specific objectives were: (i) to increase the representativeness over the study area of an existing geomechanical dataset by adding new surveys, (ii) to effectively regionalize the geomechanical properties and (iii) to evaluate the performance and the physical plausibility of a rockfall susceptibility model combining geomechanical, topographical, geomorphological, and geological predictors. We optimized new survey locations by means of Spatial Simulated Annealing (SSA) and Multivariate Envi-ronmental Similarity Surface (MESS). For the regionalization of predictors we tested several interpolation techniques and evaluated them through performance indices and leave-one-out-validation. We performed the susceptibility analysis using rockfall data from the official Italian inventory, later updated with several field -mapped rockfalls, and different combinations of predictors. We applied Generalized Additive Models, which we evaluated through spatial k-fold cross-validation in terms of model performance (AUROC) and physical plausibility. Also, we investigated the importance of the predictors in the model through penalization and the calculation of the mean decrease of deviance explained (mDD%) upon recursive removal of each predictor. Through SSA we added 25 survey locations that reduced the study area with negative MESS from 26.2 % to 15.9 %. We calculated he geomechanical predictor maps applying ordinary kriging to Jv (NRMSE = 13.7 %) and Wi (NRMSE = 14.5 %) and using Thin Plate Splines for Keq (NRMSE = 18.5 %). The model containing the geomechanical predictors resulted in acceptable rockfall discrimination capabilities (mean AUROC > 0.7), with high-susceptibility areas located in plausible geomorphological contexts, charac-terized by currently active deformations (verified by means of inSAR data), which were not revealed by the topographic predictors alone. Regarding importance, Jv showed an mDD% of 7.5 % comparable to those of secondary topographic predictors (e.g., profile curvature, northness), while Wi and Keq were penalized out of the model. Models built with the non-updated inventory resulted in physically implausible susceptibility maps and predictor behavior (unreasonable smoothing functions), highlighting a model bias.

Automated summary

The main goal of this study was to develop a reliable rockfall susceptibility map for Valchiavenna, Italy. The study used various survey methods and interpolation techniques to optimize and evaluate the predictive model, considering geomechanical, topographical, geomorphological, and geological factors. The results showed that the model had good discrimination capabilities and identified high-susceptibility areas in plausible geomorphological contexts. The importance analysis of the predictors revealed that Jv was the most important factor.