Spatial rockfall susceptibility prediction from rockwall surface classification

Rockfall both is a major process in shaping steep topography and a hazard in mountainous regions. Besides increasing thread due to thawing permafrost-stabilization in high-elevation areas, there are abundant permafrost-free over-steepened rockwalls releasing rockfall due to other triggers. General rockfall event susceptibility is addressed to frost cracking, earthquake shacking and hydrologic pressure in the walls, and to geotechnical rock properties. Spatial rockwall surface surveys or scans (delivering 3D point clouds) have been used to both deduce rock fracture patterns and to measure individual rockfall events from comparing subsequent scans. Though, the actually measured rockwall topography data has rarely been used as a general predictor of rockfall susceptibility against the background of observed events.

In this study, we use a series of dm-resolved annual (2014 to 2020) terrestrial laser scan surveys along 5km² of limestone cliffs in the Lauterbrunnen Valley, Switzerland. The annual scan data were hand-cut to remove vegetation and fringes, and then referenced to detect subsequent topographic change in the direction of the wall. From the change-detection point clouds individual rockfall event volumes were detected from cluster and filtering analyses. One surveyed rockwall section of 2014 was used as training data for our Bayesian classification model of rockfall susceptibility, while the adjacent remaining section served for model validation. We rasterized their 3D data points and calculated several surface parameters per cell, including roughness, topography, mean distances for the three main fracture systems, fracture density, local dip, percent of overhang area, normal vector change rate (called edge) and percentage of overhang area. For various parameter sets and different cell sizes (3²m², 5²m², 10²m², 15²m², 25²m², and 40²m²), we trained Naïve-Bayes-Classifier models. These were then used to predict rockfall susceptibility per cell, based on our observations of surface parameters, and assessed using Kullback-Leibler Divergence analysis and the misclassification cost score.

Results indicate the overall best model (accounting for the parameters roughness, edge, topography and overhang area) and for the lowest cell size (3²m²) could predict rockfall cells with a probability of 0.73 (against a mean of 0.3 for all cells). Predictions on another rockwall section with observed rockfall, located on the opposite side of the valley, verified the model’s applicability by both comparable probabilities (0.6 vs 0.25) and visual surveys on overhangs. We find our approach could reliably extend this spatial rockfall susceptibility classification to all Lauterbrunnen rockwalls. The classification model generally identified overhang areas and fractured zones as high rockfall risks, matching the general insight of these zones to be of major susceptibility. Interestingly, our method is based only on orientation-independent variables that are directly calculated from the 3D point cloud. Thus, it should be principally transferable to other sites of fractured limestone walls. Specifically, there is no need to determine fracture sets from the point cloud as is generally done for susceptibility studies, since we account for topography that would anyway be used to calculate fracture planes (facets). Hence, this method provides a simple means to predict spatial rockfall susceptibility, applicable for both hazard mapping and landscape evolution studies.