A statistical model for predicting rock avalanche runout using source area width from 63 representative cases

Abstract

Developing predictive models for rock avalanche runout is crucial for hazard risk mitigation and management, while also deepening the understanding of rock avalanche dynamics. In this study, a database of 63 representative rock avalanches was compiled, and the geometric characteristics of these events were analyzed. Traditional regression methods were compared with neural network regression (NNR) approaches to find the optimal model. Model parameters, loss functions, and evaluation metrics were examined to determine optimal configurations. Based on correlation analyses and model performance, this study recommends the use of source area width as an alternative to failure volume in runout prediction models, effectively addressing the common challenge of estimating avalanche volume. Ultimately, an NNR-based model, utilizing mean squared logarithmic error as the loss function and incorporating source area width and fall height as input parameters, was identified as the optimal approach. This model achieved prediction errors within a -50% to 50% range with 95.2% probability, yielding a mean absolute percentage error of 23.22% and an R² of 0.85. These findings enhance rock avalanche prediction methodologies, providing more accurate tools for assessing the potential impact zones of destructive geological events.