Predicting rockfall hazard with deep learning: Latent feature extraction from geological layers

This study introduces an innovative and comprehensive Machine Learning (ML) methodology for forecasting rockfall hazards in open-pit mining environments by integrating detailed geological stratigraphy extracted from 3D point clouds with advanced deep learning techniques. Rockfall is the sudden detachment of one or more rock blocks from a steep slope or cliff, followed by free-fall, bouncing, and rolling downslope. In open-pit mining, rockfall poses significant danger to both personnel and machinery. Unlike existing approaches that predominantly rely on basic geometric features and linear regression, this research incorporates geological stratification information into predictive models. A comprehensive synthetic dataset was created using detailed stochastic simulations of rockfall events on slope profiles, which were automatically generated from real-world 3D point clouds including stratigraphic features. A Variational Autoencoder (VAE) was developed to automatically extract latent features that effectively capture the complex relationships between geological structure and slope morphology. A thorough comparison of several regression algorithms, including Linear Regression (LR), K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGB), demonstrates the superior performance of nonlinear models combined with VAE-derived latent features. The findings indicate that incorporating these latent features significantly enhances prediction accuracy, representing a notable advancement over models based solely on traditional geometric parameters. This pioneering approach provides a reliable solution for accurately estimating key rockfall risk indicators at the base of highwalls, thereby enhancing the reliability and data-driven nature of risk assessments in open-pit mining operations. • Incorporates detailed material layers to enhance rockfall hazard predictions in open pit mines. • Generates a comprehensive synthetic dataset of rockfall indicators for robust model training. • Compares rigorously multiple regression models, including KNN, XGB, and SVM, to assess performance in predicting rockfall hazards. • Utilises a variational autoencoder for dimensionality reduction, improving the extraction of hidden features from profile data.