Travel distance prediction for rock avalanche based on machine learning

Rock avalanches are one of the most destructive geological phenomena in mountainous regions. Understanding the dynamics and characteristics of rock avalanche movement plays a crucial role in assessing the potential hazards. However, the prediction for rock avalanche propagation is still challenging. Our study used an inventory of rock avalanches from Central Asia containing 412 historical cases from 6 countries provided by A. Strom. Considering several input parameters, the machine learning-based approach of extreme gradient boosting with grid search optimization was proposed. Input parameters including confinement type, headscarp height, mean slope angle of headscrap, length and width of the headscarp base, source volume, and maximal height drop (Hmax) are analyzed and discussed. Our proposed model can multi-output the distance of propagation L and the total impacted area, which outperformed by comparison with other machine learning models. Eleven rock avalanche events in Uzbekistan were introduced to demonstrate that the proposed model can be applied to prediction for limited parameters. For future work, we intend to propose a Convolutional Neural Network (CNN) architecture that combines spatial inputs and metadata as input in machine learning. Spatial inputs including elevation, slope, aspect, curvature, and lithology were used for our proposed model. Additionally, the CNN-based deep learning approach might be possible to predict rock avalanches which are characterized by complex terrain with multiple source areas and diverging paths.