Explainable Artificial Intelligence Based Displacement Analysis and Forecasting for Unstable Rock Slopes

Geohazards such as landslides, rock avalanches or rock falls from unstable slopes can seriously threaten human life and infrastructure. Monitoring unstable slopes coupled with real-time data analyses to assess the risk they pose and mitigate this risk is thus indispensable. Machine learning-based methods for analysing monitoring data recently significantly improved the forecasting possibilities for failure events. However, one major limitation of Machine Learning-based methods is that they primarily provide "Black Box"-models. These models can, for example, transform arbitrary input into a sequence of predictions, albeit without a transparent explanation of how the output is derived from the input. Even though State-of-the-Art Machine Learning often outperforms traditional failure forecasting methods, such as the Inverse Velocity method, this limitation greatly hampers the application of these methods in practice. Recent advances in eXplainable Artificial Intelligence (XAI) have led to the development of the field of Causal Artificial Intelligence. As opposed to many Machine Learning approaches which are based on Deep Neural Networks, XAI aims to offer transparent models that provide explanations for model outputs. We therefore propose a novel forecasting approach based on XAI, leveraging Graph Neural Networks and Kolmogorov-Arnold Networks. Our approach aims to learn a causal model of an unstable slope or one particular section of it, including slope-internal and meteorological factors that can be represented as a graph, visualising cause-and-effect relationships between the variables. As such, our goal is twofold, and we aim at (1) providing insight into the mechanisms driving slope displacement, and (2) using this information for explainable short-term forecasting by selecting only causally related features from all available data. We apply our method to two case study sites for displacement driver analysis and short-term displacement prediction and compare the model performance to recent State-of-the-Art models. Our method not only aligns with but even outperforms existing models in terms of prediction accuracy and offers, in addition, superior interpretability. The proposed framework provides crucial support for geohazard assessment and monitoring network design. Furthermore, the displacement prediction has great potential as standalone predictive network as well as for hybrid failure prediction methods, for example in combination with traditional long-term failure predictions such as the Inverse Velocity method. While developed with medium-scale rock sections in mind, the method may be adapted to larger rock volumes as well as slow-moving mass movements with failure potential in general. The usage of accurate and interpretable prediction models represents a significant advancement, overcoming the transparency issues of models generated by complex Artificial Neural Networks, ultimately contributing to improving Early Warning Systems.