Interpretable co-seismic landslide prediction: Unveiling the potential of multidirectional peak ground acceleration

Current co-seismic landslide evaluations predominantly employ machine learning methods, with peak ground acceleration (PGA) serving as the primary covariate for assessing landslide impacts. However, existing research often overlooks the effects of vertical ground motions, focusing solely on horizontal PGA, which does not reflect real-world conditions. To address this gap, we utilize actual ground shaking data to calculate a more comprehensive set of multi-directional PGA parameters and explore various combinations of these directional PGAs. To investigate their impact on co-seismic landslides, we employ a generalized additive model that captures the complex relationships between environmental factors and landslide occurrence. This model not only incorporates different directional PGAs but also considers their interactions to elucidate their effects on landslide risk. A robust suite of methods is employed to validate the model's goodness-of-fit and the interpretability of covariate effects. Our experimental results demonstrate that integrating multi-directional and interactive PGA parameters significantly enhances prediction accuracy for co-seismic landslides, with results remaining interpretable. Furthermore, we examine the generalizability of this approach across multiple machine learning methods, with consistent validation outcomes across different models. This underscores the necessity of comprehensively considering multi-directional PGA parameters and their interactions in practical co-seismic landslide predictions.