Spatiotemporal landslide forecasting through machine learning and perspectives of applications for early warning: a case study in Kvam, Norway

The literature is rich with applications of machine learning techniques for assessing landslide susceptibility maps, which are limited to spatial prediction only. However, aspects related to extending the application framework to space-time landslides forecasting remain largely unexplored.

To fill this gap, this study introduces an innovative dynamic (i.e., time-dependent) application of the Random Forest (RF) algorithm. RF, among its advantages, allows the calculation of the Out-of-Bag Error (OOBE, which measures the error that would be committed if a given input variable is excluded from the RF classifier) and to visualize the Partial Dependence Plots (PDPs, depicting the relationship between each class of an input variable and the model outcome). These indices were discussed in this study to explore the algorithm's logic and verify its reliability.

The dynamic methodology proposed in this study involves using a spatially and temporally explicit landslide inventory as well as identifying non-landslide events over space and time. This procedure allows the inclusion of dynamic variables such as cumulative rainfall, snowmelt, and their seasonal variability, as model input. It also allows the inclusion of traditional static parameters such as lithology and geomorphologic attributes. Another key contribution of this study is that the RF model, once trained and tested using landslide and non-landslide events identified over space and time, produced a predictor that was subsequently applied to the entire study area before, during, and after specific landslide events. For each selected day, a specific and time-dependent landslide probability map was generated, simulating a real-time application in a warning system.

A case study in Kvam, Norway, was selected because of the availability of a comprehensive rainfall-induced landslide inventory, and the two major landslide events that occurred in June 2011 and May 2013 were selected for the simulations. Various model configurations involving the augmentation of non-landslide events were investigated to assess the model's sensitivity. The resulting pixel-based probability maps were validated using the Double Threshold Validation Tool (DTVT), a promising validation method based on the aggregation of pixels into catchment areas.

The reliability of the model was verified, and several benchmark configurations for the dynamic application of the RF model were provided. The generated landslide probability maps exhibit the ability to distinguish ordinary situations (low probability values where no critical rainfall was recorded, and no landslides occurred) from high-risk events (high probability values where highly intense rainfall triggered several landslides). The validation tool employed demonstrates the model's good performance and defines a criticality level suitable for early warning purposes. This study represents a step forward in comparison to traditional landslide susceptibility assessments and demonstrates the applicability of a novel method for spatiotemporal landslide probability mapping through machine learning, with perspectives of application to early warning systems.

Work supported by PRIN-ITALERT project (PRIN2022 call - grant number: 202248MN7N) funded by NextGenerationEU