Modeling and prediction of landslides in Norway – a machine learning approach

The spatial distribution of landslide landforms provides critical information for predicting potential slope failures and generating susceptibility maps. While this approach is confined to the spatial domain and does not account for the timing of landslide events, it is highly valuable for spatial management and landscape evolution modeling. Effective implementation, however, requires not only a robust selection of predictors but also high-quality historical data on landslide occurrences, which serve as response variables for model training. Once a local model is established, the next step involves testing its applicability to new areas characterized by differing predictor ranges and variations in landslide features, such as shape and density. This is particularly important for landslide modeling in Norway, where the landscape, significantly reshaped during the Pleistocene, exhibits distinct topography and sediment deposits. Furthermore, the region's high-latitude setting imposes unique precipitation and temperature regimes, adding complexity to landslide prediction.
We applied machine learning techniques, including Random Forest and XGBoost, to identify the optimal model for calculating landslide spatial probability. Our analysis used databases of detected and mapped landslides from two regions affected by extreme precipitation events in 2019 and 2023. Model testing revealed low spatial transferability between regions, likely due to dataset quality and predictor characteristics. We examined multiple scenarios, including a global model incorporating landslides from both events. Key factors limiting prediction accuracy include the quality and quantity of historical landslide data, the range and properties of potential predictors, and the inherent characteristics of the response variable—namely, debris flows, which are highly elongated and tend to form clustered patterns.
The study has been supported through the NAWA Bekker fellowship (No BPN/BEK/2023/1/00055).