Snow avalanche susceptibility (AS) mapping is a crucial step in predicting and mitigating avalanche risks in mountainous regions. The conditioning factors used in AS modeling are diverse, and the optimal set of factors depends on the environmental and geological characteristics of the region. Using a sub-optimal set of input features with a data-driven machine learning (ML) method can lead to challenges like dealing with high-dimensional data, overfitting, and reduced model generalization. This study implemented a robust framework involving the Sequential Backward Selection (SBS) algorithm and a decision-tree based ML model, CatBoost, for the automatic selection of predictive variables for AS mapping. A comprehensive inventory of a large avalanche period, previously derived from satellite images, was used for the investigations in three distinct catchment areas in the Swiss Alps. The integrated SBS-CatBoost approach achieved very high classification accuracies between 94% and 97% for the three catchments. In addition, the Shapley additive explanations (SHAP) method was employed to analyze the contributions of each feature to avalanche occurrences. The proposed methodology revealed the benefits of integrating advanced feature selection algorithms with ML techniques for AS assessment. We aimed to contribute to avalanche hazard knowledge by assessing the impact of each feature in model learning.

Snow avalanche refers to rapid snow mass movement under the influence of gravity and is a frequently observed natural phenomenon in mountainous regions. An area affected by a snow avalanche consists of starting, track (or transition) and runout (or deposition) zones, which have different geomorphologic characteristics that need to be considered in hazard modelling. This study attempted to analyze the impact of the different zones in producing snow avalanche susceptibility maps (ASMs) with data-driven methods. The random forest (RF) method was applied to the data from Gross Spannort Mountain region (Switzerland) for this purpose. Avalanche inventories from two dates were manually delineated to separate the three zones and two RF models were trained with learning datasets; i.e. Inventory-A which includes all snow avalanche zones (original inventory), and Inventory-B with starting and track zones. The conditioning factors were defined based on the literature, data availability and study area characteristics. The trained models (Model-A with Inventory-A and Model-B with Inventory-B) were evaluated with the area under the receiver operating characteristic (ROC), precision, recall and F1 score. The results show that Model-A has AUC of 0.97, precision of 0.89, recall of 0.95 and F1 score of 0.92 and the Model-B has AUC of 0.98, precision of 0.90, recall of 0.94 and F1 score of 0.92 that indicate high prediction performances for both cases. Furthermore, feature importance values were calculated by the Mean Decrease in Impurity (MDI) method, and elevation, aspect and valley depth were found the most influencing conditioning factors.

The quality of landslide susceptibility maps is often assessed using a part of learning data that represents geographical and land use characteristics over a quasi-fixed time. However, when validated with multi-temporal landslide inventories, more realistic insights on the susceptibility maps can be obtained. In addition, extreme events may trigger landslides in regions which are not considered as landslide-prone. The February 6, 2023, Kahramanmaras Earthquakes (Mw 7.7 and Mw 7.6), also known as the disaster of the century, triggered numerous landslides. Amanos Mountains located in southern Türkiye were also within the earthquake-affected area and had a very small amount of inventory recorded in official databases. The aim of this study was to evaluate the performance of the random forest method for producing landslide susceptibility maps. The official inventory of General Directorate of Mineral Research and Exploration (MTA) was used for map production. The resulting susceptibility map was assessed using the co-seismic landslide inventory produced in the study. The model’s performance evaluated using a part of the learning data yielded high accuracy expressed with area under receiver operating characteristics curve (AUC), precision, and recall values and F1 score using (AUC = 97%, recall = 97%, precision = 96%, F1 = 98%). However, multi-temporal evaluation with co-seismic landslides showed that 80% of the landslide pixels with moderate, high, and very high susceptibility levels could be predicted with the model. The results suggest that special attention should be given to features underrepresented in the inventory, such as low altitudes and types of lithology.