Influence of landslide inventory representation and land cover configuration on landslide susceptibility mapping: a comparative machine learning analysis from the Taslidere Catchment (Türkiye).

Landslide susceptibility mapping (LSM) is a fundamental tool for hazard mitigation and sustainable land management in mountainous regions. However, its predictive performance and reliability are strongly influenced by landslide inventory representation, land cover data quality, and validation strategy. This study evaluates how these factors affect machine learning-based LSM in the Taslidere Catchment (Rize, Türkiye) by comparing two landslide inventories and three land cover datasets. The first inventory is a conventional point-based dataset of 878 landslide locations mapped in 2015; the second is a newly developed high-resolution polygon-based inventory of 2247 shallow translational landslides. The three land cover datasets comprise CORINE (Coordination of Information on the Environment) Land Cover 2018, a standardised pan-European land cover product; a dataset derived from national forest stand maps; and the UASIS (National Land Cover Classification and Monitoring System), the national land cover product of Türkiye. Seven machine learning algorithms were implemented using conditioning factors derived from 10 m resolution raster datasets. Ensemble learning approaches, particularly extreme gradient boosting (XGB), random forest (RF), and categorical boosting (CAT), consistently achieved the highest predictive performance, with AUC values exceeding 0.94 for polygon-based inventory configurations; the best overall result was obtained with the polygon-based inventory combined with the forest stand-derived dataset (XGB AUC = 0.9471). Models based on the older point inventory produced substantially lower AUC values (generally from 0.78 to 0.83). Inventory representation therefore exerted a considerably stronger influence on model performance than land cover dataset selection, whose impact remained comparatively limited across configurations. The findings indicate that high-quality polygon-based inventories improve the robustness and spatial reliability of landslide susceptibility models, and the proposed multiconfiguration framework clarifies the relative influence of inventory representation, land cover data, and validation strategy on machine learning-based landslide susceptibility assessment.