Assessing Landslide Susceptibility Prediction Performance with an Event-Based Inventory from the 6 February 2023 Türkiye Earthquakes

The devastating earthquakes of 6 February 2023 in Türkiye (Mw 7.7 and Mw 7.6) triggered widespread co-seismic landslides across the region. This study focuses on developing and validating a landslide susceptibility map (LSM) for a 38,500 km² area in southeast Türkiye, which represents 5% of the country's landmass. Using a pre-earthquake inventory and the random forest algorithm, nine geomorphological and environmental features, including altitude, slope, lithology, and distance to faults, were integrated into the model. Validation was performed with a co-seismic landslide inventory comprising 2,611 landslides identified through pre- and post-earthquake aerial photogrammetric datasets.

Internal validation with the test data randomly split from the training dataset demonstrated high accuracy (93.67%) of the model based on the pixel-level assessments. However, the independent validation using co-seismic landslides revealed challenges, particularly in regions with rare lithological units or incomplete pre-event inventories. Despite the very limited pre-earthquake inventory, an accuracy of 76% was achieved, although it resulted in a significant number of false non-landslide labels. Thus, the co-seismic landslides highlighted the importance of accounting for unseen features, such as rare lithological units in the modeling. In addition, the resolution of the digital elevation model (EU-DEM with 25 m resolution) used for the LSM production was different from the resolution of the DEM used for post-earthquake landslide delineation. The latter one was obtained from high resolution aerial stereo images. Since the size of landslides which can be determined with the LSMs have strong correlation with the DEM quality, the difference between the internal accuracy and the external assessment results can be partly attributed to the data source used for inventory compilation. Nonetheless, the EU-DEM was found suitable for regional LSM production, and higher resolution DEMs also introduce computational complexity for such a large region.

This study outcomes revealed the potential of integrating remote sensing, machine learning, and geospatial data to enhance regional landslide susceptibility mapping. The findings provide valuable insights for disaster risk reduction, urban planning, and mitigating the impacts of latent hazards in seismically active regions; while pointing out the importance of data quality, optimization of machine learning algorithms, and multi-temporal inventory analyses to improve predictive accuracy.