Evaluating the Potential of AI-Generated Landslide Inventories for Hazard and Risk Management: Advancements and Limitations

Landslide inventories are fundamental for susceptibility mapping, hazard modeling, and risk management. For decades, the geoscientific community has relied on manual visual interpretation of satellite and aerial imagery for landslide inventory generation. However, manual methods pose significant challenges, including subjectivity in landslide boundary delineation, limited data sharing within the scientific community, and the substantial time and expertise required for accurate mapping. Recent advancements in artificial intelligence (AI) have spurred a surge in research on semi-automated and fully automated landslide inventory mapping. Despite this progress, AI-generated inventories remain in their developmental phase, with no existing models capable of consistently producing ground-truth representations of landslide events following a triggering event. Current studies utilizing AI-based models report F1-scores ranging between 50% and 80%, with only a few achieving over 80%, often limited to the same study areas used for model training. This highlights a significant research gap in the reliability and generalizability of AI-generated inventories for hazard and risk assessments. The geoscientific community must critically assess the accuracy and transferability of AI-generated landslide data to ensure their applicability in subsequent phases of landslide response and mitigation. Further collaborative efforts and benchmark datasets are needed to establish standardized protocols for validating AI-generated landslide inventories across diverse geomorphological settings.