IISER MOHALI

Abstract: Landslide inventories are critical inputs for susceptibility assessment, hazard modeling, and risk-informed decision-making. Traditionally, such inventories have been derived through manual visual interpretation of aerial or satellite imagery, a process that is labor-intensive, subjective, and often inconsistent across regions and time. The advent of artificial intelligence (AI), particularly deep learning, has enabled the development of semi-automated and fully automated landslide mapping frameworks that promise to accelerate inventory generation at unprecedented scales. Despite notable advancements, the operational integration of AI-generated inventories remains limited. Current models report F1-scores ranging from 50% to 80%, with only a minority exceeding 80% accuracy typically when tested on the same regions used for training. This indicates limited transferability across diverse geomorphological settings and event types. Moreover, inconsistencies in annotation practices, lack of benchmark datasets, and inadequate validation protocols hinder robust cross-comparisons and downstream use in hazard and risk modeling. This study critically assesses the capabilities and constraints of AI-driven landslide inventory generation. It emphasizes the urgent need for community-driven standards, open-access multi-region datasets, and rigorous evaluation protocols to improve model generalization and reliability. While AI holds transformative potential in post-disaster mapping and early warning systems, its outputs must be systematically validated and integrated with geoscientific expertise to ensure credibility and actionable utility in risk governance.