Construction and Quality Evaluation of AI-based Landslide Inventory Maps for the 2024 Noto Peninsula Events, Japan

Under a changing climate, landslides pose significant risks to human society and regional sustainability. Advances in remote sensing and Artificial Intelligence (AI) have enabled automated large-scale landslide mapping after events. However, most existing AI-based inventories treat landslides as a single class, overlooking typological differentiation and providing limited assessment of geographic accuracy, which are central concerns in Landslide Inventory Maps (LIMs) research. This oversight obscures geomorphological diversity and spatial heterogeneity, constraining their use in geomorphological investigations and analyses beyond simple detection. As a result, the lack of explicit evaluation of thematic and geographic accuracy in AI-based landslide inventories remains an unresolved scientific problem.

To bridge this gap, this study developed and evaluated AI-based multi-class landslide inventories derived from PlanetScope and SPOT-6 imagery in the northern Noto Peninsula, Japan, and validated them through two-point field surveys (Figure 1). The study area (37.40°–37.49°N, 137.01°–137.19°E) lies along the Sea of Japan coast of central Honshu, at the convergent boundary between the Okhotsk and Amurian plates, where active reverse faults have been repeatedly reactivated. In 2024, the area was affected by two major landslide-triggering events: an Mw 7.6 earthquake on 1^st January and an episode of extreme rainfall in September. These events generated widespread but contrasting slope failures and were used to construct two post-event LIMs.

Results indicate that the inventories consistently identify three failure types, falls, slides, and flows, with 1,677 landslides mapped after the earthquake and 2,511 after the rainfall. Landslide areas follow log-normal size distributions, with slides covering the largest total area and flows exhibiting the highest counts. Compared with the earthquake, rainfall triggered more numerous but generally smaller failures.

We further evaluated the thematic and geographic accuracy of the constructed LIMs against established geomorphological understanding. Falls preferentially occur near active faults and on conglomerates, where rock masses are mechanically weakened. Flows predominantly concentrate in natural valleys and headwater channels, and are associated with porous volcanic-ash deposits, favoring material mobilization. Slides mainly develop on sandstone–mudstone interbeds, reflecting contrasting mechanical properties along bedding planes. Slope is the strongest control, especially for falls and flows, with failures concentrated on steep slopes characterized by concave curvature and steep longitudinal profiles. These geomorphologically consistent patterns support both thematic and geographic accuracy.

Finally, suitability for AI applications was assessed using thirteen semantic segmentation models. DRANet and TransUNet achieved the highest accuracy (mIoU > 0.85), providing precise landslide boundaries suitable for geomorphological analysis and modeling. In contrast, SwinUNet and SegFormer offer efficient trade-offs (around 0.80 mIoU with <15 GFLOPs), making them more appropriate for rapid mapping and emergency response under limited computational resources. Overall, strong and stable model performance indicates that the validated LIMs can be effectively used for AI training and operational landslide mapping, providing a foundation for AI-based landslide inventories in both methodology and application.

Figure 1. AI-based landslide inventory maps of the northern Noto Peninsula.