ChangeFormer-Based Detection of Landslide-Damaged Areas Using Sentinel-2 Imagery in South Korea

Abstract

Landslides triggered by heavy rainfall have become increasingly frequent and severe, creating a need for the rapid and accurate detection of damaged areas for post-disaster response and recovery planning. This study developed a ChangeFormer-based landslide damage detection model using single-channel differenced Normalized Difference Vegetation Index (dNDVI) imagery derived from pre- and post-event Sentinel-2 data. Landslide reference data were used to construct a patch-based training dataset, and model generalization was evaluated in Sancheong-gun and Hapcheon-gun, Gyeongsangnam-do, Republic of Korea, where landslide damage was reported following heavy rainfall in 2025. As available reference data differed between these regions, region-specific validation strategies were applied. In Sancheong-gun, polygon and point reference data were used for quantitative validation. All 12 reference-defined damaged sites were intersected by the model predictions, corresponding to a site-level detection rate of 100%. Point-based assessment showed an increasing distance-based detection rate with increasing positional tolerance, reaching 87.9% within the 80–100 m tolerance range. This result was interpreted as positional agreement between the reference points and predicted damaged areas rather than as overall accuracy. In Hapcheon-gun, where official polygon- and point-based reference data were unavailable, qualitative external validation using drone imagery indicated that the predicted areas were generally consistent with locations interpreted as landslide damage. These results suggest that the proposed framework is effective for the post-disaster spatial assessment of landslide-damaged areas.