Work

Landslide inventory enhancement using HLS-based forest cover change and a rule-based geospatial framework in South Korea

Year: 2026

Type: article

Abstract: Accurate landslide inventory data are essential for susceptibility assessment models and early warning systems. In South Korea, however, national inventories are commonly provided at the administrative address level and over broad occurrence periods, which limits their spatial and temporal usability for research and operational applications. This study presents an earth observation-based geospatial framework for refining landslide inventory records by integrating time-series forest cover change information derived from Harmonized Landsat and Sentinel-2 (HLS) data with national geospatial datasets and publicly accessible APIs. A preprocessing pipeline was developed to generate consistent analysis-ready HLS composites for the 2016-2024 period, providing nationwide coverage for annual forest cover mapping. Attention U-Net with Focal Loss was selected as the final model for forest cover classification, achieving a recall of 0.9424 and an F1 score of 0.9372. The resulting annual forest cover change maps captured various types of surface disturbance, including forest fires, forest management activities, and urban expansion. These products were combined with parcel boundaries, topographic variables, and rainfall records to refine an initial set of 5,015 landslide inventory records into 4,487 coordinate-based records with estimated occurrence dates. Approximately 40% of the refined records were supported by detectable forest-to-non-forest transitions, which provided direct spatial evidence for location assignment. These results indicate that freely available HLS time-series data provide a practical and cost-effective basis for improving metadata-based landslide inventories at the national scale, reducing spatial and temporal uncertainty in existing records, and supporting subsequent susceptibility modeling and early warning applications. The proposed framework may also be applicable to other regions where historical landslide inventories are maintained in low-precision metadata formats.

Source: Remote Sensing Applications Society and Environment

DOI: 10.1016/j.rsase.2026.102235