An integrated approach combining advanced multi-temporal robust satellite technique and deep learning for landslide detection and prediction

Landslides are among the most hazardous and complex natural hazards, causing extensive destruction, damage to natural resources, and loss of human life and property. In this context, an accurate landslide inventory map is critically important for the effective detection and prediction of this phenomenon, which has been the focus of recent research on landslide monitoring. Supervised classification methods are commonly employed to detect landslides from satellite imagery. However, the selection and labelling of training samples for supervised classification can be both costly and time-consuming. To address this challenge, the present study proposes a method termed the Robust Satellite Technique (RST) for the automatic selection and labelling of landslide training samples across four diverse regions: Japan, Germany, Slovenia, and Taiwan. These samples are subsequently used as inputs for deep learning-based U-Net models. The RST, implemented within the Google Earth Engine (GEE), performs a time-series analysis of land cover changes from 2016 to 2024 to detect anomalies associated with land cover change, using freely available Sentinel-2 imagery to generate landslide inventory maps classified into landslide and non-landslide areas. Notably, depending on the date of landslide occurrence, the time-series analysis of Sentinel-2 data was conducted separately for each study area. Statistical evaluation of the RST yielded mean Intersection over Union (mIoU) values of 0.87, 0.88, 0.84, and 0.87 for Japan, Germany, Slovenia, and Taiwan, respectively. The Dice Coefficient was 0.89, 0.89, 0.86, and 0.88 for these regions, respectively. Based on these inventories, deep learning U-Net models were developed, trained, and tested for each case study area. The U-Net models demonstrated mIoU values of 0.91, 0.95, 0.88, and 0.92, and Dice Coefficients of 0.93, 0.97, 0.90, and 0.95, for Japan, Germany, Slovenia, and Taiwan, respectively. These results indicate the effectiveness of the RST in generating reliable labels and landslide inventory maps. Moreover, the study demonstrates that the integrated methodology enhances the accuracy of automatic landslide detection. Given the challenges associated with collecting ground-truth data in landslide-affected and often inaccessible regions, the use of RST provides a promising alternative for generating inventory maps without relying on field surveys. This approach substantially reduces dependence on on-the-ground data collection. The findings of this study have practical implications for stakeholders and researchers in the fields of risk assessment, land-use planning, geomorphology, and agriculture, enabling better simulation of landslide dynamics in prone areas, with the potential to prevent events or mitigate their associated damages.