Earthquake-induced landslides often cause greater damage than the earthquake itself, underscoring the need for rapid and accurate hazard assessments to support emergency response and post-disaster planning. Yet, in mountainous regions where cloud cover is frequent, the acquisition of reliable remote sensing data remains challenging. To address this, we propose a three-phase landslide hazard assessment framework. The first phase (near real-time assessment) delivers preliminary results within hours after an earthquake by applying rapid models (Xu2019 and Shao2023). The second and third phases progressively refine the assessments by incorporating incomplete and then complete landslide inventories, thereby improving accuracy over time. This framework is demonstrated using the 2017 Jiuzhaigou and 2022 Luding earthquakes in Sichuan, China. The results indicate that in the first phase, the Shao2023 model outperformed Xu2019, achieving 87% accuracy for Jiuzhaigou, compared to 66.7% for Xu2019. For Luding, both models showed similar accuracy, at 75% and 77%, respectively. As coseismic landslide data became available, prediction accuracy improved, though persistent cloud cover reduced the reliability of some regions. In Luding, visibility issues and data gaps resulted in an underestimation of landslide-prone areas. In the third phase, the inclusion of comprehensive landslide databases significantly enhanced model performance, with both models achieving AUC values above 0.94. Predicted landslide areas closely matched observed distributions, confirming the models’ robust predictive capabilities when supported by complete landslides data.
DownLoad: