A Spatial Prediction Architecture Utilizing Machine Learning Models for Diverse Forecasting Demands in Coseismic Landslide Scenarios

Abstract Coseismic landslides often occur extensively within a short period after an earthquake, posing severe challenges to emergency response and disaster risk reduction. The existing prediction methods generally fail to balance the dual demands of timeliness and accuracy and exhibit limited generalizability across regions with different geological and tectonic settings. In this article, a staged spatial prediction architecture (SPA) for coseismic landslides is proposed. In the rapid prediction stage, the model leverages readily available topographic, geological, and simplified seismic factors, combined with transfer learning, to rapidly generate landslide probability maps under label-scarce conditions in the target domain. In the accurate prediction stage, the measured seismic parameters and fault distance are introduced, and the rapid prediction outputs are incorporated as prior information to enhance the spatial characterization capability of the model. The results show that the rapid prediction stage achieves an area under the curve (AUC) of 0.86 even without labeled data, which can be attributed primarily to effective feature adaptation via transfer learning, whereas the accurate prediction stage further improves the AUC to 0.91 via the integration of high-resolution factors and prior information, significantly improving the delineation of high-risk zones. An ablation study confirmed the distinct contributions of both the transfer learning and prior fusion components to the overall performance. Furthermore, the extension of the proposed framework to pre-earthquake coseismic landslide forecasting is explored, achieving an AUC of 0.85 and demonstrating its potential for use in broader disaster risk management scenarios.