Balancing speed and accuracy: Deep learning for rapid coseismic landslide mapping in large earthquake-affected areas

Rapid identification of coseismic landslides over large earthquake-affected areas is critical for emergency rescue and rapid disaster assessment, yet remains challenging due to the trade-off between timeliness and mapping accuracy. Existing approaches are often limited to small-scale applications or require extensive manual intervention, hindering their practical use in time-critical scenarios. In this study, we propose and systematically evaluate a rapid mapping framework for coseismic landslides using multi-source remote sensing data and four representative methods, including image differencing, maximum likelihood classification, object-oriented classification, and deep learning. Using the 2018 Mw6.6 Iburi earthquake as a case study, we assess both mapping accuracy and time efficiency under realistic emergency constraints. Results show that deep learning significantly outperforms traditional methods in balancing accuracy and efficiency, achieving an F1-score of 0.7294 and a Kappa coefficient of 0.7139, while completing large-area mapping within 1 min 42 s. In contrast, conventional methods either suffer from severe misclassification or fail to meet time requirements. These findings demonstrate that deep learning provides a practical solution for rapid, large-scale coseismic landslide mapping in emergency contexts. The proposed framework offers important implications for real-time disaster response, and highlights the need for improving model generalization and multi-source data integration in future studies.