Abstract: Earthquakes usually trigger a large number of landslides that seriously threaten the safety of people's lives and property. How to effective use multi-source remote sensing technology to rapidly and accurately classify post-earthquake landslides is one of the key technologies for emergency response to earthquake disasters. In this paper, based on the GEE platform, the co-seismic landslides in Luding county after a magnitude 6.8 earthquake are extracted using multi-source remote sensing data (optical images, digital elevation model, synthetic aperture radar images) combined with machine learning algorithms (support vector machine (SVM), random forest (RF), gradient boosted tree (GBT)). The results show that the RF model performs best when there are few features (overall accuracy OA=93.1%, Kappa=0.859) and the GBT model performs best when there are a wealth of features (OA=96.3%, Kappa=92.3). Topographic features had the highest importance for landslide classification, followed by remote sensing spectral indices, and SAR image features had the lowest importance. Based on the best landslide classification model GBT, this study obtained the distribution map of landslides in the seismic area, with a landslide area of about 25.86 km². The results of this paper provide an important reference for the rapid identification of seismic landslides, in terms of model and feature selection.

Key words: remote sensing, landside classification, GEE, machine learning, Luding earthquake