<p>Globally, and particularly in China, coseismic landslides are regarded as one of the most devastating natural hazards, causing significant loss of life and unprecedented property damage. Therefore, the rapid availability of landslide inventories is essential for succeeding phases of landslide risk studies such as spatial distribution and susceptibility assessment. In this study, machine learning (ML) models, namely decision tree (DT) and support vector machine (SVM), and a deep learning (DL) model, namely U-Net, were experimentally used to recognize landslides in the Luding region of Sichuan, China, following a catastrophic earthquake on September 5, 2022. This analysis was based on pre- and post-earthquake high-resolution images from Gaofen-6, Sentinel-2, and Planet, as well as on 0.9-m-resolution unmanned aerial vehicle (UAV) photographs. For spatial distribution analysis, nine predisposing factors, including topography, lithology, and seismic factors, were derived to explore correlations between these factors and landslide occurrence. Progressive tree-based and ensemble ML models including DT, random forest (RF), and gradient boosting decision tree (GBDT), and their frequency ratio (FR)-enhanced models (i.e., FR–DT, FR–RF, and FR–GBDT), were implemented to conduct the landslide susceptibility assessment (LSA). The results showed that 4,636 coseismic landslides were perceived in the aftermath of the earthquake. The U-Net demonstrated outstanding performance, with accuracy of 96.05 ± 0.59% in Moxi Town and 92.04 ± 0.89% in Detuo Town. Furthermore, the receiver operating characteristic (ROC) curve showed that FR–RF achieved the best performance, followed by FR–GBDT, RF–DT, RF, GBDT, and DT. These findings confirm that incorporating well-considered non-landslide samples can significantly enhance the accuracy of LSA. The outcomes of this work have the potential to facilitate rapid and precise evaluations for effective hazard management and decision-making processes.</p>