Abstract: To address the issues of limited sample size in landslide susceptibility assessment, this study selected Zigui County as the research area and constructed an evaluation index system based on 15 landslide influencing factors. A coupled landslide susceptibility assessment method integrating Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) data augmentation and a Graph Neural Network (GNN) was proposed. WGAN-GP was employed to generate high-quality synthetic samples to expand the dataset, while a K-nearest neighbor method based on feature-space similarity was adopted to construct the graph structure, enabling both synthetic and real samples to participate jointly in the information propagation process of the GNN. GNN models were developed using both the original and augmented datasets, and three commonly used machine learning models, namely Random Forest (RF), XGBoost, and Support Vector Machine (SVM), were selected for comparative analysis. The results indicate that, under the original dataset condition, the predictive accuracies of the models showed only minor differences, and the GNN model did not exhibit a significant advantage. After the introduction of data augmentation, the predictive performance of all models improved, among which the GNN achieved the most substantial improvement, with the AUC increasing from 0.8235 to 0.9426. In the susceptibility zoning results, the GNN+WGAN model yielded an LD value of 0.11 in the very low susceptibility zone and 4.03 in the very high susceptibility zone, indicating that its susceptibility zoning results are highly consistent with the spatial distribution of landslide points.These findings verify that the proposed method can effectively alleviate the problems of small sample size, thereby improving the accuracy and reliability of landslide susceptibility assessment.