<p>Landslide susceptibility assessment serves as a key measure for government institutions to develop strategies for preventing and mitigating landslide hazards. Landslide susceptibility maps are usually prepared using landslide models to distinguish the possible occurrence of landslides. Constructing landslide models for susceptibility assessment and improving the reliability of results is necessary. However, due to landslides being a minority class in real-world scenarios, the number of landslide samples is much fewer than non-landslide samples, leading to a significant imbalance in sample classes. Sample class imbalance may reduce the accuracy and reliability of landslide susceptibility assessments, which is a common issue in susceptibility assessment modeling. In this paper, a novel model named RUSBoost is proposed for solving the issue of sample class imbalance in landslide susceptibility modelling. Furthermore, Bayesian optimization is employed to enhance the performance of the RUSBoost model. The proposed method is verified by taking the Three Gorges Reservoir area as an example. To compare the performance of different models when confronting the issue of sample class imbalance, decision tree (auc = 0.752), random forest (auc = 0.813), RUSBoost (auc = 0.828) and Bayes-RUSBoost (auc = 0.845) were used in experiments. In addition to ROC, Bayes-RUSBoost also demonstrates the best performance on model evaluation metrics focusing on the positive class (landslide), with precision (0.889), recall (0.804) and F1 score (0.844). Results indicate that the RUSBoost model, after Bayes optimization, achieved promising performance in landslide susceptibility assessment.</p>