<p>Landslide disasters frequently occur in Zigui County, Hubei Province, China, making it necessary to conduct landslide susceptibility assessment for early prevention. In machine learning-based landslide susceptibility assessment, the core task is to train both positive samples (landslide points) and negative samples (non-landslide points) to obtain accurate classification results. The selection of negative samples is crucial, as traditional random sampling methods may result in positive samples being misclassified as negative samples. Furthermore, traditional single machine learning models typically capture only a subset of the features in the data, leading to lower accuracy when applied to landslide susceptibility assessment. To address the above issues, this study proposes two negative sample sampling strategies: buffer zone sampling and Information Value (IV) model constrained sampling. Additionally, the landslide susceptibility assessment model based on a stacking strategy for heterogeneous ensemble learning is introduced. First, the susceptibility assessment indicator system is constructed based on data from 511 landslide points in the study area, along with 14 influencing factors such as elevation, slope, and aspect, etc. After the negative samples are processed using the two sampling strategies, they are combined with the positive samples to form the dataset for the machine learning models. Then, Support Vector Machine (SVM), Random Forest (RF), and Convolutional Neural Network (CNN) are used as base models, and the stacking strategy is applied to integrate these base models to obtain landslide susceptibility maps for Zigui County. Finally, the performance of each base model and ensemble model under different sampling strategies is evaluated using ROC curves, along with statistical analysis of the landslide ratio, area ratio, and landslide rate for each sub-region. The research results indicate that, regardless of whether the buffer zone sampling or the IV model constrained sampling strategy is used, the accuracy of the stacking ensemble learning model outperforms that of the three base models. When comparing the two negative sample sampling strategies, namely buffer zone sampling and IV model constrained sampling, it is found that the evaluation accuracy of all models improves after applying the IV model constrained sampling strategy. This demonstrates that the combination of stacking ensemble learning and IV model constrained sampling effectively enhances the performance of machine learning models, thereby improving the accuracy of landslide susceptibility assessment.</p>