In the mountainous region of the world, landslides are a major issue that significantly impacts the socioeconomic, infrastructure, livestock, and human lives. A landslide susceptibility map (LSM) identifies potential areas of landslide. The preparation of the LSM and the accuracy are both impacted by the training and testing datasets. Therefore, this study assesses the impact of various landslide sampling methods for LSM preparation as well as understanding the impact of the presence or absence of non-landslide points on its accuracy. Total 898 landslide points were divided into training and testing landslide points using random sampling, grid sampling, and stratified sampling (based on elevation, slope, and watershed). The training landslide points and the 10 causative factors are used for the preparation of LSM using frequency ratio (FR), information value (IoV), and multiple linear regression (MLR) methods. The testing landslide points collected by sampling techniques are validated using the precision, recall, F1 score, kappa, and area under the curve (AUC) in three cases. LSM prepared using watershed sampling produced acceptable values, the highest being AUC (0.84), F1 score (0.76), and kappa (0.82) for different cases. Furthermore, LSM prepared using random sampling obtained an acceptable accuracy in all the 3 cases. Moreover, the grid and elevation obtained lower value for all the statistical measure in all cases. Thus, this study provides major insight in application of stratified sampling based on watershed in the preparation of training and testing datasets for LSM studies.