Landslide susceptibility assessment (LSA) based on machine learning (ML) models and grid units has been gaining increasing interest all around the world. However, the quantity of landslide inventoryfor ML model training are generally limited because landslide samples are often represented by geometrical points, which reduces to a certain extent the accuracy of ML models. To solve this problem, this study proposes an adapted sampling strategy based on frequency ratio (FR) method to effectively enhance the information of both landslide and non-landslide samples to reach an improved ML-based LSA. The FR of landslide conditioning factors (LCFs) are first obtained, based on which an integrated sampling strategy is then implemented to generate enhanced datasets for ML training and testing. Two typical ML models of the random forest (RF) and support vector machine (SVM) are employed to construct LSA models based on the enhanced datasets. And, for validation, the modeling and prediction accuracy based on the traditional training dataset and improved datasets are compared. The results from a case study in Anhua County show that, compared with conventional RF and SVM models, the corresponding improved models exhibit a better performance in terms of accuracy indicators such as accuracy, precision, recall rate, and F1 value, as well as the receiver operating characteristic curve and the area under the curve. The proposed method provides a promising alternate for an accurate and reliable LSM.