The quality of the statistical-based and machine-based landslide susceptibility map depends highly on the dataset's quality for model development. When investigating the training samples in the susceptibility analysis, the unbalance area ratio between landslide and non-landslide in any given study area could be an issue in the model training procedure. Therefore, determining a suitable ratio for sampling data of landslide and none landslide can be important to optimize the modeling procedure and improve the quality of the landslide susceptibility map. So, this study introduces a practical method to reduce the uncertainty of none landslide sampling and also experiments with various ratios between landslide and none-landslide samples. The synthesis of time-series land surface disturbance index (produced by Landsat products), the bivariate statistical Frequency ratio (FR) with a budget of landslide, and the experience is considered trustworthy data for reducing the uncertainty when extracting non-landslide samples. In addition, to investigate the suitable ratio of the sample subset, the range from 1:1 to 1:10 of respective landslides and none-landslide are examined. The hybrid of Frequency ratio (FR) and artificial neural network (ANN) is applied in this study to conduct the landslide susceptibility analysis in the Thu Lum watershed in Lai Chau province, Viet Nam. Comparatively, for accuracy assessment, increasing the number of absence samples leads to the problem of specificity value (true negative rate) increase, but sensitivity (true positive rate) value change downward. Overall, the Area under ROC (receiver operating characteristic) curve decreases while we increase the portion of the non-landslide sample of the training dataset. Eventually, this research shows that the unbalance sample ratio does not produce a satisfying model. For example, the unbalance ratio can be obtained when directly using the actual landslide and non-landslide area ratio. On the other hand, a balanced ratio is recommended in this study for statistical-based and machine-based landslide susceptibility analysis because it generally produces a landslide susceptibility map with better model performance.