Step-like landslide displacement prediction is influenced by multiple characteristic factors, and the feature selection models using different sampling methods significantly affect the prediction results. Taking the Linbang accumulation body landslide in the Houziyan reservoir area of the Dadu River as an example. Step-like landslide displacement was accurately predicted through the establishment of a novel integrated prediction model that combined stratified sampling-based sequential forward and backward selection(SFS, SBS), LSTM(long short-term memory network), and an attention mechanism. Traditionally, the random sampling method in feature selection was abandoned, and a stratified sampling approach was innovatively adopted for the first time. The stratified sampling method specifically addressed the regularity of periodic displacement, particularly the data distribution characteristics near step inflection points, which significantly enhanced the model’s ability to characterize features during critical deformation stages. Additionally, an attention mechanism was introduced, endowing the model with the capability to dynamically adjust its focus on historical time-series information, thereby enabling more precise prediction of displacement mutations. The results show evaluation metrics for monitoring points 1-5 and 2-5 as RMSE(root mean square error) values of 3.16 and 1.56, MAPE(mean absolute percentage error) values of 0.27 and 0.83, and R² values of 0.978 and 0.984.The model’s robustness was further validated using two typical step-like landslides in the Three Gorges Reservoir area: Baishuihe and Bazimen. It is evident that the stratified sampling approach combined with the attention mechanism can effectively enhance the prediction accuracy of step-like landslides.