Rainfall is one of the most important factors controlling landslide deformation and failure. State-of-art rainfall data collection is a common practice in modern landslide research world-wide. Nevertheless, in spite of the availability of high-accuracy rainfall data, it is not a trivial process to diligently incorporate rainfall data in predicting landslide stability due to large quantity, tremendous variety, and wealth multiplicity of rainfall data. Up to date, most of the pre-process procedure of rainfall data only use mean value, maxima and minima to characterize the rainfall feature. This practice significantly overlooks many important and intrinsic features contained in the rainfall data. In this paper, we employ cluster analysis (CA)-based feature analysis to rainfall data for rainfall feature extraction. This method effectively extracts the most significant features of a rainfall sequence and greatly reduced rainfall data quantities. Meanwhile it also improves rainfall data availability. <br><br> For showing the efficiency of using the CA characterized rainfall data input, we present three schemes to input rainfall data in back propagation (BP) neural network to forecast landslide displacement. These three schemes are: the original daily rainfall, monthly rainfall, and CA extracted rainfall features. Based on the examination of the root mean square error (RMSE) of the landslide displacement prediction, it is clear that using the CA extracted rainfall features input significantly improve the ability of accurate landslide prediction.