Sequent data mining analysis for rainfall-based landslide events with the refinement of landslide samples and feature reduction

The objective of this study is to adopt the Bayesian network algorithm to predict sequent rainfall-induced landslides of Shimen reservoir watershed in Taiwan since 2004 to 2008. Previous landslide events are used as training data to classify the samples of next event. Two subjects are further explored in this study. The first is eliminating landslide runout from landslide samples, and the other is feature (variable) reduction. The former is performed in order to refine landslide samples for prediction improvement, because image-based interpretation cannot discriminate landslide and landslide runout area. The latter is to reduce redundancy of landslide events and variables. Experimental results demonstrate that the landslide runout elimination and feature reduction can improve the prediction accuracy and the computation efficiency while maintaining acceptable results in landslide detection and prediction.