A new interpretable prediction framework for step-like landslide displacement

Machine learning models perform satisfactorily in landslide displacement prediction, but they are generally black-box models that are difficult to gain the trust of decision-makers. Therefore, a three-stage prediction framework based on the Hodrick–Prescott (HP) filter, double exponential smoothing (DES), natural gradient boosting (NGBoost), and Shapley additive explanations (SHAP) was proposed. The framework quantifies the uncertainty in the predictions and provides fully transparent outputs. In the first stage, the HP filter decomposes cumulative displacements into trend and period displacements. The second stage uses DES and NGBoost to predict them separately. In the third stage, we compute the SHAP values of the features to analyze the impact of the features on the model output. It is applied to the Bazimen and Baishuihe landslides in the Three Gorges Reservoir area and compared with other literature. The results show that the framework can achieve high accuracy in both point and interval prediction, and its performance is similar to or even better than other models. And it is easier to operate and applicable to a wider range of people. Most importantly, the framework can interpret the model, allows users to verify the consistency of the model prediction basis with the landslide evolution mechanism, and reduces random errors by optimizing the data. Overall, the prediction framework can capture the complex nonlinear relationships between features, solving the current challenge of machine learning models that cannot be applied in the final optimization scenario due to the lack of convincing, and providing a reliable tool for intelligent disaster prevention.