Abstract: Rainfall-induced landslides commonly display distinct,step-like deformation patterns.However,in practical early warning scenarios,each deformation step observed in monitoring data may be misinterpreted as the onset of a critical sliding phase,potentially resulting in false alarms.To enhance early warning accuracy,this study investigates the Machicun No.1 and Zaoshuwa landslides as representative cases of sudden slope failure and prolonged creep,respectively.First,the evolution of landslide displacement curves under rainfall is systematically analyzed.Characteristic inflection points are identified to delineate step-like intervals within these curves.Second,a landslide inflection point identification model is developed,integrating the single-slope change-point method with the Automatic Multiscale Peak Detection (AMPD) algorithm.The identified inflection points are fitted using the Hill function.A comprehensive early warning analysis is conducted using an improved tangent angle theory applied to the fitted curves.Results demonstrate that the proposed method outperforms conventional peak detection models in accurately identifying inflection points.The fitted curves effectively mitigate the influence of step-like deformation,enabling precise identification of the creep phase and robust trend forecasting.By employing cumulative acceleration and the enhanced tangent angle as key indicators,a four-level integrated early warning model is constructed.These findings offer theoretical insights and technical support for improving the accuracy of rainfall-induced landslide early warnings.