<p>Conventional rainfall thresholds for debris flows rely solely on rainfall parameters, neglecting the influence of environmental factors. Moreover, the samples of rainfall events that did not trigger debris flows (majority) significantly exceeded those of the rainfall events that did (minority). Most studies overlooked the impact of imbalanced data on model construction. This study proposes a novel rainfall threshold model based on logistic regression and resampling methods. The model integrates rainfall intensity, duration, vegetation, and material distribution, overcoming the limitations of conventional rainfall thresholds. It further employs resampling to achieve data balance between majority and minority classes, enhancing the reliability of debris flow predictions. Applying this framework, we developed rainfall threshold models for typical debris flow catchments in the Wenchuan earthquake-affected area from 2011 to 2015. Validation results for the developed rainfall thresholds demonstrate robust predictive capability. External validation further confirms the models’ robust generalizability. Subsequently, a quantitative analysis of rainfall thresholds revealed that the 1-h and 24-h rainfall required to trigger debris flows increased 88.43% and 38.08%, respectively, from 2011 to 2013, before decreasing in 2015. Additionally, a power-law function was adopted to fit the calibration data and project the future change in rainfall thresholds. Results suggest that the 1-h and 24-h rainfall required to trigger debris flows may recover to pre-seismic levels in approximately 2032 and 2039, respectively. This research contributes to the technological innovation of geohazard early warning systems and provides a scientific foundation for long-term disaster prevention.</p>