Abstract: Monitoring and early warning represent a critical technological approach for landslide hazard mitigation and disaster prevention.However, high false alarm rates remain widespread due to issues with monitoring data and threshold settings.By summarizing the main factors affecting the accuracy of landslide warnings, this study established a sample database of typical anomalous monitoring curves, employed machine learning methods to identify abnormal monitoring data in real time, and introduced a dynamic adjustment mechanism for warning thresholds.An intelligent warning methodology that incorporates landslide warning level corrections was developed and applied in the Aba Tibetan and Qiang Autonomous Prefecture.A comparative analysis on warning events from 2021 to 2024 was conducted.The results demonstrated that identifying anomalous data and accordingly adjusting warning levels can effectively reduce false alarms caused by such data.The false alarm rate for GNSS monitoring decreased from 42.58% in 2021 to 11.72% in 2024, while that for crack meters dropped from 42.58% to 19.32% during the same period.The overall effective warning rate improved by 16.90%, 14.79%, 17.64%, and 19.10% from 2021 to 2024, respectively.Owing to strong performance of machine learning in feature extraction from displacement monitoring data, the accuracy of identifying anomalous GNSS and crack meter data reached between 97.43% and 98.93%, validating the effectiveness of both the typical anomaly curve database and the dynamic threshold algorithm.The research outcomes can provide valuable references for similar landslide monitoring and early warning initiatives.