Accurate prediction of soil freezing depth is critical for ensuring the operational safety of the Sichuan-Tibet Railway (STR),as frozen soil dynamics play a significant role in subgrade deformation.This study developed a high-resolution freezing depth forecasting system for the STR,leveraging a random forest machine learning model combined with multi-factor fusion analysis and integrated numerical meteorological predictions.The system utilizes 11 key predictive variables—including multi-year accumulated temperature,air temperature,humidity,elevation,soil composition,and real-time freezing depth—to generate 5 km-resolution gridded forecasts of freezing depth up to 7 days in advance.Based on an extensive analysis of meteorological risk factors and 40 years (1980—2021) of climate data from the Xizang region,three key findings emerged:1) The spatial distribution of frozen soil in the Tibet section of the STR is strongly corelated with altitude.Seasonally frozen soil—characterized by winter freezing and summer thawing—is primarily distributed above 4 000 m in regions such as Bangda Grassland,Guoging (Baxoi County),and Lajiu (Lhorong County).These areas exhibit annual mean ground temperatures below -5.0 ℃,with seasonal freezing depths ranging from 2.0 to 3.5 m,significantly impacting embankment sections,particularly in the Bangda Grassland.2) Over the past four decades,the region has experienced a warming and wetting climate trend,accompanied by declining wind speeds.The maximum annual freezing depth has decreased at a rate of 0.37—11.55 cm per decade,with the greatest reduction at Qamdo Station and the smallest at Nyingchi Station.3) The random forest-based freezing depth prediction model demonstrated high accuracy and reliability.Validation using field measurements yielded a prediction accuracy of 96％,a TS score of 0.96,a 3％ missed detection rate,and zero false alarms.The model achieved a R² value of 0.74 and a RMSE (root mean squared error) of 66.34 cm,indicating strong consistency between predicted and observed values.