Integrated data from multiple public databases and literature records on permafrost and non permafrost sites, primarily sourced from the Global Terrestrial Permafrost Monitoring Network (GTN-P); National Field Scientific Observation and Research Station for Special Environment and Disasters in the Cryosphere of the Northern Tibetan Plateau (CRS); Swiss Permot Monitoring Network (PERMOS); Observation data extracted from published literature; The Global Historical Climate Network (GHCN) is a meteorological station with an average annual temperature of 2-10 ° C, used to expand non permafrost locations. The environmental factor data mainly comes from the following high-resolution global datasets: temperature and precipitation are from WorldClim 2.1. To reduce the interference of interannual climate fluctuations on the results, the data are smoothed using a 9-year moving average (i.e. centered on the target year and extended by 4 years before and after); The altitude is determined using the Copernicus Digital Elevation Model (DEM); The number of snow covered days is obtained from the MODIS Northern Hemisphere daily cloud free snow coverage product statistics; The soil clay content comes from SoilGrids 2.0, with an original resolution of 250 m, which has been resampled to 1 km

 Based on data from permafrost and non permafrost stations, corresponding environmental factor information was extracted, including 20 environmental factors such as temperature, precipitation, snow days, altitude, latitude and longitude, and soil properties. During the model evaluation phase, 13 common machine learning classification algorithms were selected and compared through grid search and 40 rounds of 5-fold cross validation, ultimately determining the random forest model as the best solution. Based on the ranking of model feature importance, combined with feature combination testing, the optimal feature combination is determined, and the final model is constructed by training with data from all sites. Using a partitioning strategy to predict the Eurasian continent. The final generation includes five periods of 1 km resolution permafrost distribution maps for the years 2000, 2005, 2010, 2015, and 2020, each corresponding to a GeoTIFF formatted raster data file.

 (1) In the process of organizing site data, strictly screen out points with abnormal coordinates, unclear labels, or inconsistent regional climate conditions to ensure spatial distribution representativeness and classification accuracy. After resampling various grid factors, align them according to a unified grid system and remove missing value areas to ensure the integrity and consistency of input data.

    (2) The random forest model exhibits good stability and generalization ability. After adopting the optimal feature combination, multiple 5-fold cross validation results showed that its average accuracy on the validation set was 0.936, F1 score was 0.936, and both AUC-ROC and AUC-PR exceeded 0.93. The final prediction model has better evaluation results on the entire sample.

    (3) The predicted distribution of permafrost shows obvious characteristics of high latitude and high altitude control, which is consistent with the existing distribution pattern of permafrost. The boundary area transitions smoothly and has good spatial continuity. Effectively controlling error propagation and memory overhead during large-scale inference through block processing, ensuring the stability and practicality of the results.