Spatial information of particle size fractions (PSFs) is primary for understanding the thermal state of permafrost in the Qinghai-Tibet Plateau (QTP) in response to climate change. However, the limitation of field observations and the tremendous spatial heterogeneity hamper the digital mapping of PSF. This study integrated log-ratio transformation approaches, variable searching methods, and machine learning techniques to map the surficial soil PSF distribution of two typical permafrost regions. Results showed that the Boruta technique identified different covariates but retained those covariates of vegetation and land surface temperature in both regions. Variable selection techniques effectively decreased the data redundancy and improved model performance. In addition, the spatial distribution of soil PSFs generated by four log-ratio models presented similar patterns. Isometric log-ratio random forest (ILR-RF) outperformed the other models in both regions (i.e., R2 ranged between 0.36 to 0.56, RMSE ranged between 0.02 and 0.10). Compared with three legacy datasets, our prediction better captured the spatial pattern of PSFs with higher accuracy. Although this study largely improved the accuracy of spatial distribution of soil PSFs, further endeavors should also be made to improve model accuracy and interpretability for a better understanding of the interaction and processes between environmental predictors and soil PSFs at permafrost regions.