The spatial heterogeneity of soil thickness in karst regions characterized by extremely shallow soils is significant, making it challenging to accurately predict soil thickness using traditional environmental covariates such as topography, land use, and vegetation cover. This study applied three machine learning models-Gradient Boosted Tree (BOOST), Random Forest (RF), and Support Vector Machine Regression (SVR)-to the Huajiang area, a typical karst landscape with extremely shallow soils. The results indicate that when relying solely on traditional environmental covariates, all three models yielded low R 2 values, with BOOST performing best at R 2 = 0.42, while RF and SVR recorded R 2 values of 0.21 and 0.22, respectively. The root mean square error (RMSE) values for all models were high, approximately 20. However, incorporating remote sensing band data as additional covariates improved the accuracy of all three models. Further inclusion of dynamic vegetation characteristics as covariates led to additional enhancements in model accuracy, with BOOST achieving the best performance, resulting in an R 2 of 0.74 and an RMSE of 13. Training each model 100 times and validating them demonstrated that the introduction of dynamic vegetation characteristics significantly enhances predictive performance across all machine learning models. Notably, the BOOST model exhibited the highest effectiveness, achieving R 2 values around 0.7 across the 100 validations, with optimal accuracy attained at a tree number of 300 (R 2 = 0.7, RMSE = 14.9). In conclusion, the integration of dynamic vegetation characteristics substantially improves the prediction of soil thickness spatial distribution in karst areas with extremely shallow soils.