Abstract: Aiming at the problem of insufficient accuracy of joint roughness coefficient evaluated by single statistical parameter
and traditional machine learning,an integrated tree estimation method of rock joint roughness coefficient based on multiple
statistical parameters is proposed. Based on the data set of 112 rock joint roughness profiles,eight statistical parameters
representing the geometric shape of joint profiles are selected,and six representative ensemble tree models,including bag method,
random forest,extreme random tree,adaptive lifting,extreme gradient lifting tree and lightweight gradient lifting tree,are
constructed. Combined with Bayesian algorithm to optimize its hyperparameters,and compared with five traditional single machine
learning models,the applicability and performance differences of six ensemble trees in joint roughness prediction are systematically
analyzed. Finally,the influence of each statistical parameter in the ensemble tree model on the prediction results of
joint roughness is revealed by Shapley additive feature interpretation method. The experimental results show that the prediction
effect of the integrated tree model is better than that of the traditional single machine learning model,especially the extreme
random tree model. The mean square error is 0. 081 0,the average absolute error is 0. 066 6,the coefficient of determination is
as high as 0. 995 6,the prediction accuracy is high,and the generalization ability is strong. This study provides a method basis
and reference for the determination of rock joint roughness,and it is recommended to use the extreme random tree algorithm to
predict the joint roughness coefficient.

Key words: joint roughness,statistical parameters,ensemble tree,Bayesian optimization,interpretability analysis