Abstract: A series of potential safety hazards caused by leakage have seriously threatened the construction and normal operation of underground concealed works. In order to develop new seepage measurement methods to reduce the seepage control accidents, a classification model of sonar seepage detection results based on gradient boosting decision tree (GBDT) is proposed. The ReliefF algorithm is used to select the features with significant contribution weights as the training data set, and the expert annotated data are used to develop the gradient lifting tree model for distinguishing reservoir seepage, well seepage and noise. The experimental results show that the classification model proposed in this paper has good recognition performance, and the accuracy in the training set is as high as 96.6%.