针对碎石土边坡监测过程中滑坡稳定变形期与临滑阶段监测数据量严重不匹配，导致临滑阶段数据量偏小，从而产生的非平衡数据集造成预判不准确的问题，提出了一种基于DST随机森林的碎石土边坡临滑阶段地表位移的预测算法。首先，采用过采样和欠采样相结合的双采样技术（double sampling technique, DST）对地表位移中的非平衡数据集进行采集，然后，通过随机森林预测算法有放回的随机抽样进行预测，最后，通过实验得出预测结果。结果表明：DST随机森林预测算法相比于普通随机森林预测算法预测误差率降低到3.39%，证明双采样技术（DST）采集临滑阶段非平衡数据集的必要性。

In the monitoring process of gravel soil slope, there is a serious mismatch between the monitoring data of the landslide stable deformation stage and the pre-sliding stage, which leads to the small amount of data in the pre-sliding stage, and the unbalanced data set resulting in the inaccurate prediction. A prediction algorithm of surface displacement in the pre-slip stage of gravel soil slope based on DST random forest was proposed. Firstly, the threshold comparison method of surface displacement pre-slip was used to determine whether there is an unbalanced data set. If there is an unbalanced data set, the double sampling technology (DST) was used to collect the unbalanced data set, and the negative samples were oversampled to improve the proportion of negative sample data set. Undersampling of positive sample data is carried out to reduce the proportion of positive sample data set. Equal amount of random positive and negative samples were selected as the training set to be processed, and the random forest algorithm was built to test the processed training set, and the test set after training was compared with the predicted results before the collection of non-equilibrium data set. Finally, by comparing the error values and error rate before and after sampling, it is verified that the prediction error rate of DST random forest prediction algorithm is reduced to 3.39% compared with the ordinary random forest prediction algorithm (the prediction error rate is 4.66%), which proves the necessity of double sampling technology (DST) to collect non-equilibrium data set in the pre-slip stage. Finally, it is concluded that DST random forest algorithm can obviously improve the pre-warning effect of the pre-slip stage, and overcome the problems of voting average, stagnation and increasing error rate.