This study employed a complete pipeline to build several ML algorithms accounting also for class imbalance and the effects that different oversampling algorithms have on model performance. Introduction and state of the art are clear and well described. The different ML algorithms are individually reported and described in perhaps too much detail. The monitored landslide is presented only in geographical terms. It might be useful to provide more details on the characteristics of the landslide. Different predictive features from an in-situ monitoring station were used to train and predict future landslide movements. The soil movements were split into four classes. Different SMOTE variants were then used to oversample the minority class. However, it is not clear if the other two minority classes were oversampled, or if they were removed in subsequent analyses. The main results are synthesised in Tables 5 and 6. In my opinion, these two tables are not enough to convey the effect of oversampling. In most cases the data without oversampling returns better scores than the oversampled data in all the metrics, not allowing the reader to understand the cause. Furthermore, scores so close to 1 might suggest a data leakage between training and model testing. It could be worth it to revise the data-splitting procedure and implement the pipeline with cross-validation to avoid this issue. Chapter 7 is just conclusions; the critical investigation of results (i.e., the discussion) is completely missing.

This paper presents the development of machine learning (ML) models with oversampling techniques to address the class imbalance issue, essential to developing a robust soil movement prediction system.

The paper is well-written and easy to follow. I have some significant questions regarding the proposed methods:

(i) How much does the model parameters value change with different training data sets? Also, authors should take different training sets for their method evaluation

(ii) The results should also contain the accuracy of each class. A truth table will be beneficial to understanding the method's performance.

(iii) The RF model has 100 % performance for training that might be overfitting in the model. Please check the overfitting in the model

(iv) The paper should include the precautions authors took to ensure no information from training samples was mixed with the testing samples.

(v) It will be good to compare results without a balanced dataset versus a balanced dataset (using oversampling techniques) in a plot. Also, discuss the reasons why no oversampling performs well over oversampling in some cases.

Some minor comments:

Figure 1 text is hard to read. Please increase the font size of figure 1