Supervised Machine Learning Algorithms for Ground Motion Time Series Classification from InSAR Data

Round 1

Reviewer 1 Report

The authors present an approach to Ground Motion Time Series Classification from InSAR Data based on supervised machine learning. The study is interesting and well organized. However, there are some flaws in this study, and the details of the reviews of this manuscript are as follows:

1. Introduction: I am missing a good introduction, please consider citing the following works that were related to the application of geohazards monitoring using remote sensing and artificial intelligence, such as LiDAR and photogrammetry (you will find others in scientific databases):

The application of terrestrial LiDAR for geohazard mapping, monitoring and modelling in the British Geological Survey

Combining temporal 3-D remote sensing data with spatial rockfall simulations for improved understanding of hazardous slopes within rail corridors

Deformation monitoring of earth fissure hazards using terrestrial laser scanning

Validation of a new UAV magnetic prospecting tool for volcano monitoring and geohazard assessment

Rock Discontinuities Identification from 3D Point Clouds Using Artificial Neural Network

2. Please provide a location map of the study and area. In addition, please provide more information about the geological setting of the study area, which is highly related to the ground motion.

3. Why did the authors classify the ground motion into five categories: Stable, Linear, Quadratic, Bilinear, and PUE.

4. Could the authors apply the trained model to a new case to predict the ground motion class? It will be a good way to verify your proposed model.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This is an interesting paper to apply multiple supervised and unsupervised machine learning models to train and verify the newly released EGMS. The authors figured out that RF and EGB models could classify reference samples accurately and indicated shortcomings of the proposed models for classifying non-moving targets and lower accuracy for shorter TS. This paper will be attractive to those interested in ML and EGMS data. Here are some improvements the authors could make before publication:

1.      For the introduction, rewrite lines 36-39. Ground displacement is not introduced because of the release of EGMS. Additionally, why the authors choose ML to classify the TS is not explained clearly. Add more references to describe the strength of ML in classifying the TS.

2.      Line 106-107, “whereas the three other datasets were used for accuracy assessment and validation purposes”? How many datasets do the authors use in this paper?

3.      For the method, the MB approach may need more description, such as how it performs in other studies.

4.      Table 2 needs a more detailed caption. Such as what ACF_1 and DACF_1 mean is so hard for readers to understand.

5.      Table 3. Could you add a column to describe if the model is supervised ML, unsupervised ML, or DL?

6.      For section 4.1, could you add the time cost for each model so that readers can figure out the time difference among models?

7.      Line 346, “exceeding 0.11 for XGB and 0.9 for RF and SVM”, I assume 0.9 is a typo and should be 0.09.

8.      Figure 5. Add an appropriate legend for the first image (the representative of green and red dots). Also, what the highlight in the second image means is not clear.

9.      Figure 7. XGM should be XGB.

1.   For limitations and future works, the authors mention that “ An unsupervised learning approach is highly recommended,” but from the paper itself, I cannot make this conclusion. Do I miss any?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The authors propose machine learning methods for  ground motion timeseries classification from InSAR data. The analysis is quite extensive, however there are some aspects that need further attention:

1) have you used K-fold cross-validation for your models' evaluation?

2) the authors should consider to use methods to tune their models (e.g. bayesian optimization). In literature there are various works that use hyperparameters tuning method for timeseries problems (e.g., Kaselimi, M., Doulamis, N., Doulamis, A., Voulodimos, A. and Protopapadakis, E., 2019, May. Bayesian-optimized bidirectional LSTM regression model for non-intrusive load monitoring. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 2747-2751). )

3) Fig. 4 shows the normalized feature importance for the RF method, with XGBoost algorithm one can also extract the feature importance. Have you also tried this? What are the outcomes?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

My concerns have been addressed by the authors.

Reviewer 3 Report

My comments are addressed. I recommend this article for publication.