the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Earthquake Building Damage Detection based on Synthetic Aperture Radar Imagery and Machine Learning
- 1Global Earthquake Model Foundation, Pavia, Italy
- 2Jet Propulsion Laboratory, California Institute of Technology, CA, USA
- 3World Bank Group, Washington DC, USA
- 4Earth Observatory of Singapore, Nanyang Technological University, Singapore
- 5Asian School of the Environment, Nanyang Technological University, Singapore
- 6School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore
- 1Global Earthquake Model Foundation, Pavia, Italy
- 2Jet Propulsion Laboratory, California Institute of Technology, CA, USA
- 3World Bank Group, Washington DC, USA
- 4Earth Observatory of Singapore, Nanyang Technological University, Singapore
- 5Asian School of the Environment, Nanyang Technological University, Singapore
- 6School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore
Abstract. This article presents a framework for semi-automated building damage assessment due to earthquakes from remote sensing data and other supplementary datasets, while also leveraging recent advances in machine-learning algorithms. The framework integrates high-resolution building inventory data with earthquake ground shaking intensity maps and surface-level changes detected by comparing pre- and post-event InSAR (interferometric synthetic aperture radar) images. We demonstrate the use of ensemble models in a machine-learning approach to classify the damage state of buildings in the area affected by an earthquake. Both multi-class and binary damage classification are attempted for four recent earthquakes and we compare the predicted damage labels with ground truth damage grade labels reported in field surveys. For three out of the four earthquakes studied, the model is able to identify over fifty percent or nearly half of the damaged buildings successfully when using binary classification. Multi-class damage grade classification using InSAR data has rarely been attempted previously, and the case studies presented in this report represent one of the first such attempts using InSAR data.
Anirudh Rao et al.
Status: final response (author comments only)
-
RC1: 'Comment on nhess-2022-125', Samuel Roeslin, 29 May 2022
Overall, a very well written paper that will appeal to earthquake engineers active in building damage assessment and the wider audience interested in the use of novel technologies (InSAR) and machine learning. Nevertheless, some points would benefit from additional explanations/clarifications. My comments are provided below.
Comment 1
Lines 175-176: “The SAR-derived DPMs published by the ARIA project are used as the primary remote-sensing proxy to identify surface-level changes that are potentially attributable to earthquake-induced building damage.”
Could you please clarify how you interpret the correlation between surface-level changes and earthquake-induced building damage? Would it be possible to have seismic building damage without significant change in the ground surface level? If so, how would you proceed to remotely detect building damage using InSAR or EO?
Comment 2
Lines 185-188: “The problem presented is one of multi-class classification, and two ensemble machine-learning classification algorithms are employed—the Random Forest classifier for the cases involving only numeric features, and Histogram-Based Gradient Boosting classifier for the cases which also involve categorical features amongst the selected building attributes.”
Could you please explain why you selected the random forest algorithm and histrogram-based gradient boosting classifier? Did you try any other algorithms? How did they perform?
Comment 3
Lines 188-189: “The models are trained with a 70% subset of the available data, and then the best-fit models are tested against the 30% hold-out subset.”
Could you please explain why you chose a 70%/30% for the training and testing set? Did you try 80%/20% for example? Was 70%/30% giving the best performance?
Comment 4
Lines 211-213: “Another important reason to undertake a binary damage classification exercise is that it permits the aggregation of building damage datasets from different events into a larger training pool.”
Could you please clarify what you understand under a “larger training pool” for a machine learning model?
Lines 325-329: “Cross-regional training datasets will also help greatly improve the performance of these models for earthquakes in new regions previously unseen by the model. By expanding the datasets used to train the ML damage classification models, we can transfer the learning from regions with more damage data availability to data sparse regions. Cross-regional training is also critical as it will ultimately make such damage classification models more robust as they can be more confidently applied to future disasters, which may affect regions the model has not been trained on.”
Could you comment on the performance of a ML model applied to region on which it has not been trained on? When creating a “larger training pool”, could you please explain how you would capture regional specificities?
This reviewer is of the opinion that each location/region has specificities (e.g., construction practices, seismic setting, ground conditions) and thus has concerns regarding the applicability of a "one-fits-all" ML damage prediction model.
Comment 5
Lines 320-323: " The training of the machine learning models happens prior to the disaster event, and the trained model can be deployed for damage detection following an earthquake as soon as the pre-event building inventory, ShakeMap, and DPM become available ".
Could you please clarify the process? Why does this information only appear in the conclusion? Did you try to train a ML model for a region prior to a disaster and test the ML model after the earthquake event?
Comment 6
Lines 342-343: “Code availability. The Python code and Jupyter notebooks used for the analysis are available at https://github.com/gemsciencetools/eodamage-detection under the GNU Affero General Public License (v3.0).”
GitHub repo accessed by the reviewer on 29 May 2022
Data and code were found for the Gorkha, Puebla, and Zagreb earthquake, as well as the central Italy earthquakes. However, no folder/code related to the Puerto Rico earthquake could be found.
-
RC2: 'Comment on nhess-2022-125', Rui Jesus, 20 Jun 2022
The paper describes a framework for building damage classifications after an earthquake that combines InSAR data, high-resolution building inventory data and earthquake ground shaking intensity maps. The classification is performed using two different strategies (multi-class and bynary classification) and two different machine learning algorithms.
I am a researcher in the field of computer science, particularly in machine learning. Therefore, this review will focus on issues related to machine learning techniques used to solve a problem in the area of natural disasters.
The article is well structured and organised. It is not difficult to understand the main ideas of the work done and it is well written. The work described represents a very valid proposal but the original contributions are minor. Anyway, if the proposed framework proposed is well evaluated and validated, for me, it can be accepted for publication. However, as the paper is, in terms of validation, does not follows the standards and requirements of this journal.
The authors claim innovation in the use of the Multi-class damage grade classification using InSAR data. I have nothing against but the results presents are weak to validated the method. I think authors should present results that clearly show the benefits of the strategy used when compared to other techniques proposed. The multi-class strategy seems to me very important but the results are weak and are not compared. Also, the selection building damage categories is a critical point, regarding the urgency of action. This should be more discussed and explained in the paper.
Deep learning techniques are been used to solve many problems in many fields and presenting good results or better than the previous methods. I found it very strange that the authors would talk about recent advances in machine learning and then not use deep learning techniques for classification.
I understand that in some situations it can be difficult to use deep learning techniques, particularly when there is not much data available for training. If that's the case, I think the authors should present results demonstrating this. Also, the data augmentation methods should be considered.
Anirudh Rao et al.
Anirudh Rao et al.
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 236 | 65 | 10 | 311 | 2 | 2 |
- HTML: 236
- PDF: 65
- XML: 10
- Total: 311
- BibTeX: 2
- EndNote: 2
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1