the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 26 Feb 2025
Sentinel-1 SAR-based Globally Distributed Co-Seismic Landslide Detection by Deep Neural Networks
Abstract. Rapid response to multiple landslide events (MLEs) demands accurate, all-weather, day-and-night detection capabilities. Optical remote sensing has advanced landslide detection but remains limited under adverse weather and lighting conditions. Synthetic Aperture Radar (SAR), resilient to these constraints, remains underexplored for automated landslide detection due to challenges such as complex pre-processing and geometric distortions. This study integrates Deep Neural Networks (DNNs) with SAR backscatter data for co-seismic landslide detection, utilizing a data-centric approach. We inform the models using 11 earthquake-induced MLEs, covering ≈ 73000 landslides across diverse geologic and climatic settings. Inference on unseen MLEs in Haiti (2021) and Sumatra (2022) demonstrates robust transferability, achieving F1-scores up to 82 %. Using explainable artificial intelligence, we highlight the discriminative capability of change detection bands over backscatter alone. Our findings emphasize the potential of SAR-based DNN models for worldwide, generalized, and rapid landslide detection, addressing critical gaps in current methods that solely use optical data. This research lays a foundation for broader applications in automated SAR-based earth surface change detection, particularly in complex, hilly and mountainous terrains.
- Preprint
(25430 KB) - Metadata XML
-
Supplement
(389 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on gmd-2024-230', Anonymous Referee #1, 24 Mar 2025
-
CEC1: 'Comment on gmd-2024-230', Astrid Kerkweg, 04 Apr 2025
Dear Authors,
it seems to me that you are using here a lot of datasets from different sources, and some processing with Google Earth Engine.
Please share the input data in an acceptable repository. If there are data products you are not able to share for good reasons, provide these reasons in the data availability section.
Best regards, Astrid Kerkweg (GMD executive Editor)
Citation: https://doi.org/10.5194/gmd-2024-230-CEC1 -
AC1: 'Reply on CEC1', Lorenzo Nava, 06 Apr 2025
Dear Astrid Kerkweg,
Thank you for your comment.
We have now updated the Zenodo repository to ensure all relevant input data are properly archived and accessible (https://doi.org/10.5281/zenodo.15159492). The repository includes:
- Shapefiles of all inventories used in the analysis
- SAR datasets used for the rapid assessment combination
- Model weights
- Processing code (already previously available)
We hope this addresses your concerns and are happy to provide further clarifications if needed. We will update the Data Availability section of the manuscript accordingly in the first review round.
Best regards,
Lorenzo NavaCitation: https://doi.org/10.5194/gmd-2024-230-AC1
-
AC1: 'Reply on CEC1', Lorenzo Nava, 06 Apr 2025
-
RC2: 'Comment on gmd-2024-230', Anonymous Referee #2, 10 Sep 2025
Review on “Sentinel-1 SAR-based Globally Distributed Co-seismic Landslide Detection by Deep Neural Networks”.
I have enjoyed reading this paper and believe it to be a valuable addition to the literature. As far as I am aware it is the first attempt to build a globally applicable landslide detection model using SAR data and deep learning techniques and the results look very promising. There are a few minor revisions that I think are needed before the paper can be published. The main one is the use of the Haiti earthquake as a test case – I am concerned that since the optical inventory was compiled after a tropical storm passed over the area, the landslides detectable in the SAR images acquired before this storm might not match the landslides mapped with the optical (see my comment on line 266 for more detail)
Line 55 “the issue of transferability in different settings and with different satellite data persists” you have addressed the first part, but since you only use Sentinel-1, the second limitation remains
Line 59 “Instances where SAR and DL are combined remain rare.” This is true, but there are a few more examples you could include
- Liang et al. (2025) use deep learning with polarimetric ALOS-2 SAR data to detect landslides - although the requirement for quad-pol SAR makes their work less widely relevant than yours since such images are not available for many earthquakes https://doi.org/10.1016/j.rse.2025.114904
- Chen et al. (2024) use deep learning for landslide detection with Sentinel-1 images https://doi.org/10.1080/17538947.2024.2393261
Then there are several studies using more basic machine learning methodologies that may or not be relevant here
- Ohki et al (2020) use Random Forests with SAR and terrain variables for landslide detection for two events in Japan https://doi.org/10.1186/s40623-020-01191-5
- Burrows et al. (2021) use Random Forests with SAR and attempt a somewhat “global” model, although it only includes 3 events https://doi.org/10.5194/nhess-21-2993-2021
- Lin et al. (2021) use Object based image analysis and SAR images for landslide detection 1109/IGARSS47720.2021.9554248
Line 76 “across diverse geographic and geologic settings”. The majority of your events are in vegetated areas and you do not include any cases where snowfall might complicate your signal (e.g. the 06/02/2023 Turkey earthquake.
Line 127 “Notably, Sentinel-1b has been inactive since 2022 and it is in the process of being substituted by an equivalent platform” Sentinel-1c has been launched since you originally submitted this manuscript so you could update this sentence.
Line 172 “side of look” “Look direction” is more commonly used for this.
Line 177 The GEE script by Vollrath 2020 also carries out radiometric terrain correction converting from sigma0 (normalised in the ellipsoid plane) to gamma0 (normalised in the plane perpendicular to the local satellite look direction) did you also carry out this processing step on your data? Or did you only use the shadow and layover mask?
Line 178 “inventory filtered with ascending/descending distortion masks”. It would be useful to know how much of each study area is masked (could be as a supplement rather than in the main text)
Line 212 “VH data is not available for these three locations” as far as I know, VH data was not regularly acquired until late 2016, so this will not be a problem for any future events you applied your model to.
Figure 3 It is quite hard to see the blue boxes overlying the grey images. Maybe they could be a brighter shade like cyan
Line 266 “The Haiti case is particularly challenging due to its topographical and environmental variability”
Another reason Haiti is challenging is that it was followed a few days later by a tropical storm. Studies of this event noted that many landslides increased in size during this storm (making them easier to detect using your SAR methods) (e.g. Havenith et al. 2021 https://doi.org/10.5194/nhess-22-3361-2022).
The exact images used to compile the inventory are not given in the inventory of Martinez et al., but in a different study on this event, Havenith et al. (2021) state that only 10% of the study area is visible in images acquired between the storm and the earthquake, so it can be assumed that most of the inventory is done using images acquired after the storm. The inventory includes both earthquake-triggered and storm-triggered landslides.
On your ascending track, the first post-seismic SAR image was acquired after the storm so also includes both earthquake- and storm-triggered landslides. However, on the descending track, the first post-seismic SAR image was acquired before the storm and so only includes earthquake-triggered landslides.
In my opinion, it would be better to consider the earthquake and storm as a single trigger and so start the period for your post-event stacks on the 17th of August (when the storm passed over) rather than on the 14th (when the earthquake occurred). Otherwise you are comparing earthquake and storm-triggered landslides in the optical imagery with earthquake-triggered landslides only in the descending-track SAR image.
Line 272 Define the acronym SHAP
Line 287 “Moreover, our model’s performance does not improve with an increase in the number of pre-event temporal stacks, contrasting with findings reported by Handwerger et al. (2022)” Maybe this depends on which events are used as test cases? For example if the landscape experiences widespread snowmelt during springtime in the 2 months prior to an earthquake, then using a full year of amplitude data would be beneficial to mute this signal.
Line 289 “Increasing the difference between pre- and post-event stacks” it would be clearer to say “increasing the difference in size between pre- and post-event stacks”
Line 323 “However, it is important to note that the location of landslide-related information in SAR imagery does not always align with the location of landslides in optical imagery due to geometric distortions, which is a current inherent limitation of SAR data” This is true, but optical images do not necessarily represent the “true” location of landslides while SAR images give a “false” one. Studies such as Pokharel et al. (2021) https://doi.org/10.1038/s41598-021-00780-y demonstrate that different inventories of the same event do not agree even when all the landslides are mapped using optical imagery.
Line 367 “160 SE” these are therefore slopes facing away from the SAR sensor – it would be useful to state this.
Line 407 “generalized rapid co-seismic landslide detection” you should specify here “in vegetated areas” since you do not test your model on any earthquake in a more arid environment.
Typographic corrections
Line 34 “Numerous research” should be “Numerous studies” or “Extensive research”
Line 48 “using the amplitude information to identify and rapid landslide failures” maybe there is a word missing here?
Line 165 “Landslide polygons from the available inventories are de using optical data” “de” should be “detected”?
Line 245 “recently occurred” you can just say “recent”
Line 302 “Insights by Spatial XAI” should be “Insights from Spatial XAI”
Citation: https://doi.org/10.5194/gmd-2024-230-RC2
Model code and software
SAR-LRA: A Synthetic Aperture Radar-Based Landslide Rapid Assessment Tool Lorenzo Nava, Alessandro Mondini, Kushanav Bhuyan, Chengyong Fang, Oriol Monserrat, Alessandro Novellino, and Filippo Catani https://zenodo.org/records/14898556
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 531 | 192 | 21 | 744 | 37 | 26 | 57 |
- HTML: 531
- PDF: 192
- XML: 21
- Total: 744
- Supplement: 37
- BibTeX: 26
- EndNote: 57
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
See attached pdf for comments