the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 27 May 2025
Review article: Deep Learning for Potential Landslide Identification: Data, Models, Applications, Challenges, and Opportunities
Abstract. As global climate change and human activities escalate, the frequency and severity of landslide hazards have been increasing. Early identification, as an important prerequisite for monitoring, evaluation, and prevention, has become increasingly critical. Deep learning, as a powerful tool for data processing and analysis, has shown significant potential in advancing landslide identification, particularly in the automated processing and analysis of remote sensing, geological, and terrain data. This review provides an overview of recent advancements in the utilization of deep learning for potential landslide identification. First, the sources and characteristics of landslide data are summarized, including satellite observation data, airborne remote sensing data, and ground-based observation data. Next, several commonly used deep learning models are classified based on their roles in potential landslide identification, such as image analysis and processing, time series analysis. Then, the role of deep learning in identifying rainfall-induced landslides, earthquake-induced landslides, human activity-induced landslides, and multi factor-induced landslides is summarized. Although deep learning has shown certain successes in landslide identification, it still faces several challenges, such as data imbalance, insufficient generalization capabilities of the models, and the complexity of landslide mechanism research. Finally, the future directions in the field are discussed. It is suggested that by combining knowledge-driven and data-driven methods for potential landslide identification, deep learning holds broad prospects for future applications in this field.
- Preprint
(3917 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-2158', Anonymous Referee #1, 20 Aug 2025
Confidence in my review: I am confident in my review regarding the deep learning and remote sensing topics. However, I am not a landslide expert, and have less confidence regarding that aspect of my review. The editor and authors should keep this in mind while weighing my comments.
Summary:
This review article presents an overview of deep learning applications to landslide identification. The authors review data sources, deep learning models, applications to different types of landslides, current challenges, and opportunities in the area.
Strengths:
The paper is well-organized and includes all the key elements I would expect from a review article on this topic. The topic itself (the use of deep learning in identifying landslides) is a worthwhile area to have a review on. The discussion of landslide mechanisms in Section 4 is quite detailed and interesting (although needs more discussion of how deep learning has been applied, see Weaknesses). The recommendations in Section 6, particularly regarding data fusion, feel well-motivated and salient. The figures are mostly helpful and illustrative overall (see specific comments below), and I think would be useful for a landslide researcher who wants to know more about deep learning. The review of previous literature is extensive and can point curious readers in the right direction.
Weaknesses:
Throughout the paper, many claims made by the authors are unsupported by citations to relevant works (see below for several examples).
The figures are helpful but there are not enough references to them in the text. I believe I counted exactly one in-text reference per figure, and often they seemed out of place. It would be helpful to refer the reader to the appropriate figure more often.
I felt the paper provided good background on landslides, and good background on deep learning, but was missing emphasis on the intersection: the application of deep learning to landslides. I would have expected this to appear in Section 3 or 4. However, section 3 discusses deep learning methods (without much specific discussion of how previous works have utilized them for landslides), and section 4 is almost entirely about the actual mechanisms of landslides, rather than how deep learning can be beneficial here. Section 4 seems almost mis-titled in this regard. To be precise, discussions similar to lines 632-638 are what I would have expected to see more of: specific examples of specific methods applied to specific problems in landslide applications.
The most notable weakness of this work is the grammar and writing style, which is well below the acceptable standards of a journal paper. I found many grammar errors and several sentences which essentially repeated sentences just before them, among other issues (see below). The language was often vague, passive, and lacking focus. The authors need to carefully proofread their paper. This reads like a rough draft, not a publication-ready submission.
Specific Comments and Questions:
L28: I’m not sure what you mean by the “relativity…of potential landslides”. Could you clarify what is “relative” about potential landslides?
Figure 1 contains reference to several important satellites that are not explained or mentioned anywhere in the text (Sentinel, for example). The authors should briefly describe the satellites and what kind of images they produce, either in the caption or the main body of the text, where relevant.
While I appreciate the brevity of Section 1, I feel it would be improved by adding a paragraph summarizing the authors’ overall takeaways and findings from this review.
Section 2.1.1: I would recommend the section on SAR also discuss NISAR. This is quite timely; the satellite just launched, is the most expensive earth-imaging satellite ever, and part of its mission is to monitor and better understand natural processes on Earth. Further, the authors should mention that another benefit of SAR is that it can image Earth regardless of illumination (ie day or night) and weather conditions (eg cloudy), which is not true of optical remote sensing.
Section 2.1.2 would benefit from more citations to previous work. In particular, it needs more citations to support its statements. This also benefits readers who want to learn more. For example “Its application in geological hazard investigations dates back to the 1970s” and “currently capable of achieving spatial resolutions as fine as 0.3 meters or better” are both claims that have no supporting citation.
Section 2.1.2: I would recommend more detailed discussion of multi- and hyper-spectral images and their application to landslides. You briefly mention it but I feel that more discussion is warranted given their prevalence in earth monitoring (especially via deep learning).
Section 2 would benefit from a more discussion comparing and contrasting these different data sources. Lines 173-176 do a good job of this with SAR and GB-SAR. More discussion similar to this for other methods would improve this section, in my opinion.
L167: “However, due to the influence of various factors, the identification results may not always be fully accurate, leading to potential misjudgments.” As a review article, your job is to tell the reader what the “various factors” are! You should elaborate and cite sources about what can cause inaccuracies in identification.
Figure 2(b) and 2(c) are nice illustrations. However, it is not clear to me Figure 2(a) is trying to convey, particularly the top and bottom panels.
With regards to data, are there any benchmark datasets for landslide identification? If so, which models/methods are state-of-the-art on these datasets? this would be worthwhile to discuss, and if not, a strong recommendation to the community would be to construct such benchmark datasets to encourage further deep learning research in the area.
Throughout Section 3, when introducing a new architecture, the authors should specifically cite the paper that introduced that architecture. This both credits the original work and provides the reader with references to important papers. For example, ResNet should be attributed to Kaiming He et al., GCN to Kipf and Welling, and so on.
L273: What does “feature reuse” mean? Potentially rephrase or clarify.
L276: “...even with limited landslide training samples”. This claim is not supported by a citation. These large networks typically need a lot of training samples. Is there previous work that successfully applied ResNet or DenseNet with limited training samples and was successful? If so, how much training data did they use?
L277: “semantic segmentation” should be defined - it is an important concept that an ML novice may not know the meaning of.
L285-287: Why is DeepLab preferable to U-Net? Can you elaborate on what ASPP is or does?
L289-290: The reference to Figure 2 here is confusing. You are in the middle of discussing DeepLab, but there is no mention of deeplab in figure 2.
L297: What are temporal variation curves?
L298-299: Which studies have used attention mechanisms? This paragraph has no citations, so it is impossible for the reader to refer to work that has been done regarding attention.
L319: You should mention that one reason RNNs struggle to model long term dependencies is exploding and vanishing gradients, and that LSTMs avoid this via their special gate design.
L352: Transformers are hugely important in ML right now, so it should warrant more discussion here than the authors have given it, in my opinion. For example, elaborate on what self-attention is, and mention that that transformer-based architectures are state of the art in several areas right now (language, imaging, etc). Are there more works in landslides that have used transformers? I appreciate the author’s discussion of its computational limitations; this is key, and it may not be the right choice for all practitioners. You should mention that the main drawback of transformers comes from its quadratic complexity, and that this is a current area of research to alleviate this issue.
L445: No citations in this paragraph to back up your claim.
L448: To me, it’s a little backwards to introduce AEs here when you’ve already introduced VAEs. Organizationally, it would make more sense to introduce AEs earlier before you talk about VAEs.
Section 3.4: Based on the way this section is written, it should be titled “Anomaly Detection” rather than “data cleaning”. These are somewhat similar, but distinct topics. Anomaly detection is about identifying data that is outside the norm of its dataset. Data cleaning is about filling in missing values and ensuring overall data quality. To me, anomaly detection fits more with what the authors discussed, and seems more relevant to landslides (ie, detecting when a landslide is imminent due to some abnormality)
L505: Should cite the original GCN paper by Kipf and Welling
L513: Similar to my comment about AEs/VAEs, you should move this paragraph to when you first introduce Transformers, rather than re-introducing them here.
The description of the mechanisms of landslides in Section 4 is extensive and appreciated. However, not enough time is spent discussing the main point of the section: the application of DEEP LEARNING to identifying these landslides. Lines 632-638 are a good example of what I would have expected more of. This section needs to do a better job answering the questions: What kinds of model (and what data) are generally used for what applications? Why? And so on.
L690: “gated fusion units” are never defined, and I’m not familiar with what they are. Could you define them?
L692: Paragraph may benefit from discussion of “fine-tuning” from the ML literature.
Section 5.1 makes good points, but the concept of limited training data should be discussed earlier, for example when discussing data in section 2.
L805: Paragraph may benefit from discussion of Physics-Informed Neural Networks, which may be of interest to scientists in this area.
L833-841: Perhaps I am misunderstanding, but this information is already conveyed in Section 4 and Figure 7.
Figures 9 and 10 seem redundant to me. It is not clear why you need both instead of just one or the other. Moreover, Figure 10(d) seems completely unnecessary.
Technical Corrections (spelling, grammar, wording/writing style):
L9: This is not a complete sentence. “...such as image analysis and processing, time series analysis.” Did you mean to include an “and” before “time”?
L44: “Through the training of large-scale and multi-source data,” is not correct. You don’t train data, you train “with” or “on” data. A better way to phrase this would be “Through training on large scale and multi-source data…”
L46: You use “potential” twice in short succession. I would replace the first use of it with a different word. Varying the wording, while only a stylistic choice, will improve readability and flow in the article.
L86: Not a complete sentence.
L158: “Then, assisting…” is not a complete sentence.
L193: “TLS scanner” is redundant - without the acronym it would be “Terrestrial Laser Scanning scanner”.
L193: The comma after “mass” should be a dash “...landslide mass - that is, …”
L218-219: “...geotechnical strength parameters can be inverted.” The “inverted” should be “inferred”. Otherwise, I’m not sure what you mean by that.
L223-227: The sentences “With continuous…” and “Especially when…” essentially repeat the exact same point. I would remove one of them.
L257: This sentence repeats an earlier sentence (L237-238).
L258: grammar error, “arise” should be removed or the sentence should be reworded
L294: Typo
L297: should be “outputs”.
L304: “an” should be “can”
L315-317: Bit of a run-on sentence, you could just remove “for calculation to gain an understanding of the data at the current moment”
L318-319: multiple grammar errors, “struggle” should be “struggling” and “limit their” should be “their limited”, or need to completely rephrase.
L344-345: This repeats the exact same point as in L337-339.
L348: need an “is” after “GRU”.
L352: typo, you repeat “employs” twice
L365: “data distribution of data” is redundant
L434-435: “Although” is the wrong word choice here. Both clauses of the sentence highlight the strengths of diffusion.
L448: missing a period
L470: The comma after “factors” should be a period.
L485-486: These two sentences repeat the same point.
L521: Typo/incomplete sentence?
L522: Another typo/incomplete sentence?
L542, 692, 708: multi-factor
L752: grammar: landslides (plural)
L780: needs a space after the period
L913: “, the combination of static and dynamic data is realized” is redundant, you already say static data earlier in the sentence.
L928: “underperforming” should be “underperform”
L944-945: not a complete sentence. Maybe you meant “to” instead of “can”?
L958: “leverage” should be “leveraging”, “combine” should be “combining”
L1033: should be “advancements” (plural)
Citation: https://doi.org/10.5194/egusphere-2025-2158-RC1 -
RC2: 'Comment on egusphere-2025-2158', Anonymous Referee #2, 09 Sep 2025
The objective of this work is to review recent advances in the application of deep learning to landslide prediction and to highlight the challenges and opportunities in this field. The authors provide an extensive overview of existing model types, but the manuscript does not go into sufficient detail on the actual application of deep learning techniques to landslide prediction. I see potential in this review; however, it requires a thorough revision.
Along the manuscript, I noticed several unsupported statements and a consistent lack of citations. There is also redundancy in the information presented, numerous grammar errors, and a confusing structure. For example, definitions and mechanisms of landslides appear scattered across different sections, rather than being organized logically. Since the manuscript focuses on landslides, I recommend a restructuring of the paper along the following lines:
1. Introduction
2. Landslide definition
a. Landslide mechanisms
b. Type of landslides
3. Deep learning for potential landslides
a. Data sources and models
b. Applications
c. Challenges and Limitation
d. Opportunities
4. ConclusionsIn addition, there is excessive discussion on the general use of deep learning, without providing sufficient concrete examples of its application to landslide prediction. I recommend focusing on the models currently used (3.a, 3.b) and on the models that could be used and how they will improve the landslide identification in 3.d. Below I provide some detailed comments (note that I did not highlight all grammar errors):
• Introduction: The section lacks sufficient citations, and the objectives are unclear. I recommend rephrasing them for clarity.
• Line 28. What do you mean by relativity?
• Line 33. What do you mean by potentials? Do you mean driving factors?
• Line 57. What do you mean by remainder? Do you mean the structure of this paper? You don’t need to mention it.
• Line 61. Chapter 2. It is unclear whether the data sources mentioned are actually used in deep learning for landslide detection, or whether they could be used. If they are used, please provide specific examples.
• Line 84. The phrase stops in the middle of the sentence.
• Line 219-227. Theis read as introduction.
• Line 280. The citation refers to medical research. While cross-disciplinary examples can be useful, this seems out of scope in the current context.
• Line 300. Chapter 3.2. You talk a lot about each model but not the application to landslides. For example, give more details on the studies cited at line 336.
• Line 364. Chapter 3.3. There is a lot of information but not related to landslides.
• Line 434. Although: this expects something negative after.
• Line 436: widely applied: you need to give reference on what they are applied to.
• Line 510: missing reference
• Line 579. Thus?
• Line 612. Missing reference in the first phrase: “The Newmark model is..”
• Line 788. Although deep leaning model. Needs referenceCitation: https://doi.org/10.5194/egusphere-2025-2158-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,098 | 80 | 10 | 1,188 | 19 | 38 |
- HTML: 1,098
- PDF: 80
- XML: 10
- Total: 1,188
- BibTeX: 19
- EndNote: 38
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1