A Study of Small-scale Landslides Susceptibility Dynamic Assessment in South Chian: Merging SBAS-InSAR and Machine Learning

Liang, Zhu; Hou, Jingxin; Liu, Yang; Wang, Ting; Liu, Guochao; Si, Chunshuai; Wu, Jun

doi:10.5194/egusphere-2026-218

Preprints

https://doi.org/10.5194/egusphere-2026-218

Preprints

23 Feb 2026

| 23 Feb 2026

A Study of Small-scale Landslides Susceptibility Dynamic Assessment in South Chian: Merging SBAS-InSAR and Machine Learning

Zhu Liang, Jingxin Hou, Yang Liu, Ting Wang, Guochao Liu, Chunshuai Si, and Jun Wu

Abstract. Shallow landslides, triggered by factors such as heavy rainfall and human engineering activities, are characterized by their sudden occurrence, wide distribution, and small scale, which pose significant threats to human life, especially in southern China. Landslide susceptibility assessment (LSA) is crucial for disaster prevention, mitigation, and land-use planning. Traditional assessment methods, such as field surveys and statistical models, predominantly depend on static geological environment factors, which exhibit inherent limitations in capturing spatiotemporal information on dynamic surface deformation. This deficiency directly leads to a lack of timeliness in hazard assessment, which is particularly important for sudden disasters. In this study, small Baseline Subset (SBAS – InSAR) and machine learning were married to explore a dynamic method in LSA, and the Conghua district in Guangzhou was selected as the study area. First, a dataset consisting of 326 historical landslides and 10 static factors, and 2 dynamic factors in the area was prepared. Then, the dataset was divided into two parts, one for modeling training (80 %) and the other for testing (20 %). Third, factors and samples were involved in the modeling of Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost). Finally, the performance of these models was validated through the area under the curve (AUC) and compared with the models analyzed by 10 static factors only. The results show that the XGBoost model exhibits the best performance, with an AUC of 0.933, and the models considered dynamic factors all performed better than that of static factors only. This study demonstrates that merging SBAS-InSAR and machine learning can provide a reliable technical approach for dynamic assessment of small-scale LSA, which is of great significance for targeted disaster prevention and mitigation in rural and hilly areas of South China.

Received: 16 Jan 2026 – Discussion started: 23 Feb 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Zhu Liang, Jingxin Hou, Yang Liu, Ting Wang, Guochao Liu, Chunshuai Si, and Jun Wu

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-218', Anonymous Referee #1, 15 Mar 2026
This manuscript presents an approach to landslide susceptibility assessment that integrates SBAS-InSAR-derived surface deformation data with tree-based ensemble machine learning models (RF, LightGBM, XGBoost) in the Conghua district of Guangzhou, southern China. The central premise that dynamic deformation factors can improve upon static-factor-only susceptibility models is reasonable and addresses a recognized gap in the landslide susceptibility literature. However, the manuscript suffers from several methodological shortcomings and presentation issues that substantially weaken the reliability of the reported findings. Major revisions are required before this work can be considered for publication.
Major Comments:
The manuscript provides no information whatsoever on how non-landslide samples were generated. This is a critical omission. The selection of negative samples is known to significantly influence the performance and spatial predictions of machine learning-based susceptibility models (e.g., Dou et al., 2023, which the authors themselves cite; and https://doi.org/10.1007/s10346-020-01473-9; https://doi.org/10.3390/rs15123200). Were non-landslide points selected randomly across the study area? Were buffer zones applied around known landslide locations to avoid spatial contamination? Was any consideration given to ensuring geomorphological plausibility of the negative samples? Without this information, it is impossible to evaluate whether the reported AUC values and accuracy metrics are meaningful or artificially inflated. For instance, if non-landslide points were drawn predominantly from flat, low-elevation terrain far from faults and roads, the models would trivially learn to separate landslide from non-landslide locations based on topographic setting alone, inflating performance metrics without reflecting genuine predictive skill. The authors must clearly describe the negative sampling protocol and ideally test sensitivity to alternative sampling strategies.

The authors rely on a single random train-test split (stated as 80/20 in the abstract but 70/30 in Section 3.2.1, see inconsistency comment below) to evaluate model performance. A single hold-out partition is highly sensitive to the particular random split and does not provide a reliable estimate of model generalizability. The landslide susceptibility modeling community has broadly adopted k-fold cross-validation (or spatial cross-validation) as a minimum standard for performance evaluation. The absence of any cross-validation scheme means the reported accuracy metrics may not be reproducible under alternative data partitions. Furthermore, spatial autocorrelation between training and testing samples is not addressed. If landslide and non-landslide points in the training and test sets are spatially proximate, the models may benefit from information leakage rather than learning generalizable patterns. The authors should implement at a minimum stratified k-fold cross-validation (e.g., 5-fold or 10-fold), and ideally spatial cross-validation to account for spatial dependence.

The abstract states the dataset was divided into 80% training and 20% testing, whereas Section 3.2.1 states 70% training and 30% validation. The authors must clarify and reconcile this discrepancy.

The abstract and Section 3.2.2 refer to 326 historical landslides and 10 static factors, while Section 2.1 reports 361 landslides and Section 3.2.1 also states 361 samples. The number of static factors listed in Table 3 is 9 (slope, aspect, elevation, distance to faults, distance to rivers, distance to roads, profile curvature, maximum elevation difference, NDVI), not 10 as claimed. These numerical discrepancies must be resolved.

The manuscript provides no information on the hyperparameters used for RF, LightGBM, or XGBoost, nor on whether any hyperparameter optimization was performed (e.g., grid search, random search, Bayesian optimization). Machine learning model performance is highly sensitive to hyperparameter settings, and the absence of this information undermines reproducibility. The authors should report all key hyperparameters and describe the tuning procedure.

The authors frame their approach as a "dynamic" susceptibility assessment method. However, the SBAS-InSAR deformation data are summarized as a single cumulative deformation value and a single average deformation rate over the entire two-year monitoring period (Jan 2024–Dec 2025). This is, in effect, a static representation of a dynamic process. A truly dynamic assessment would involve time-varying susceptibility maps that update as new deformation data become available. For instance, producing monthly or seasonal susceptibility maps that reflect evolving deformation patterns. As currently implemented, the InSAR-derived variables are simply two additional static predictors appended to the existing feature set. The authors should moderate their claims accordingly or demonstrate a genuinely time-varying assessment framework.

The InSAR methodology section (3.1) presents textbook equations but omits essential processing parameters: coherence thresholds, unwrapping algorithm and parameters, atmospheric phase screen correction method, reference point selection, multi-looking factors, and filtering approach. These details are critical for evaluating the quality and reliability of the derived deformation products, particularly in a heavily vegetated subtropical environment where decorrelation is a major concern.

The SBAS-InSAR processing in Section 3.1 extracts only line-of-sight deformation, which represents a one-dimensional projection of the true three-dimensional displacement vector. The manuscript does not specify whether ascending, descending, or both orbital geometries were used. This is a significant limitation, particularly given the authors' own observation (Section 4.3) that high-deformation zones do not fully correspond to high-susceptibility areas and that some highly unstable areas occur in residential plains, a pattern likely reflecting anthropogenic vertical subsidence rather than slope-related displacement. Three-dimensional displacement decomposition, achievable by fusing ascending and descending InSAR observations, would help disentangle landslide-related motion from unrelated ground deformation processes such as consolidation or groundwater extraction. Moreover, decomposed horizontal and vertical displacement components could serve as more physically meaningful and discriminative conditioning factors for the machine learning models than a single LOS measurement, potentially reducing the misclassification issues evident in the current results. Since Sentinel-1 data are freely available in both orbital geometries over the study area, the authors should at minimum acknowledge the limitations of single-geometry LOS measurements and discuss how three-dimensional displacement retrieval could strengthen the proposed framework. Relevant references include, for instance, doi:10.1016/j.earscirev.2014.02.005, doi:10.3390/rs11030241, and doi:10.1016/j.geoai.2026.100061.

Minor Comments:
The title contains a typo: "South Chian" should be "South China."

Table 5 labels the last row as "ROC" rather than "AUC." The ROC is the curve itself; AUC is the scalar metric.

The claim that overall accuracy improved by 15–20% relative to traditional static methods (line 385) is not supported by the reported results. The largest AUC difference between a static-only and SAR-integrated model on the validation set is 0.933 − 0.882 = 0.051 (XGBoost), which corresponds to roughly a 5 percentage-point improvement, not 15–20%.

The reference list contains formatting inconsistencies, and some citations in the text do not match reference list entries. For example, "Breiman et al., 2021" is cited in the text (line 192), but the reference list entry is Breiman (2001). The references should be carefully checked for accuracy.
Citation: https://doi.org/10.5194/egusphere-2026-218-RC1
RC2: 'Comment on egusphere-2026-218', Anonymous Referee #2, 16 Apr 2026

Dear Authors,
I carefully read your manuscript on the use of InSAR data to improve landslide susceptibility assessment.
Although the idea is interesting, the topic should be addressed in a more rigorous and appropriate way.
You used Sentinel data, claiming that you can detect antecedent movements of slopes affected by small and shallow landslides, but you did not prove it. You used only two deformation maps calculated over a 2-year period, without any verification of the temporal correlation between slope deformations (if any) and landslide triggering. the deformation rate appears to be a relevant factor in your machine learning models, but no clear explanation is provided on how it is incorporated. Based solely on the manuscript, it is also possible that the models identify slow deformation rates as indicative of high susceptibility. Models set up is not well described and the reported AUC values can suggest some overfitting issues, furthermore, you presented only some statistics of the results, rather then the results themselves (what about the number of TP, TN, FP, FN?)
I am somehow doubtful on the possibility of a correlation between long-term slope deformation and shallow-landslide, since the latter are often triggered by out-of-scale events (extreme rainfall, earthquackes), that are not detectable by SAR data. Moreover, even if you recognized the influence of rainfall in the triggering of the landslide, you did not analyze it.
You did not describe the parameters to create the deformation maps: did you consider any coherence thresholds?
Overall,your claims are not supported by the results your presented and the method and data you used are not described in enoughn detail to sggest the manuscript for publication of further reviewing
Detaild comments are rpvided in the attached file.

Citation: https://doi.org/10.5194/egusphere-2026-218-RC2

Zhu Liang, Jingxin Hou, Yang Liu, Ting Wang, Guochao Liu, Chunshuai Si, and Jun Wu

Viewed

Total article views: 273 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
184	73	16	273	14	15

HTML: 184
PDF: 73
XML: 16
Total: 273
BibTeX: 14
EndNote: 15

Views and downloads (calculated since 23 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	86	21	6	113
Mar 2026	77	36	10	123
Apr 2026	21	16	0	37

Cumulative views and downloads (calculated since 23 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	86	21	6	113
Mar 2026	77	36	10	123
Apr 2026	21	16	0	37

Viewed (geographical distribution)

Total article views: 269 (including HTML, PDF, and XML) Thereof 269 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 27 Apr 2026

Short summary

Small, sudden landslides are common in South China, often triggered by heavy rain. Traditional ways of assessing such risks rely on fixed land characteristics and lack the ability to track real-time ground changes. The study combines advanced satellite monitoring technology and intelligent data analysis to develop a dynamic risk assessment method. This method provides a reliable way to assess small landslide risks dynamically.


Total:	0
HTML:	0
PDF:	0
XML:	0