the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 01 Jul 2025
A Semi-Automatic Iterative Method for Freeze-Thaw Landslide Identification in the Permafrost Region of the Qilian Mountains
Abstract. In permafrost regions, freeze-thaw landslides (FTLs) are a typical geological hazard that poses significant threats to environments and infrastructure at local to regional scales. However, traditional visual interpretation and also new deep learning methods still have limitations in their ability to detect and recognize FTLs at high precision, especially for hidden and small FTLs. Here we propose a semi-automatic iterative recognition method that combines InSAR surface deformation, multi-source images, and topographic factors to achieve a more accurate FTLs dataset for the Qilian Mountain permafrost region. The methodology involves four key steps: (1) acquiring surface deformation data from SBAS-InSAR with a deformation rate threshold of ≥50 mm·a⁻¹; (2) statistically analyzing topographic factors based on an existing FTLs inventory to determine initial threshold ranges; (3) extracting overlapping mask regions of these factors; and (4) verifying FTL boundaries through visual interpretation of multi-source remote sensing images and iteratively optimizing the sample database until deformation rates stabilize. Results indicate that after five iterations, 98 new FTLs were identified, primarily consisting of hidden and small-scale FTLs. The method achieved a true positive rate of 93.3 %, indicating high accuracy. In addition, we found that areas with larger absolute values of deformation rate and higher seasonal deformations are more prone to FTLs. The application of this method demonstrates highly accurate and efficient FTL identification, providing a new technical approach for monitoring and assessing the FTLs.
- Preprint
(13057 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-2726', Anonymous Referee #1, 24 Jul 2025
-
RC3: 'Comment on egusphere-2025-2726', Anonymous Referee #3, 01 Jan 2026
The manuscript highlights a new semi-automatic iterative method for identifying freeze-thaw landslides (FTLs), but it is framed primarily as an inventory-generation and detection-method paper. Although similar semi-automatic or hybrid approaches already exist (e.g., Xia et al., 2022; Jiao et al., 2023), the iterative masking strategy is a good example of iterative refinement of existing inventories (Fig. 2). The iterative logic is communicated clearly (n.b. Table 2). Integration of SBAS-InSAR + NDVI + topography is pragmatic and evidently provides additional discovery power for locating FTLs.
Overall, the focus on hidden and small (<10 ha) FTLs is a notable contribution, and a validation (93.3% TPR) is unusually good for landslide inventory papers. That said, only 15 FTLs are used for accuracy assessment out of >260 features. This is small for claiming “high accuracy” at landscape scale. Indeed, the term “high accuracy” should be better qualified throughout the paper.
The paper’s primary contribution is methodological, however, and not cryospheric. Freeze–thaw landslides (FTLs) are treated largely as detection targets, not as manifestations of permafrost thaw mechanics. Crysophere readership will be particularly interested in the process questions. Evidence of seasonal deformation is used diagnostically, but it is not interpreted mechanistically (e.g., active-layer thickening vs. ice-rich thaw settlement). There is an opportunity here to explicitly tie InSAR signals to cryospheric processes, for example, not just landslide likelihood.
The iterative masking thresholds (elevation, slope, aspect, deformation rate) are derived from an existing FTL inventory and then used to identify additional FTLs. This leads to some potential for circularity and self-reinforcement, whereby the method preferentially identifies features that resemble the original training set. This is not explicitly acknowledged. For instance, a false positive rate of 0% is unusual, and may reflect the validation design rather than true performance. It is unclear how representative the validated sites are of the full range of terrain, deformation magnitudes, and landslide sizes. The paper would be strengthened if the method could be deployed across a greater range of terrain types.
At present, the paper mostly offers better mapping but not better process understanding. A greater engagement with cryospheric process literature would substantially strengthen the paper. By deepening the Discussion, it should be possible to provide more substantive insights into the FTL distribution problem by augmenting the process interpretation of freeze–thaw mechanics, and the cryospheric significance of deformation signals - beyond detection efficiency. Note that the Conclusions tend to repeat results, rather than synthesizing insight - this section could be refined further, as well.
Some further questions to consider:
What does a seasonal deformation amplitude of 9–21 mm (ln. 278) correspond to, in terms of active-layer thickness changes or thaw settlement?
How do deformation signals in FTL areas compare with stable permafrost terrain undergoing seasonal freeze–thaw?
To what extent is the approach transferable to other permafrost regions with different geomorphic or climatic conditions?
Are the observed signals indicative of progressive instability and/or mass wasting processes, or could they also reflect benign permafrost heave–subsidence cycles?
Citation: https://doi.org/10.5194/egusphere-2025-2726-RC3 -
AC3: 'Reply on RC3', gang wei, 10 Jan 2026
We thank the reviewer for this constructive and insightful comment.
1. Overall, the focus on hidden and small (<10 ha) FTLs is a notable contribution, and a validation (93.3% TPR) is unusually good for landslide inventory papers. That said, only 15 FTLs are used for accuracy assessment out of >260 features. This is small for claiming “high accuracy” at landscape scale. Indeed, the term “high accuracy” should be better qualified throughout the paper.
Response: We sincerely thank the reviewer for the valuable feedback. We agree and in response, we have clarified in Section 4.2 “Inventory Validation” that the number of validation samples is limited due to the complex topography and restricted accessibility of the study area. We have also qualified the term “high accuracy” to emphasize that it reflects the method’s strong identification performance under the current validation conditions, rather than implying an absolute, landscape-wide statistical precision. In future applications and broader implementation of the method, further validation with additional field data will be necessary to confirm its robustness.
Line228-233: To evaluate the accuracy of the InSAR-based semi-automatic iterative identification method, this study combined optical remote sensing imagery with field verification. The validation process consisted of two complementary parts: optical imagery verification and field-based investigation. Due to the complex topography and limited accessibility of the study area, validation was conducted on 15 representative freeze-thaw landslides across three different regions. Although the sample size is limited, these sites cover a variety of terrain conditions (including different slopes, aspects, elevations, and spatial scales), thereby providing a representative accuracy assessment. The two verification approaches are detailed below:
2. The paper’s primary contribution is methodological, however, and not cryospheric. Freeze-thaw landslides (FTLs) are treated largely as detection targets, not as manifestations of permafrost thaw mechanics. Crysophere readership will be particularly interested in the process questions. Evidence of seasonal deformation is used diagnostically, but it is not interpreted mechanistically (e.g., active-layer thickening vs. ice-rich thaw settlement). There is an opportunity here to explicitly tie InSAR signals to cryospheric processes, for example, not just landslide likelihood.
Response: Thank you for this insightful suggestion. To more explicitly link InSAR-derived deformation to cryospheric processes rather than treating landslides merely as detection targets, we have added Section 5.3, "Interpretation of Cryospheric Mechanisms Underlying InSAR Deformation Signals," to the Discussion. In this section, we provide a mechanistic interpretation of the observed signals: (1) Seasonal deformation amplitude is discussed as a reflection of the intensity of moisture phase changes and migration within the active layer (Chen et al., 2020), where high values indicate periodic structural disturbances to the soil, i.e., a "pre-damage" indicator for landslide initiation (Gruber and Haeberli, 2007). (2) Negative annual deformation rates (subsidence) are linked to long-term permafrost degradation, potentially resulting from ground ice melt or active layer thickening (Daout et al., 2017; Chen et al., 2022). By transitioning from "deformation detection" to "process correlation," we provide a mechanistic basis for understanding the permafrost mechanics driving these landslides. We also acknowledge the current limitations, specifically the challenge of distinguishing the relative contributions of thaw settlement versus active layer thickening, as well as the need for direct quantification of hydrothermal coupling, areas that define our future research direction.
Line346-358:
5.3 Interpretation of Cryospheric Mechanisms Underlying InSAR Deformation
Research findings indicate that FTLs are more prevalent in regions with high seasonal deformation amplitudes (>10 mm) and substantial absolute annual deformation rates (>20 mm·a⁻¹). This spatial distribution suggests that landslide occurrence is closely linked to surface deformation processes operating at two distinct spatiotemporal scales. The seasonal deformation amplitude primarily reflects the intensity of the freeze-thaw cycle within the active layer, physically rooted in the volumetric changes of soil water (Chen et al., 2020). Higher seasonal amplitudes correspond to more pronounced moisture migration and phase change processes, which exert periodic structural impacts on slope soils (Gruber and Haeberli, 2007). Conversely, negative annual deformation rates indicate ongoing permafrost degradation, typically manifested as ground subsidence resulting from ground ice melting or active layer deepening (Daout et al., 2017; Chen et al., 2022).By utilizing the InSAR annual deformation rate as a key diagnostic, this study provides an observational foundation for understanding the connection between landslides and permafrost mechanics. However, limitations remain in distinguishing the relative contributions of thaw settlement versus active layer deepening, and in directly quantifying hydrothermal processes. These aspects represent the primary scope and limitations of the current approach.
3. The iterative masking thresholds (elevation, slope, aspect, deformation rate) are derived from an existing FTL inventory and then used to identify additional FTLs. This leads to some potential for circularity and self-reinforcement, whereby the method preferentially identifies features that resemble the original training set. This is not explicitly acknowledged. For instance, a false positive rate of 0% is unusual, and may reflect the validation design rather than true performance. It is unclear how representative the validated sites are of the full range of terrain, deformation magnitudes, and landslide sizes. The paper would be strengthened if the method could be deployed across a greater range of terrain types.
Response: Thank you for pointing out the potential circularity of the method, which is an important consideration for any sample-based identification approach. The primary objective of this method is to address the challenge of limited initial samples in freeze-thaw landslide research. Its value lies in employing a semi-automated process to systematically construct a larger and more reliable inventory of freeze-thaw landslides using a limited set of high-confidence initial samples.
The design of the method incorporates the following two key aspects: 1) Statistical representativeness of the initial samples: The existing inventory used to determine the initial thresholds includes freeze-thaw landslides across different slopes, aspects, and elevations within the permafrost environment of the study area. We have tried to cover landslide types and distributions as comprehensively as possible, so that the statistically derived thresholds reflect the diversity of known landslides in the region. 2) Core screening role of InSAR deformation data: In the iterative process, the most important screening criterion is the regional surface deformation rate obtained from InSAR. This physical observation is independent of the morphological characteristics of the initial samples and covers the entire study area. Terrain factor thresholds then serve as a secondary, physically based constraint to further focus on areas with active freeze-thaw processes.
Thus, the iterative workflow can be summarized as follows: first, identification of surface deformation anomaly zones via InSAR; second, efficient focusing using terrain statistical patterns derived from the initial samples; and finally, confirmation of new landslides through visual interpretation. This process not only iteratively validates the initial features but also retains the ability to detect new landslides, including those with potentially atypical characteristics, through independent physical observation.
We agree that further validation of the method’s generalizability is needed. As noted in Section 5.2, the field validation in this study remains spatially limited. Metrics such as the reported 0% false positive rate should be interpreted within the context of the current validation design. While the validation sites were selected to cover a range of slopes, aspects, elevations, and landslide scales within the study area, we recognize that the representativeness relative to the entire complex and heterogeneous permafrost region could be further improved. In future applications of the method, emphasis should be placed on the diversity of initial samples and extending the approach to broader and more varied permafrost environments to systematically evaluate its robustness and adaptability.
We have added the following text to the manuscript to acknowledge these issues:
Line333-342:
Additionally, the thresholds used in the iterative identification process are derived from a statistical analysis of existing landslide characteristics. This approach carries a potential risk of circularity that may result in the method preferentially identifying landslides that resemble those in the initial training set. As a result, these thresholds may not encompass the full range of characteristics of all freeze-thaw landslides, especially those with atypical geomorphological or kinematic signatures (Kim and Park, 2020). For validation, sites should be selected to cover varied terrain and environmental conditions, and the results should be interpreted within the context of the specific validation design, as they may not fully represent the method’s universal performance across all terrain types and landslide forms. Therefore, future applications of the method should consider the diversity of the initial samples, and efforts should be made to cover a broad range of topographies, geomorphic settings, and permafrost environments to systematically evaluate its generalizability and robustness.
4. At present, the paper mostly offers better mapping but not better process understanding. A greater engagement with cryospheric process literature would substantially strengthen the paper. By deepening the Discussion, it should be possible to provide more substantive insights into the FTL distribution problem by augmenting the process interpretation of freeze-thaw mechanics, and the cryospheric significance of deformation signals - beyond detection efficiency. Note that the Conclusions tend to repeat results, rather than synthesizing insight - this section could be refined further, as well.
Response: We appreciate the reviewer’s suggestions. We have implemented the following substantive improvements in the revised manuscript: 1) We have enhanced the Discussion section by adding section "5.3 Interpretation of Cryospheric Mechanisms Underlying InSAR Deformation" where ee explicitly link the observed InSAR deformation results (high seasonal amplitudes and negative annual rates) with established cryospheric processes: interpreting seasonal deformation as a representation of the intensity of freeze-thaw cycles in the active layer, and annual subsidence as a signal of permafrost degradation (such as ground ice melting or active-layer thickening). We hope this addition strengthens the mechanistic interpretation of frost-thaw mechanics and elucidates the cryospheric significance of our InSAR observations. 2) We have completely rewritten the Conclusions section to provide a more in-depth synthesis. We now elaborate on utilizing InSAR deformation as a key diagnostic factor and also present key process-level insights, the systematic correlation between FTL susceptibility and the deformation results of cryospheric processes (freeze-thaw cycles and permafrost degradation). We furthermore address the limitations of the current method in quantitatively distinguishing the contributions of different mechanisms (thaw settlement vs. active-layer thickening) and identify this as a future direction for research focusing on the coupling of InSAR with thermal-hydro-mechanical models.
Line359-384:
6 Conclusions
This study introduces an InSAR-based semi-automated iterative method for identifying FTLs in the permafrost regions of the Qilian Mountains, effectively addressing the limitations of traditional optical interpretation and data-driven approaches in detecting small-scale and obscured landslides. By integrating surface deformation data, topographic constraints, and multi-source remote sensing interpretations, this method significantly enhances the completeness and reliability of regional FTL inventories. Beyond improving identification accuracy, the results further elucidate the cryospheric significance of InSAR deformation signals in permafrost environments. Our findings indicate that regions characterized by high absolute annual deformation rates (particularly >20 mm·a⁻¹) and substantial seasonal deformation amplitudes (>10 mm) are more susceptible to FTLs. These deformation characteristics are consistent with freeze-thaw dynamic processes, such as seasonal volumetric changes within the active layer and long-term ground subsidence induced by permafrost degradation. The spatial correspondence between deformation anomalies and FTL distribution demonstrates that InSAR provides an effective observational basis for linking surface deformation to permafrost mechanical processes.
However, the current methodology focuses on enhancing the identification accuracy and spatial characterization of FTLs rather than quantitatively resolving specific cryospheric processes. Although deformation effectively indicates slope instability related to freeze-thaw processes, relying solely on InSAR observations makes it difficult to distinguish the relative contributions of various mechanisms to surface deformation. These mechanisms include active-layer deepening, ice-rich permafrost thawing, and hydrothermal processes. Strengthening the coupling between deformation observations and permafrost thermal-hydro-mechanical processes is an opportunity for future research. Additionally, our iterative threshold identification framework relies on existing landslide inventories with a potential risk of circular reasoning and self-reinforcement. While this design improved identification efficiency within the study area, its applicability across diverse geomorphological and permafrost conditions requires further validation.
Overall, this study represents a preliminary application of an InSAR iterative identification method in a specific permafrost region. Future research should integrate ground temperature observations, ground ice content, and hydrothermal process models to strengthen the connection between InSAR observations and cryospheric mechanisms. Furthermore, experiments across a broader range of terrain types and permafrost conditions are necessary to systematically evaluate the robustness and generalizability of this method.
5. What does a seasonal deformation amplitude of 9-21 mm (ln. 278) correspond to, in terms of active-layer thickness changes or thaw settlement?
Response: Thank you for this important question. Our seasonal deformation amplitude on the order of 9-21 mm corresponds to the magnitude of surface heave-subsidence cycles commonly associated with freeze-thaw processes in the active layer of permafrost terrains. Previous InSAR-based studies have shown that seasonal deformation amplitudes of several millimeters to a few centimeters can result from volumetric changes driven by soil freezing and thawing, moisture phase transitions, and water redistribution within the active layer, without requiring large changes in total active-layer thickness (Daout et al., 2017; Chen et al., 2020).
Specifically, seasonal vertical deformation of ~10-20 mm can reflect frost heave during freezing periods due to ice lens growth, and thaw consolidation and partial settlement during the thaw season caused by ice melt and pore-water drainage. These processes may occur within an active layer that thickens by only several centimeters, particularly in ice-rich or fine-grained soils, where small changes in ice volume can produce measurable surface displacement (Gruber and Haeberli, 2007; Chen et al., 2022).
It is important to note that InSAR-derived seasonal deformation amplitudes represent integrated surface kinematic responses rather than direct measurements of active-layer thickness change. As such, the observed 9-21 mm deformation should be interpreted as an indicator of the intensity of seasonal freeze-thaw activity and associated mechanical weakening, rather than a direct proxy for active-layer deepening or net thaw settlement. Disentangling the relative contributions of active-layer thickening, excess ice melt, and hydrothermal forcing would require joint analysis with ground temperature observations, soil ice content measurements, and thermal-hydrological-mechanical modeling.
We have clarified this interpretation in the revised Discussion (Section 5.3), emphasizing that seasonal deformation amplitude serves as a physically meaningful but non-unique indicator of cryospheric processes influencing freeze-thaw landslide susceptibility.
6. How do deformation signals in FTL areas compare with stable permafrost terrain undergoing seasonal freeze-thaw?
Response: Thank you for this important question. In stable permafrost terrain undergoing seasonal freeze-thaw, InSAR observations typically show relatively small-amplitude with spatially coherent and temporally repeatable deformation patterns. These signals reflect frost heave during freezing and elastic or quasi-elastic settlement during thawing. Previous studies report that seasonal deformation amplitudes in such stable areas are commonly on the order of a few millimeters to approximately 10 mm, with limited interannual variability (Daout et al., 2017; Chen et al., 2020).
In contrast, deformation signals observed in FTL-prone areas in this study exhibit systematically larger seasonal deformation amplitudes (commonly 9-21 mm), higher absolute annual deformation rates, and spatial heterogeneity at the slope scale. This could suggest that seasonal freeze-thaw deformation is superimposed on the longer-term, partially irreversible deformation associated with slope instability. Such deformation patterns are consistent with enhanced mechanical weakening related to active layer thickening, excess ice degradation, increased pore-water pressure, and gravity-driven creep on steep slopes (Gruber and Haeberli, 2007; Zhou et al., 2022).Therefore both stable permafrost terrain and FTL areas will experience seasonal freeze-thaw deformation, but FTLs have amplified deformation magnitudes, less spatial coherence, and stronger coupling with topographic and gravitational controls.
7. To what extent is the approach transferable to other permafrost regions with different geomorphic or climatic conditions?
Response: The core concept is generally transferable, although the specific threshold ranges require adjustment depending on regional environmental conditions. The workflow (the InSAR-derived deformation signals, topographic constraints, and iterative masking) does not rely on region-specific parameters and is thus largely transferable to other permafrost regions, particularly in terms of using elevated annual deformation rates as indicators of freeze-thaw-related slope instability.However, the specific threshold ranges for the iterative identification process (deformation magnitude, elevation, slope, and aspect) are influenced by regional factors such as permafrost thermal regime, ground ice content, topographic setting, and climatic conditions. Consequently, if the method is applied to permafrost regions with substantially different geomorphic or climatic characteristics, these thresholds would need to be recalibrated using local landslide inventories or reference datasets. To address this issue, we have added a description of the technical basis for threshold determination in Section 3.1.2 (Semi-Automatic Iterative Method), including a subsection on the determination of physically based thresholds.
Line166-193:
(1) Determination of Physically-Based ThresholdsThe core of our method is to determine iterative thresholds based on physical constraints and statistical patterns to efficiently concentrate on potential FTL areas. Using the existing FTL dataset, we first conducted a statistical analysis of the distribution characteristics of topographic factors: aspect, slope, and elevation. Combined with regional environmental physical constraints, the masking thresholds for each factor were determined (Table 2).
The method for determining each parameter threshold is as follows:Slope Threshold (1-40°): Slope is a critical factor for FTL occurrence. Engineering-significant FTLs typically initiate on slopes steeper than 1° (Niu et al., 2016). The initial iteration used a range of 1-30°, covering 96.22% of the known samples. Considering studies indicate significant frequency ratios for landslide occurrence even in areas with slopes >40°, despite lower pixel counts (Akgun, 2012), the upper slope limit was expanded to 40° in later iterations to more comprehensively include potential landslides in marginal areas, ultimately achieving a coverage of 99.46% of samples.
Aspect Threshold (0-112.5° & 135-360°): The initial threshold was set to 0-315° to cover over 90% of the sample data. To avoid areas where excessive solar radiation significantly reduces freeze-thaw effects, the 112.5-135° range (south-facing slopes) was excluded in the final optimized threshold (Burnett et al., 2008; Fan et al., 2023). This optimized range covers 95.65% of FTL samples, effectively focusing on areas with strong freeze-thaw activity.
Elevation Threshold (3600-4400 m): This range corresponds to the permafrost distribution in the Qilian Mountains, with its lower limit based on the known lower boundary of permafrost (Wang et al., 2015). An elevation step size of 100 m was set to expand the identification range of potential areas. The upper limit of 4400 m covers 99.98% of samples while effectively excluding non-freeze-thaw-dominated mass movements in high-altitude regions (>4500 m).
Deformation Rate Threshold (|rate| ≥10 mm·a⁻¹): The initial threshold was set at 50 mm·a⁻¹ because areas with an absolute deformation rate ≥50 mm·a⁻¹ accounted for less than 1% of the known FTL area, effectively identifying anomalous deformation. Based on existing studies and the regional context, the stable deformation range in permafrost regions is defined as −10–10 mm·a⁻¹ (Meng et al., 2015; Chen et al., 2019). Consequently, the threshold was progressively relaxed in steps of 10 mm·a⁻¹, with the final minimum deformation threshold determined as 10 mm·a⁻¹.
By extracting the overlapping areas of these single-factor masks, a relatively small but high-probability prediction zone encompassing the vast majority of potential FTLs was obtained. This step significantly narrows the target area for subsequent analyses, providing a precise focus for the visual interpretation of multi-source remote sensing imagery and thereby greatly enhancing identification efficiency.
8. Are the observed signals indicative of progressive instability and/or mass wasting processes, or could they also reflect benign permafrost heave–subsidence cycles?
Response: Thank you for this important question In stable permafrost terrain, seasonal heave-subsidence cycles have small deformation amplitudes, with high spatial coherence, as well as strong temporal repeatability and low annual deformation rates, reflecting largely reversible freeze-thaw processes.By contrast, in FTL-prone permafrost regions there will be larger seasonal deformation amplitudes and higher absolute annual deformation rates, in addition to lower spatial coherence at the slope scale and strong spatial correspondence with steep slopes and landslide morphologies. These characteristics suggest that the observed deformation reflects seasonal freeze-thaw processes superimposed on longer-term, partially irreversible deformation associated with progressive slope weakening and gravitational adjustment, rather than general freeze-thaw cycles.
Nevertheless, we acknowledge that distinguishing general freeze-thaw deformation from early-stage instability remains challenging when relying on InSAR observations alone. Our approach is therefore intended to identify areas with elevated susceptibility to freeze-thaw-related slope instability, rather than to provide definitive diagnoses of active mass wasting processes. These limitations and the non-uniqueness of deformation signals are discussed in Sections 5.2 and 5.3 of the revised manuscript.
References:
Chen, J., Wu, Y., O’Connor, M., Cardenas, M. B., Schaefer, K., Michaelides, R., and Kling, G.: Active layer freeze-thaw and water storage dynamics in permafrost environments inferred from InSAR, Remote Sensing of Environment, 248, 112007, https://doi.org/10.1016/j.rse.2020.112007, 2020.
Gruber, S. and Haeberli, W.: Permafrost in steep bedrock slopes and its temperature-related destabilization following climate change, J. Geophys. Res., 112, 2006JF000547, https://doi.org/10.1029/2006JF000547, 2007. Daout, S., Doin, M., Peltzer, G., Socquet, A., and Lasserre, C.: Large-scale InSAR monitoring of permafrost freeze-thaw cycles on the Tibetan Plateau, Geophysical Research Letters, 44, 901-909, https://doi.org/10.1002/2016GL070781, 2017.
Chen, J., Wu, T., Liu, L., Gong, W., Zwieback, S., Zou, D., Zhu, X., Hu, G., Du, E., Wu, X., Li, R., and Yang, S.: Increased Water Content in the Active Layer Revealed by Regional-Scale InSAR and Independent Component Analysis on the Central Qinghai-Tibet Plateau, Geophysical Research Letters, 49, e2021GL097586, https://doi.org/10.1029/2021GL097586, 2022.
Zhou, B., Zhang, Y., Wei, S., Wang, Z., Zhu, W., and Xue, Z.: Slope Instability Analysis in Permafrost Regions by Shear Strength Parameters and Numerical Simulation, Sustainability, 14, 9401, https://doi.org/10.3390/su14159401, 2022.
Citation: https://doi.org/10.5194/egusphere-2025-2726-AC3
-
AC3: 'Reply on RC3', gang wei, 10 Jan 2026
-
AC1: 'Reply on RC1', gang wei, 10 Jan 2026
We thank the reviewer for this constructive and insightful comment.
General comment
1.This manuscript presents a novel semi-automatic iterative method for identifying freeze-thaw landslides (FTLs) in permafrost regions, particularly focusing on the Qilian Mountains. The method uses derivative data extracted from InSAR data and other multi-modal data, including ERA5, DEM, and optical RS imagery. The validation data set has 167 FTLs and 17 RTSs. They found 98 new FTLs in the study region. Essentially, this is a threshold-based, manual and iterative screening process to discover new FTLs. Although the method has novelty and succeeded in the study region, it remains questionable if this method could be easily adopted to other regions due to highly heterogeneous permafrost environments.
Response: We thank and fully acknowledge the reviewer's valid concern regarding permafrost heterogeneity. In response, we emphasize that our methodology was deliberately designed to address this challenge through a two-stage adaptive approach.(1) Regional-specific threshold calibration: Given the significant environmental heterogeneity inherent in permafrost landscapes, our methodology addresses this challenge through a multi-tiered statistical and physically-constrained approach. These thresholds were specifically tailored to the study region's characteristics.
(2) Iterative spatial optimization: The threshold-based iterative process serves as a necessary adaptation mechanism rather than a limitation. By progressively adjusting thresholds, it gradually expands potential hazard zones, enhances the efficiency of visual interpretation, and reduces false positive rates.
We agree that broader validation is required. This method is a semi-automatic iterative method, not a fully automatic method, given the heterogeneity of freeze-thaw landslides in permafrost regions. Nonetheless, this method is an effective semi-automatic approach that. saves the cost of traditional interpolation. Thus, it can still be used in other regions.
Specific comment
2.The biggest concern of this method is the ability to generalise to other regions. As the authors has claimed in the Limitation section, the thresholding values for Elevation, Slope, Aspect and Deformation rate are analysed and determined manually for each iteration with subjectivity, which means for almost every environmentally-distinct region, the manual tuning of these thresholds needs to be done again, and even in different years in the same region, these values won’t be guaranteed to be stable. This method worked in a tiny region, as it has been tested in the manuscript; however, compared to the entire Tibetan permafrost or Arctic permafrost, the region is way too small. It will be very questionable to apply this method to a vast region without extensive manual analysis of the values and thresholds.
Response: We sincerely thank the reviewer for raising this important concern regarding the generalizability of our method. We fully acknowledge that parameter thresholds require regional calibration due to permafrost heterogeneity. However, we emphasize that our threshold determination is not purely subjective but follows a rigorous process combining statistical analysis with physical constraints. While initial threshold ranges can effectively identify FTLs in similar permafrost environments, the optimized thresholds developed through our iterative process provide significant advantages in detection accuracy and efficiency. The refined ranges were determined based on the following scientifically-grounded technical rationale:The slope threshold was determined through a histogram-based statistical analysis of known freeze-thaw landslide (FTL) samples. The initial iteration employed a range of 1-30° slopes, encompassing 96.22% of the samples, which was progressively expanded to 1-40° in later stages, ultimately achieving a coverage of 99.46%. Engineering-significant freeze-thaw landslides typically initiate on slopes steeper than 1°, with a predominant concentration between 6° and 10°, as gentler slopes provide insufficient gravitational driving force to trigger failure under thaw-weakened conditions (Niu et al., 2016). This pattern is corroborated by regional landslide inventories: Peng et al. (2024) demonstrated that retrogressive thaw slumps (RTSs) in the permafrost region of the Qilian Mountains are primarily distributed on slopes between 5° and 20° or steeper. Similarly, Zhang et al. (2024) reported that FTLs along the engineering corridors in the northeastern margin of the Qinghai-Tibet Plateau predominantly occur on gentle slopes ranging from 7° to 25°. From a broader perspective, Xia et al. (2024) observed that RTSs across the entire Qinghai-Tibet Plateau tend to develop on gentle slopes, mostly below 8°. Notably, Deng et al. (2024) reported that 14 thaw slumps developed on gentle slopes of only 2°-8° also exhibited significant deformation. Given that slopes of 1° or less generate inadequate driving force for failure—even under significant strength reduction induced by freeze-thaw cycles—landslides are unlikely to occur in such terrain. Thus, slope gradient serves as a critical precondition for FTL occurrence, and 1° can be regarded as a practical lower bound below which landslide probability is negligible. During the iterative process, the upper slope threshold was maintained at 35° in multiple rounds, as this range already encompassed over 98% of the training samples. This allowed for prioritized optimization of other parameters—such as annual deformation rate and elevation—to enhance identification efficiency and accuracy. Furthermore, the slope gradient classification was informed by the framework proposed by Akgun (2012), which categorizes slopes into intervals of 0-10°, 10-20°, 20-30°, 30-40°, and >40° for landslide susceptibility assessment. Their findings indicated that even the >40° class, despite having a lower pixel count, exhibited a notable frequency ratio (0.47), confirming its relevance to landslide occurrence. Accordingly, in the final iteration of the present study, the upper slope threshold was extended from 35° to 40°. This adjustment was implemented to better incorporate potential landslide areas often located in steep slope-toe transition zones. Thus, the comprehensiveness of the landslide inventory is improved, ensuring the adequate identification of landslides occurring on the steepest terrain facets.
The aspect threshold was initially set to 0-315° (covering over 90% of the sample data) to avoid omitting potential landslide areas. Based on further analysis using solar radiation models and sample distribution patterns, the threshold was optimized to 0-112.5° and 135-360°. This range covers 95.65% of FTL samples, effectively focusing on areas with significant freeze-thaw activity. The 112.5-135° range (south-facing slopes) was excluded due to excessive solar radiation and significantly reduced freeze-thaw effects (Burnett et al., 2008; Fan et al., 2023).
The elevation threshold was set to 3600-4400 m. The Lower limit of 3600 m is determined based on the lower boundary of permafrost in the Qilian Mountains (Wang et al., 2015). The elevation step size was set to 100 m, and gradually expanding the identification range of potential hazard areas. The upper limit of 4400 m covers 99.98% of samples while effectively excluding non-freeze-thaw-dominated mass movements in high-altitude regions (>4500 m).
The deformation rate threshold (|rate| ≥10 mm·a⁻¹) was determined based on regional stability characteristics. For the initial threshold, it was found that areas with absolute deformation ≥50 mm·a⁻¹ accounted for less than 1% of FTL regions. Therefore, this value can effectively identify anomalous deformation areas. The step size was set to 10 mm·a⁻¹ to gradually explore potential landslides. Finally, the termination threshold was 10 mm·a⁻¹. Based on existing studies and regional context (Chen et al., 2019; Meng et al., 2015), the stable deformation range in permafrost regions is defined as -10-10 mm·a⁻¹.
The determination of thresholds for all parameters follows these three unified optimization principles to ensure a balance between scientific rigor and practical applicability:
1) Statistics-Guided, Physics-Constrained: Initial threshold ranges for each parameter are established based on sample quantiles (e.g., 85%-99%) and bounded by physically meaningful constraints. This ensures that the resulting thresholds are both practical and reliable.
2) Iterative Optimization from High to Low Confidence: Initial thresholds are set using high-confidence intervals (e.g., >90%) to ensure accuracy, and then progressively relaxed in fixed steps (e.g., 100 m for elevation, 10 mm·a⁻¹ for deformation) to improve landslide detection while controlling uncertainty.
3) Region-Specific Adaptation: Parameters are adjusted based on the study area's environment, using examples like the 3600-meter permafrost lower limit in the Qilian Mountains and the high solar radiation on southeast slopes. This process ensures consistency with regional environmental characteristics while supporting potential application in other permafrost regions.Based on the above, we added the below text on lines 166-193. We hope that this offers an adaptive identification method for permafrost regions, rather than for a specific area. This approach is to generate a comprehensive sample base, which can significantly benefit subsequent machine learning applications.
Lines 166-193:
(1) Determination of Physically-Based Thresholds
The core of our method is to determine iterative thresholds based on physical constraints and statistical patterns to efficiently concentrate on potential FTL areas. Using the existing FTL dataset, we first conducted a statistical analysis of the distribution characteristics of topographic factors: aspect, slope, and elevation. Combined with regional environmental physical constraints, the masking thresholds for each factor were determined (Table 2).
The method for determining each parameter threshold is as follows:
Slope Threshold (1-40°): Slope is a critical factor for FTL occurrence. Engineering-significant FTLs typically initiate on slopes steeper than 1° (Niu et al., 2016). The initial iteration used a range of 1-30°, covering 96.22% of the known samples. Considering studies indicate significant frequency ratios for landslide occurrence even in areas with slopes >40°, despite lower pixel counts (Akgun, 2012), the upper slope limit was expanded to 40° in later iterations to more comprehensively include potential landslides in marginal areas, ultimately achieving a coverage of 99.46% of samples.
Aspect Threshold (0-112.5° & 135-360°): The initial threshold was set to 0-315° to cover over 90% of the sample data. To avoid areas where excessive solar radiation significantly reduces freeze-thaw effects, the 112.5-135° range (south-facing slopes) was excluded in the final optimized threshold (Burnett et al., 2008; Fan et al., 2023). This optimized range covers 95.65% of FTL samples, effectively focusing on areas with strong freeze-thaw activity.
Elevation Threshold (3600-4400 m): This range corresponds to the permafrost distribution in the Qilian Mountains, with its lower limit based on the known lower boundary of permafrost (Wang et al., 2015). An elevation step size of 100 m was set to expand the identification range of potential areas. The upper limit of 4400 m covers 99.98% of samples while effectively excluding non-freeze-thaw-dominated mass movements in high-altitude regions (>4500 m).
Deformation Rate Threshold (|rate| ≥10 mm·a⁻¹): The initial threshold was set at 50 mm·a⁻¹ because areas with an absolute deformation rate ≥50 mm·a⁻¹ accounted for less than 1% of the known FTL area, effectively identifying anomalous deformation. Based on existing studies and the regional context, the stable deformation range in permafrost regions is defined as -10 - 10 mm·a⁻¹ (Meng et al., 2015; Chen et al., 2019). Consequently, the threshold was progressively relaxed in steps of 10 mm·a⁻¹, with the final minimum deformation threshold determined as 10 mm·a⁻¹.
By extracting the overlapping areas of these single-factor masks, a relatively small but high-probability prediction zone encompassing the vast majority of potential FTLs was obtained. This step significantly narrows the target area for subsequent analyses, providing a precise focus for the visual interpretation of multi-source remote sensing imagery and thereby greatly enhancing identification efficiency.3.Secondly, this iterative, threshold-based method is very similar to the process of a decision tree-based machine learning model. One could train an ML model using the same variables (Elevation, slope, aspect, deformation rate) with some FTL ground truths. The model will have the exact same input-output as your methods. The downside of the ML method is that it requires at least a few hundred/thousand ground truth FTLs for training, but the resulting ML model will have much better generalisability than manual thresholding. I would strongly recommend comparing your proposed method with a properly trained ML model on a significantly larger area to see the difference in performance and geospatial extrapolation ability.
Response: Many thanks for this suggestion. We fully agree that a well-trained machine learning (ML) model on a large sample set holds potential for superior generalizability. To directly address this concern, we conducted a comparison following the reviewer's suggestion, which rigorously evaluates the performance of an ML model against our proposed semi-automatic iterative method under the condition of limited initial samples.In this experiment, we trained a random forest model using all available initial FTL samples (184 polygons). Following the data standards of geospatial machine learning, we first implemented strict quality control of the samples, which included two key steps: (1) coordinate system and spatial alignment check to ensure precise registration between each polygon and the raster data, and (2) data integrity checks requiring that each sample must cover a sufficient number of valid pixels to support feature extraction. Some samples were excluded due to cartographic boundary deviations, coordinate transformation errors, or location within raster “no data” areas (e.g., InSAR decorrelation zones). The reduction in sample count from 184 to 121 visually demonstrates the sample filtering process in the training dataset construction pipeline (Figure S1). The resulting 121 high-quality samples formed the basis for model construction.
The experimental results reveal:The ML model exhibited limited generalization ability. When applied to an independent test set containing 79 newly identified FTLs, it achieved a recall of only 26.6% (Figure S2), indicating that over 73% of the new FTLs were not identified. This empirically demonstrates the significant limitations of directly applying ML methods in the early stages of permafrost landslide research under conditions of sample scarcity.
Feature importance analysis further revealed that the deformation rate was one of the least discriminative features for the model (importance: 0.184), significantly lower than elevation (0.449) and slope (0.223) (Figure S3). This contrasts sharply with our approach where the deformation rate serves as a core physical discriminant, indicating that ML models struggle to effectively learn this key physical mechanism when training samples are scarce.In conclusion, the value of our study lies in proposing an effective strategy to address the problem of sample scarcity. Our proposed semi-automatic iterative method acts as an efficient data augmentation tool specifically for the initial phase of research where samples are limited. By integrating physical mechanisms such as deformation thresholds with expert knowledge, this method can rapidly build a high-quality sample library, laying a solid foundation for training robust ML models with strong generalization capabilities in the future.
4.Another concern is the iterative nature of this method, and manual verification makes it very labour-intensive. It requires further assessment to understand the cost-effectiveness of this method.
Response: We thank the reviewer for raising this point. We fully understand the concern regarding methodological efficiency. As the reviewer correctly observes, our method does involve manual interaction steps, and a core part of our argument is that this "cost" is necessary but highly cost-effective for specific research stages.The method was not designed for established application scenarios with abundant labeled samples, but for the initial research phases characterized by sample scarcity. Compared to pure large-scale manual visual interpretation, our method significantly narrows down the candidate areas requiring expert attention through InSAR deformation data and multi-factor thresholds. Each iteration only requires interpreting a small number of high-probability areas, which represents an order-of-magnitude improvement in efficiency compared to systematic visual interpretation of the entire vast permafrost region.
In terms of a cost-benefit analysis, the "benefits" we emphasize are primarily manifested in the following two aspects. The core output of our method is a significantly expanded and high-quality FTL inventory. This study added 98 new FTLs through five iterations, expanding the sample library by approximately 50%. These verified new samples themselves hold high scientific value and provide key foundational data for potential future large-sample ML model training. Additionally, there is a scientific discovery benefit. The method enables the identification of hidden and small-scale landslides that are difficult to detect by traditional visual interpretation or small-sample ML models. This is crucial for understanding permafrost degradation processes and associated hazards, leading to new scientific insights. Therefore, we consider our method to bridge the gap between traditional visual interpretation and future fully automated ML applications.
5.Line140: 2x7(unit?) window
Response: During InSAR data processing, we applied a “multi-look” to the SAR images. Because raw SAR data is typically made up of rectangular pixels (with different resolutions in range and azimuth directions) and DEM data consists of square-pixel rasters, we adopted a 2:7 multi-look ratio (azimuth:range) to make the SAR pixels as square as possible. This processing yields SAR data with approximately 30 m resolution, matching the 30 m resolution of the DEM data and ensuring geometric consistency for subsequent deformation analysis. We have clarified this in the manuscript by indicating that it is “multi-look processing with two-by-seven in range and azimuth direction”6.Line 169 specify how you ‘statistically analyzing’ the factors, and how you determine the thresholds
Response: Thanks for allowing us to clarify the statistical analyses. For each topographic factor (elevation, slope, aspect), we conducted histogram-based statistical analysis using the known FTL points from our initial inventory. This involved generating frequency distribution histograms for each parameter, calculating the percentage coverage of FTL samples within different value ranges, and identifying the value ranges that contained the majority (typically >90%) of known FTLs. The specific thresholds were determined through a combination of statistical dominance and physical constraints. Initial thresholds were set to cover high-percentage ranges of known FTLs (e.g., slope 1-30° covering 96.22% of samples). Thresholds were physically validated against permafrost mechanics (e.g., slope >1° for sufficient gravitational driving force), and thresholds were progressively refined through 5 iterations to balance precision and coverage. The complete technical rationale for each parameter threshold has been provided in our response to Question 2 and added to the manuscript, hopefully demonstrating the rigorous, data-driven approach used in our methodology.7.Eq(1) and Eq (2) unit of the variables are not specified, and which of these are dimensionless quantities?
Response: We appreciate this careful observation. The parameter units for deformation-related terms (a, c, d, Aseasonal) are in mm; temporal terms b in mm/day and t in days; cycle T is dimensionless. All trigonometric inputs (2πt/T) are normalized and therefore dimensionless.This was modified in the manuscript (lines 218-220) as follows:
“…where a represents the initial deformation value (mm); b is the long-term trend term (mm/day); c and d (mm) are the amplitudes of the cosine and sine terms, respectively; T is the soil freeze-thaw cycle (dimensionless, fixed at 365 days); the 2π/T term is dimensionless; and t denotes the time span with respect to the first SAR acquisition time (days).”And on line 223: “where Aseasonal (mm) represents the seasonal deformation value of each pixel.”
References:
Akgun, A.: A comparison of landslide susceptibility maps produced by logistic regression, multi-criteria decision, and likelihood ratio methods: a case study at İzmir, Turkey, Landslides, 9, 93-106, https://doi.org/10.1007/s10346-011-0283-7, 2012.Burnett, B. N., Meyer, G. A., and McFadden, L. D.: Aspect-related microclimatic influences on slope forms and processes, northeastern Arizona, J. Geophys. Res., 113, 2007JF000789, https://doi.org/10.1029/2007JF000789, 2008.
Chen, Y. X., Jiang, L. M., and Liang, L. L.: Monitoring permafrost deformation in the upstream Heihe River, Qilian Mountain by using multi-temporal Sentinel-1 InSAR dataset, Chinese J. Geophys., 62, 2441-2454, https://doi.org/10.6038/cjg2019M0255, 2019.
Deng, H., Zhang, Z., and Wu, Y.: Accelerated permafrost degradation in thermokarst landforms in Qilian Mountains from 2007 to 2020 observed by SBAS-InSAR, Ecological Indicators, 159, 111724, https://doi.org/10.1016/j.ecolind.2024.111724, 2024.
Fan, X., Li, W., Wu, X., Yao, M., Niu, F., and Lin, Z.: Heterogeneity of Surface Heat Exchange of Slopes and Potential Drivers of the Initiation of Thaw Slump, Qinghai-Tibet Plateau, Int J Disaster Risk Sci, 14, 549-565, https://doi.org/10.1007/s13753-023-00508-8, 2023.
Meng, Y., Lan, H., Li, L., Wu, Y., and Li, Q.: Characteristics of Surface Deformation Detected by X-band SAR Interferometry over Sichuan-Tibet Grid Connection Project Area, China, Remote Sensing, 7, 12265-12281, https://doi.org/10.3390/rs70912265, 2015.
Niu, F., Luo, J., Lin, Z., Fang, J., and Liu, M.: Thaw-induced slope failures and stability analyses in permafrost regions of the Qinghai-Tibet Plateau, China, Landslides, 13, 55-65, https://doi.org/10.1007/s10346-014-0545-2, 2016.
Peng, X., Yang, G., Frauenfeld, O. W., Li, X., Tian, W., Chen, G., Huang, Y., Wei, G., Luo, J., Mu, C., and Niu, F.: The first hillslope thermokarst inventory for the permafrost region of the Qilian Mountains, Earth Syst. Sci. Data, 16, 2033-2045, https://doi.org/10.5194/essd-16-2033-2024, 2024.
Sheng: Map of permafrost distribution in the Qilian Mountains, https://doi.org/10.11888/Geocry.tpdc.270456, 2020.
Wang, S. T., Sheng, Y., Wu, J. C., Li, J., and Zhang, S. F.: The characteristics and changing tendency of permafrost in the source regions of the Datong River, Qilian Mountains, Journal of Glaciology and Geocryology, 37, 27-37, https://doi.org/10.7522/j.issn.1000-0240.2015.0003, 2015.
Xia, Z., Liu, L., Mu, C., Peng, X., Zhao, Z., Huang, L., Luo, J., and Fan, C.: Widespread and Rapid Activities of Retrogressive Thaw Slumps on the Qinghai-Tibet Plateau From 2016 to 2022, Geophysical Research Letters, 51, e2024GL109616, https://doi.org/10.1029/2024GL109616, 2024.
Zhang, J., Chen, J., Li, C., Lu, W., Hao, J., Niu, P., Li, K., Ma, S., and Yuan, R.: Landslides along the Engineering Corridors in the Northeastern Margin of the Qinghai-Tibet Plateau of China: Comprehensive Inventory and Mechanism Analysis, Landslides, 21, 3049-3067, https://doi.org/10.1007/s10346-024-02341-6, 2024.
-
RC3: 'Comment on egusphere-2025-2726', Anonymous Referee #3, 01 Jan 2026
-
RC2: 'Comment on egusphere-2025-2726', Anonymous Referee #2, 05 Sep 2025
This manuscript addresses the important problem of detecting freeze–thaw landslides (FTLs) in permafrost regions, focusing on the Qilian Mountains. The authors propose a semi-automatic iterative recognition framework that integrates SBAS-InSAR surface deformation, multi-source imagery, and topographic factors. The study is of practical significance, aiming to improve the accuracy of identifying small and hidden FTLs, which are often difficult to capture using conventional methods. The methodology is clearly structured into four steps, and the results, including a reported true positive rate of 93.3%, demonstrate promising performance. Nevertheless, several aspects could be improved:
L86: Change “red points” to “red regions/polygons”.
L238: In the Results section, it is recommended to add a dedicated subsection to specifically present the InSAR deformation results of hidden and small-scale FTLs (partially), along with a comparison to non-FTL areas.
L242–243, Fig. 3: Please add a distribution map of permafrost and seasonally frozen ground as the basemap for Figure 3.
L299: There is a spelling error that needs correction.
Overall, this paper represents a valuable attempt to advance the identification of FTLs and has potential for publication after minor revision.
Citation: https://doi.org/10.5194/egusphere-2025-2726-RC2 -
AC2: 'Reply on RC2', gang wei, 10 Jan 2026
We thank the reviewer for this constructive and insightful comment.
1. This manuscript addresses the important problem of detecting freeze-thaw landslides (FTLs) in permafrost regions, focusing on the Qilian Mountains. The authors propose a semi-automatic iterative recognition framework that integrates SBAS-InSAR surface deformation, multi-source imagery, and topographic factors. The study is of practical significance, aiming to improve the accuracy of identifying small and hidden FTLs, which are often difficult to capture using conventional methods. The methodology is clearly structured into four steps, and the results, including a reported true positive rate of 93.3%, demonstrate promising performance. Nevertheless, several aspects could be improved: 1. L86: Change “red points” to “red regions/polygons”.
Response: We have revised the expression as suggested. The term “red points” has been changed to “red regions” throughout the manuscript to more accurately reflect the spatial nature of the detected FTLs.
2. L238: In the Results section, it is recommended to add a dedicated subsection to specifically present the InSAR deformation results of hidden and small-scale FTLs (partially), along with a comparison to non-FTL areas.
Response: We thank the reviewer for this valuable suggestion. In direct response, we have revised the existing Section 4.3 "Surface Deformation of Freeze-Thaw Landslides" by appending a detailed paragraph that provides a concrete example of a small and hidden FTL, explicitly comparing its deformation characteristics to the surrounding stable, non-FTL areas. The specific addition to the end of Section 4.3 includes a representative case study of a specific FTL with an area of 42,450 m2, serving as a typical example of small and hidden landslides in the region. The concealment metric (ΔNDVI) values for this landslide are predominantly clustered within the narrow range of -0.1 to 0.1, quantitatively confirming its "hidden" nature due to minimal vegetation disturbance (see new Fig. S8a). A maximum annual deformation rate of 27.4 mm·a⁻¹ is found at the landslide head, which starkly contrasts with the deformation values of the adjacent stable terrain, providing a clear and quantifiable comparison as requested (see new Fig. S8b). The morphological correlation is such that the area of maximum deformation correlates with developing tension cracks visible in high-resolution imagery, linking the geophysical measurement to physical ground truth (see new Fig. S8c). This was added to our manuscript on lines 294-300: To exemplify the deformation characteristics of a typical small and hidden FTL, we analyzed a representative case with an area of 42,450 m². This feature exhibits diagnostic characteristics of such landslides: its ΔNDVI values are predominantly clustered within the narrow range of -0.1 to 0.1, indicating minimal vegetation disturbance and high concealment (Fig. S8a). Interferometric analysis reveals a maximum annual deformation rate of 27.4 mm·a⁻¹ at the landslide head (Fig. S8b), where developing tension cracks are clearly visible in high-resolution imagery (Fig. S8c). Most notably, the deformation pattern of this FTL demonstrates a stark contrast with the surrounding stable, non-FTL terrain, highlighting the capability of InSAR monitoring to detect these subtle but dynamically active geomorphological processes.
3. L242-243, Fig. 3: Please add a distribution map of permafrost and seasonally frozen ground as the basemap for Figure 3.
Responses: We thank the reviewer for this constructive suggestion. In response, we have updated both Figure S1 (Study Area Location) and Figure S3 (Spatial Distribution of Identified FTLs) to include a basemap delineating the distribution of continuous permafrost, discontinuous permafrost, and seasonally frozen ground within the study area. This classification is based on the most recent regional permafrost map of the Qilian Mountains (Sheng, 2020). The addition of this cryospheric context provides a clearer understanding of the relationship between FTL occurrence and the underlying permafrost environment.
4. L299: There is a spelling error that needs correction.
Responses: We have carefully reviewed Line 299 and corrected the spelling error (“the” was misspelled). The text now reads accurately.
Line325: For instance, the method relies heavily on the temporal coverage, data quality, and resolution of InSAR. Some FTLs occurred
References:
Sheng: Map of permafrost distribution in the Qilian Mountains, https://doi.org/10.11888/Geocry.tpdc.270456, 2020.
-
AC2: 'Reply on RC2', gang wei, 10 Jan 2026
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 840 | 102 | 33 | 975 | 26 | 31 |
- HTML: 840
- PDF: 102
- XML: 33
- Total: 975
- BibTeX: 26
- EndNote: 31
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General comment
This manuscript presents a novel semi-automatic iterative method for identifying freeze-thaw landslides (FTLs) in permafrost regions, particularly focusing on the Qilian Mountains. The method uses derivative data extracted from InSAR data and other multi-modal data, including ERA5, DEM, and optical RS imagery. The validation data set has 167 FTLs and 17 RTSs. They found 98 new FTLs in the study region. Essentially, this is a threshold-based, manual and iterative screening process to discover new FTLs. Although the method has novelty and succeeded in the study region, it remains questionable if this method could be easily adopted to other regions due to highly heterogeneous permafrost environments.
Specific comment
The biggest concern of this method is the ability to generalise to other regions. As the authors has claimed in the Limitation section, the thresholding values for Elevation, Slope, Aspect and Deformation rate are analysed and determined manually for each iteration with subjectivity, which means for almost every environmentally-distinct region, the manual tuning of these thresholds needs to be done again, and even in different years in the same region, these values won’t be guaranteed to be stable. This method worked in a tiny region, as it has been tested in the manuscript; however, compared to the entire Tibetan permafrost or Arctic permafrost, the region is way too small. It will be very questionable to apply this method to a vast region without extensive manual analysis of the values and thresholds.
Secondly, this iterative, threshold-based method is very similar to the process of a decision tree-based machine learning model. One could train an ML model using the same variables (Elevation, slope, aspect, deformation rate) with some FTL ground truths. The model will have the exact same input-output as your methods. The downside of the ML method is that it requires at least a few hundred/thousand ground truth FTLs for training, but the resulting ML model will have much better generalisability than manual thresholding. I would strongly recommend comparing your proposed method with a properly trained ML model on a significantly larger area to see the difference in performance and geospatial extrapolation ability.
Another concern is the iterative nature of this method, and manual verification makes it very labour-intensive. It requires further assessment to understand the cost-effectiveness of this method.
Line140: 2x7(unit?) window
Line 169 specify how you ‘statistically analyzing’ the factors, and how you determine the thresholds
Eq(1) and Eq (2) unit of the variables are not specified, and which of these are dimensionless quantities?