the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 18 Sep 2026
Initiation Timing Improves Retrogressive Thaw Slump Susceptibility Mapping on the Qinghai–Tibet Plateau
Abstract. Accelerating retrogressive thaw slump (RTS) activity is reshaping permafrost landscapes on the Qinghai-Tibet Plateau (QTP) and threatening infrastructure. Forward-looking assessment requires identifying slopes approaching failure rather than mapping past RTS occurrence. Conventional susceptibility models treat RTS initiation as a binary outcome, making pre-initiation locations effectively equivalent regardless of when they subsequently fail. Yet under progressive environmental forcing, initiation timing can provide an implicit ordering of susceptibility: earlier initiation may indicate greater susceptibility than later initiation. We therefore defined Years Before Initiation (YBI) from annual inventories and developed a weakly supervised ranking framework that uses initiation-time ordering to learn susceptibility gradients. The Annual-YBI model was compared with two conventional models (Static Binary and Annual Binary) sharing the same neural network architecture. Although all models performed similarly in five-fold cross-validation (ROC-AUC ≈ 0.988), the Annual-YBI model captured 89.45 % of 199 temporally held-out RTS initiations in 2021–2022 within the top 10 % susceptibility area, compared with 79.90 % and 63.82 % for Annual Binary and Static Binary models, respectively. SHAP analysis identified NDVI, slope, and ground ice content as primary predisposing factors, with modulation by annual hydroclimatic variables. The 2021–2022 maximum susceptibility composite classified 1.70 × 105 km2 (15.6 % of QTP permafrost) as high or very high susceptibility, including 42.8 % of permafrost within the 20-km Qinghai–Tibet Engineering Corridor buffer. These results show that initiation timing provides weak supervision for susceptibility gradients that binary labels cannot resolve, improving identification of locations prone to future RTS initiation and offering potential for transfer to other permafrost regions.
- Preprint
(5301 KB) - Metadata XML
-
Supplement
(7611 KB) - BibTeX
- EndNote
Status: open (until 30 Oct 2026)
- RC1: 'Comment on egusphere-2026-5035', Anonymous Referee #1, 16 Oct 2026 reply
-
RC2: 'Comment on egusphere-2026-5035', Anonymous Referee #2, 19 Oct 2026
reply
This manuscript presents an innovative framework for mapping retrogressive thaw slump (RTS) susceptibility across the Qinghai–Tibet Plateau. Its main contribution is the use of relative initiation timing, expressed as Years Before Initiation (YBI), as a weak supervision signal through an auxiliary pairwise LambdaRank loss. This framework extends conventional binary classification by incorporating temporal information about pre-initiation conditions. The results demonstrate improved identification of subsequent RTS initiation locations in the temporally held-out evaluation while maintaining comparable cross-validation performance. The approach also has potential relevance to broader landslide susceptibility applications where suitable multiyear inventories are available.
The manuscript is well written and logically organized. The comparisons with conventional binary models, together with the temporally held-out evaluation, provide valuable support for the proposed framework. I recommend minor revision, primarily to clarify the interpretation of the temporal-ordering assumption, the treatment of uncertainty in initiation timing, and several methodological details.
General comments
1) Physical interpretation of initiation time ordering
The central premise is that, within the same observation year and a locally comparable spatial domain, slopes on which RTSs initiate earlier tend to have higher relative susceptibility than those on which RTSs initiate later. The authors appropriately acknowledge in Sections 5.1 and 5.4 that this relationship is probabilistic rather than deterministic. Nevertheless, its physical justification could be presented more explicitly in the methodological description.
RTS initiation depends on both predisposing conditions, such as slope and ground ice content, and the timing and intensity of triggering events, such as extreme precipitation, anomalous thawing, or local disturbances. A highly susceptible slope may remain stable until a sufficiently strong trigger occurs, whereas a less susceptible slope may fail earlier because it experiences a stronger local disturbance. Restricting comparisons to the same year and DBSCAN cluster improves comparability but does not necessarily ensure comparable triggering conditions. Please explain more explicitly the conditions under which initiation order is expected to provide a meaningful proxy for relative susceptibility, and distinguish this conditional interpretation from a general ordering of intrinsic slope susceptibility.2) First-detection uncertainty and robustness of the ranking constraints
The inventory uses the year of first detection as a proxy for initiation year. Although this limitation is acknowledged in Section 5.4, please clarify how errors or delays in first-detection dates affect the YBI ranking pairs and model robustness. A simple sensitivity analysis could help quantify these effects.
The 2016 cohort warrants particular attention because 2016 is the first year covered by the inventory. RTSs already visible in 2016 may have initiated before the observation period and therefore cannot necessarily be interpreted as new initiations in that year. Please clarify how these baseline features were treated when assigning initiation-year labels and constructing ranking constraints.3) The use of YBI provides an interesting way to incorporate temporal information. However, time-to-event approaches, including survival models, also provide established frameworks for using event timing and accounting for locations that have not failed during the observation period. A brief conceptual comparison explaining why pairwise ranking was selected over other models for the present study would strengthen the rationale for this methodological choice.
4) Pseudo-absence sampling
Pseudo-absence samples were selected at least 1,000 m from mapped RTS polygons. Given that permafrost terrain on the QTP often shares similar background characteristics over short distances, please clarify whether any additional topographic or land-cover criteria were applied, or whether sampling was otherwise unrestricted throughout the mapped permafrost domain outside this buffer.
The exclusion buffer reduces proximity to known RTSs but does not by itself ensure that the background samples represent environmentally comparable slopes. For example, extensive sampling from very gentle terrain or other settings with limited potential for RTS development could make occurrence–background discrimination relatively easy. Please provide the rationale for the sampling strategy and briefly discuss how the environmental composition of the pseudo-absence pool may influence classification performance and the learned susceptibility gradient.Specific comments
1) Section 2.3: Please specify the months represented by “summer” and “warm season” and indicate whether these definitions are consistent across predictors.
2) Section 2.4: ERA5-Land predictors have a stated spatial resolution of approximately 11.1 km. Please specify the interpolation or resampling method used to resample these data to the 100-m modeling grid. The corresponding method for SMAP soil moisture should also be reported.
3) Line 190: Replace “earlies” with “earliest.”
4) Line 230: Replace “performance differences among the three models reflects” with “performance differences among the three models reflect.”
5) Line 350: Replace “validation points was summarized” with “validation points were summarized.”Citation: https://doi.org/10.5194/egusphere-2026-5035-RC2
Model code and software
nanzt/YBIRTS-Susceptibility: v1.0.0 Pengfei Li and Zhuotong Nan https://doi.org/10.5281/zenodo.20603769
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 430 | 129 | 59 | 618 | 88 | 81 | 76 |
- HTML: 430
- PDF: 129
- XML: 59
- Total: 618
- Supplement: 88
- BibTeX: 81
- EndNote: 76
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The use of initiation timing to inform RTS susceptibility is interesting, and the annual inventory compiled by Xia et al. (2024) provides a valuable basis for this work. My main concern is the methodological justification. The manuscript attributes limitations of inventory aggregation and sample construction too broadly to binary classification, while the experiments do not yet isolate the contribution of temporal ordering. I recommend major revision.