the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 Mar 2026
A 2020 permafrost distribution map over the Qinghai-Tibet Plateau
Abstract. Permafrost on the Qinghai-Tibet Plateau (QTP) is undergoing rapid degradation, yet most existing distribution maps reflect long-term historical averages, failing to capture the current thermal state required for accurate ecological and hydrological modeling. To address this temporal mismatch, this study presents a 1-km QTP permafrost distribution map for the year 2020. We employed an extended ground surface frost number model (FROSTNUM) driven by satellite-derived freezing/thawing indices. To overcome the lack of concurrent field surveys for parameter calibration, we implemented a space-for-time substitution strategy, utilizing a Random Forest regression to robustly estimate the empirical soil parameter (E) based on environmental covariates. The resulting map reveals that in 2020, permafrost covered approximately 1.038 × 10⁶ km² (39.35 % of the plateau), while seasonally frozen ground (SFG) covered 1.466 × 10⁶ km² (55.57 %). Compared to the 2010 baseline, the permafrost area declined by 4.8×104 km2 (a 1.82 % decrease). Spatially, the degradation of permafrost to SFG extensively occurred in the central QTP (accounting for 7.41 % of the total change), and a significant marginal contraction of SFG to non-frozen ground in the southern margin (accounting for 39.62 % of the total change). Validations against 109 independent borehole records from the 2020 period confirms the map’s reliability, achieving an overall accuracy of 0.84 and a Kappa of 0.58. This 2020 map provides an essential, up-to-date resource for quantifying the recent cryospheric shifts and supporting engineering risk assessments in this climate-sensitive region.
- Preprint
(6771 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 22 May 2026)
- RC1: 'Comment on egusphere-2026-345', Anonymous Referee #1, 09 Apr 2026 reply
-
CC1: 'Comment on egusphere-2026-345', Mamoru Ishikawa, 22 Apr 2026
reply
General comments
I have reviewed EGUSPHERE-2026-345 by Chen et al. This paper has mapped permafrost distribution in 2020 over Tibetan plateau, using new strategy of a space-for-time substitution, using neural network and machine learning approaches. The map and estimated permafrost change since 2010 well explained ecohydrological changes in high resolution. It also explained well the limitation of applying neural network scheme for permafrost mapping. Accordingly I would be very positive to publish this in TC after following minor revisions.
Specific comments
174-175: any reference for this threshold?
181-184: As for hypothesis of E determination, soil moisture would be changed over the decadal scale, and statistical relation between E and environmental drivers might alter.
184: a space-for-time substitution strategy, needs for more detail explanation. It is hard to understand only from Fig.2(c).
209: Any other environmental factors for multilinear regression? Any correlation between NDVI and latitude?
322: Is usage of ther term 'risk' appropriate?
Fig.5: Can you say that these boxplots differ between Negative and Positive? Any statistics?
334-335: How we see Fig 5 to confirm this description?
Table 2: Include citations for MAGT and TTOP-based maps. Can you show borehole location in any of the figures?
Section 4.6: better to move to Discussions
333-340: Better to move to Discussions
Citation: https://doi.org/10.5194/egusphere-2026-345-CC1 -
RC2: 'Comment on egusphere-2026-345', Anonymous Referee #2, 27 Apr 2026
reply
Chen et al. present a 1x1 km map of permafrost conditions on the Qinghai–Tibet Plateau for the 2020 period, along with a novel methodology for deriving said map without access to extensive field data for this period. The authors also perform extensive validation of their methodology and output against available observational data and permafrost maps. I applaud Chen et al. for including the MLP model despite its deficiencies, which can help guide similar applications in the future. Overall, my judgement is that this study is well suited for publication in The Cryosphere.
Specific comments:
- How well does the presented methodology for estimating the soil parameter E perform for the 2010 period? If I understand correctly, the authors compare estimated E for 2020 period to the E determined from the field surveys in 2010. It would be interesting to see how well the two methods are in estimating E for 2010, and how the estimated E has changed from 2010 period to the 2020 period.
- The authors consistently refer to the permafrost distribution being for 2020, while in they are using data aggregated over 5 years to predict this distribution. Moreover, regions are classified as seasonally frozen ground if conditions are unfavourable for permafrost during this temporal period, while it can take substantially longer for deep permafrost to completely disappear. I would thus encourage a more precise language regarding the time of validity of the map, and how the classifications relate to the transient thermal state of the ground.
- I would welcome a discussion on the transferability of the presented mapping approach to other climates/environments. It would be interesting to hear the authors thoughts on which aspects of the method, including assumptions, are generalizable, and which are specific to the context of the Qinghai–Tibet Plateau
Technical comments:
Line 56-62: This paragraph warrants some reference(s), e.g. for the mentioned “previous mapping efforts” or for the applications requiring “year-specific benchmarks”
Line 97: please use “per decade” rather than “/10a”
Line 162-163: This does not fit to the subsection “Existing permafrost datasets for comparison”, please move or revise header.
Section 4.1: It would be interesting to know if there is also a temporal trend in DDT_{LST}, or if the presented cooling pattern is only present in the corrected DDT_{GST} values.
Line 294: Please indicate on this or the overview map where the “southern Himalayan margin” is.
Line 306: Do you mean “temporally extrapolate”?
Figure 4: It is not clear to me if the “E value in 2010” refers to the values from Cao et al. (2023). Please be specific here and elsewhere in the manuscript whether the “E values” and “Permafrost maps” from 2010 refer to the original Cao et al. publication or if they refer to your estimated E values and the thereof calculated permafrost maps.
Line 366: It’s unclear to me where the “Three-River Headwater Region” is. Please include the placenames used in the text in the relevant figure of in the overview map in Figure 1.
Line 407: You state that you identify “four out of five confirmed SFG boreholes”, but there appears to be eight purple “borehole with seasonal frost” markers in Figure 8c.
Figure 8 & 9: I find it hard to distinguish the borehole markers, please consider revising the symbols for readability.
Line 446: it is unclear where the classifications “continuous and discontinuous permafrost” come from, and if they are relevant for this context.
Line 466: Where do the datasets of thermokarst lakes and engineering instabilities come from? Remote sensing, field mapping?
Line 492-502 and elsewhere: I’m surprised that so much attention is given to the transition of seasonally frozen ground to non-frozen ground. While this is an interesting transition, I think this can be deemphasized as the manuscript is about permafrost mapping.
Thank you for an interesting read!
Robin B. Zweigel
Citation: https://doi.org/10.5194/egusphere-2026-345-RC2 -
EC1: 'Comment on egusphere-2026-345', Anne Morgenstern, 29 Apr 2026
reply
Please note that CC1 should be considered equivalent to a full review. The authors are expected to address this comment with the same level of detail as the referee reports in their reply.
Citation: https://doi.org/10.5194/egusphere-2026-345-EC1
Data sets
Dataset associated with "A 2020 permafrost distribution map over the Qinghai-Tibet Plateau" submitted to The Cryosphere Yuhong Chen, Zhuotong Nan, Wenbiao Tian, Yi Zhao, Shuping Zhao, Dongkai Yang, Guifei Jing, and Fujun Niu https://doi.org/10.6084/m9.figshare.30997375
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 650 | 308 | 51 | 1,009 | 36 | 67 |
- HTML: 650
- PDF: 308
- XML: 51
- Total: 1,009
- BibTeX: 36
- EndNote: 67
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors present in this manuscript a new high-resolution permafrost distribution map for the Qinghai–Tibet Plateau for the 2020 period. The main novelty lies in the gridded estimation of the empirical soil parameter E under the extended FROSTNUM framework, using a space-for-time substitution strategy in the absence of concurrent large-scale field surveys. Overall, the results of this study have good practical value and provide a useful reference for this community.
Main points:
(1) The final results lack an explicit uncertainty analysis. At present, the manuscript selects a single optimal scheme from multiple methods and configurations to generate the final map, but does not further quantify the uncertainty of the final results. I recommend that the authors include an assessment of the robustness of both the predicted E parameter and the resulting permafrost distribution.
(2) The methodological implications of the space-for-time substitution strategy are not discussed in sufficient depth. In particular, the manuscript would benefit from a clearer discussion of where this strategy is likely to be most reliable, where it may break down, and how this affects the interpretation of the final map.
Minor points:
(1) P7 L172–175. A brief explanation of why F>0.5 is used as the threshold for permafrost classification would improve the clarity of the method description.
(2) P2 L63–68; P4 L110–115. The manuscript refers to the map as being “for 2020” or an “instantaneous snapshot”, while the forcing data actually cover 2016–2020. A more cautious and consistent expression, such as “for the 2020 period,” is recommended.
(3) P12, L308–315. For the discussion of the MLP failure, it would be better to avoid attributing the issue simply to the “black box” nature of deep learning, and instead refer more specifically to the lack of physical constraints and poor generalization.