the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 24 Aug 2026
A first 90 m resolution active layer moisture dataset across the Qinghai–Tibet Plateau permafrost region
Abstract. Active layer moisture (ALM) plays a fundamental role in regulating the hydrothermal dynamics, freeze–thaw processes, carbon cycling, and ecosystem and load-bearing functions of permafrost. However, spatially continuous, high-resolution datasets that characterize soil moisture across the entire active layer remain largely unavailable for the Qinghai–Tibet Plateau (QTP), which hosts the largest low- to mid-latitude permafrost region globally. Here, we present the first 90 m resolution, spatially complete dataset of ALM – the depth-averaged volumetric water content of the entire active layer – for the entire permafrost region of the QTP. The dataset was produced by integrating 342 in situ samples collected during peak-thaw seasons from 2009 to 2024 with multi-source remote sensing environmental predictors, topographic factors, and soil physicochemical properties. Four ensemble machine learning models, include Random Forest, Extra Trees, XGBoost, and CatBoost, were trained, and a bias-aware multi-model fusion strategy was applied to generate the final ALM product. SHAP-based recursive feature elimination was used to identify the dominant environmental controls and optimize feature selection. Random five-fold cross-validation showed high apparent accuracy (R² =0.62–0.63, RMSE ≈ 0.08 m³ m⁻³), and a group-based spatial cross-validation provided a conservative transferability estimate (pooled R² = 0.30–0.38). Within the permafrost region, ALM ranges from 0.02 to 0.70 m³ m⁻³, with a mean of 0.21 m³ m⁻³ and a standard deviation of 0.09 m³ m⁻³, exhibiting a distinct southeast-to-northwest decreasing gradient. The dataset is freely available with DOI: https://doi.org/10.12072/ncdc.permafrost.db7312.2026. The dataset is accompanied by an area-of-applicability (AOA) mask delineating the region – approximately 60 % of the permafrost region – where the cross-validated performance can be expected to hold, together with per-pixel uncertainty estimates derived from the DI–RMSE relationship and the divergence among the four models.
- Preprint
(12378 KB) - Metadata XML
-
Supplement
(9128 KB) - BibTeX
- EndNote
Status: open (until 28 Oct 2026)
-
CC1: 'Comment on essd-2026-330', Dongliang Luo, 07 Sep 2026
reply
-
AC1: 'Reply on CC1', Erji Du, 09 Oct 2026
reply
Dear Editor and Professor Luo,
We thank the referee for the careful reading of our manuscript and for the seven constructive comments, which all concern the physical interpretation and the usability of the dataset rather than the underlying analyses. We are especially grateful that you chose to review our manuscript under your public identity; we greatly appreciate this open and transparent form of scientific exchange. We have addressed every comment in the revised manuscript, which is attached to this response with all textual changes shown as tracked changes; the two figure-related comments have been implemented in the revised figure files, which are also provided in the attachment. Point-by-point responses are given below.
Specific Comments 1
The feature importance analysis reveals that surface-sensitive remote sensing indices contribute 58%–69% of the predictive weight across models. While this reflects strong surface–subsurface hydrothermal coupling on the plateau, the authors should briefly acknowledge scenarios where decoupling occurs—such as deep active layers (>2–3 m) with coarse surface layers overlying perched water tables (supra-permafrost groundwater) or thick organic mats, where surface signals might have weaker sensitivity to bottom-layer moisture dynamics.
Response: We thank the referee for this perceptive comment on the interpretation of the feature-importance results. To begin with, the 58–69% cumulative contribution refers to the share of the surface-sensitive remote sensing indices relative to the other predictor groups (terrain, soil physicochemical and climate); within each model the contributions of all retained predictors sum to 100%. Our field experience on the plateau indicates that ALM is closely linked to the surface moisture state and to soil bulk density: wetter surface conditions favour higher ALM, and a lower bulk density—to which thick organic mats contribute substantially—implies higher porosity and hence a larger water-storage capacity. We agree that, for very deep active layers, the moisture of the lower part of the profile is increasingly controlled by supra-permafrost groundwater; for the depth-averaged moisture of the thawed column, however, the surface moisture regime remains a major controlling factor. We attribute the high cumulative contribution of the remote sensing indices to two circumstances: (i) these indices carry genuine information on ALM, and (ii) the bulk-density layer (BD_60–100) that we expected to enter the models was not retained in any of the four final subsets, and only the 0–5 cm organic-matter content (OC_0–5; Fig. 4) was selected in the CatBoost, Extra Trees and XGBoost models—an indication that the texture-related attributes of the CSDLv2 dataset do not yet resolve the true spatial variability of active-layer soil texture at the 90 m grid used here. These two circumstances explain the 58–69% share and, in our view, also underlie the moderate accuracy of the simulated ALM, especially under spatial cross-validation (pooled R² = 0.368; Table 3). Further gains will require more spatially distributed sampling points and high-resolution data characterising the soil texture within the active layer. Following the referee’s suggestion, the Discussion (Sect. 4.1) now acknowledges the situations in which the surface–subsurface coupling weakens—very deep active layers, coarse-textured surface horizons above shallow supra-permafrost water and thick organic mats—and states how the predictions should be interpreted in such areas: the product is most reliable where the surface state and active-layer moisture co-vary strongly and least reliable where the two are decoupled, while the DI–AOA framework flags pixels that are environmentally novel relative to the calibration data, although decoupling within the calibration domain itself cannot be detected by the current predictors.
Specific Comments 2
Given that in situ samples were pooled from 2009 to 2024 while remote sensing predictors were compiled as 2014–2024 composites, the resulting product is explicitly a long-term climatological representation of the annual maximum thaw state (September–October). Ensure this distinction is consistently highlighted in the Abstract and User Guidance to avoid users misinterpreting the product as a time-slice snapshot of any single year.
Response: We fully agree that the pooled temporal design must be flagged wherever users encounter the product. The Abstract now states explicitly that, as a multi-year composite of peak-thaw-season observations, the dataset represents the long-term climatological state of ALM rather than the moisture of any specific year. The same clarification is given in the Data Availability section (Sect. 5), which is the first place users of the data will consult, and Sect. 2.1 adds a caveat that pooling inevitably suppresses year-specific departures; the per-year sample counts (56 samples in 2009–2013 versus 286 in 2014–2024) are provided in Supplementary Table S3. The manuscript does not contain a separate “User Guidance” section; this role is fulfilled by the Data Availability section and by the readme documentation accompanying the released dataset, and the climatological statement will be kept in line with the paper in the final release, so that the distinction is highlighted consistently.
Specific Comments 3
Because the four-model intersection AOA covers ~60% of the permafrost domain, approximately 40% of the region represents an environmental extrapolation. The authors have handled this commendably by publishing the companion AOA and expected-RMSE layers. A sentence should be added to the Data Availability section reminding users to apply the AOA mask when reliable empirical bounds are required.
Response: We appreciate the referee’s recognition of the AOA and expected-RMSE companion layers. Following the suggestion, the Data Availability section (Sect. 5) now contains an explicit reminder, set in bold in the revised manuscript, that users should mask the ALM product to the fused AOA (provided as a companion layer in the same archive) before any pixel-level interpretation or model ingestion, and that pixels outside the AOA represent extrapolation with substantially higher expected error and should not be read as point estimates. We have further reported the outside-AOA fraction by watershed in Supplementary Table S3; as the referee anticipates, most of the uncovered area lies in the interior of the Qiangtang Plateau.
Specific Comments 4
Line 254: Remove the placeholder “...ChinaMet 0.01° (approximately 1 km) gridded dataset of annual Penman-Monteith PET (petPM; citation to be added) for 2009-2024” and supply the missing citation.
Response: The placeholder has been removed and the missing citation supplied. Sect. 2.8 now cites the ChinaMet dataset in full—Hu, Y., Zhang, L., Zhao, Y., and Che, T.: ChinaMet: A multisource integrated high-resolution multi-variable meteorological dataset for China, Zenodo [dataset], https://doi.org/10.5281/zenodo.14580706, 2024—which has also been added to the reference list (Hu et al., 2024). We fully agree that the bilinear resampling of the 1 km PET field to the 90 m grid only aligns the grids and adds no information at the 90 m scale; Sect. 2.8 now states this explicitly and clarifies that pet_mean mainly controls the overall spatial distribution pattern of ALM across the study region at the regional scale, rather than representing a temporal trend.
Specific Comments 5
Line 322: Correct the typo in sample size: “G4_Peripheral (Tarim + Hexi + Salween R., n=2?)”.
Response: We thank the referee for the careful check. The sample size of the G4_Peripheral fold is n = 23 (Tarim + Hexi + Salween R.), as stated in Sect. 3.2; we have re-verified this value against the archived sample dataset and confirmed that it is consistent with the per-fold statistics reported in Table 3 and Supplementary Table S3.
Specific Comments 6
Figure 11: The axes for both panels currently display raw grid row/column counts. Please replace these with projected coordinates (km) or geographic coordinates (°E/°N) to match the other maps in the paper.
Response: Figure 11 has been revised: the axes of both panels now display geographic coordinates (°E/°N), consistent with the other maps in the paper.
Specific Comments 7
Figures 1 & 9: Ensure consistent typography for river and lake labels across maps (e.g., standardizing spaces in “Bangong Co”, “Yibug Caka”, and checking “Brhmaputra R.” for spelling).
Response: The river and lake labels in Figures 1 and 9 have been harmonised in the figure files accompanying the revised submission: the spacing of the lake names (e.g., “Bangong Co”, “Yibug Caka”) has been standardised and the spelling of “Brhmaputra R.” has been corrected to “Brahmaputra R.”. We have also checked the corresponding toponyms in the text and in Table 4 and unified their spacing, so that “Yibug Caka”, “Bangong Co”, “Zhari Nam Co” and “Dogai Coring L.” are now written consistently throughout the manuscript.
In summary, all seven comments have been addressed: five by textual revisions that are fully marked in the tracked-changes version of the revised manuscript, and two (Comments 6 and 7) by the updated figure files. None of the changes alters the data products, the numerical results or the core conclusions of the paper. We believe the revised manuscript is clearer and easier to use correctly as a result, and we hope it is now suitable for publication in Earth System Science Data. We thank the referee once again for the time and expertise invested in this review.
Sincerely,
Erji Du, Tonghua Wu, Liyun Dai, Youhua Ran, and co-authors
on behalf of all authors of essd-2026-330
-
AC1: 'Reply on CC1', Erji Du, 09 Oct 2026
reply
-
RC1: 'Comment on essd-2026-330', Anonymous Referee #1, 05 Oct 2026
reply
General comment
A 90 m, depth-averaged active-layer moisture field for the entire TP permafrost region fills a real need. It also comes with an honest uncertainty story, combining the AOA with per-pixel layers. The authors openly flag the LST artefact rather than hide it. The field dataset, the multi-ensemble framework, the spatial cross-validation and the open data and code practice are all solid. The comments below are mostly about making the product easier to use correctly and harder to misinterpret. None of them requires new analysis.
One scientific concern unifies the specific comments below. The transferable accuracy of the product is modest and strongly conditioned. The manuscript should make those limits as visible as its strengths. Pixel-level spatial cross-validation (R² ≈ 0.37) tells a far more cautious story than the random-CV and bin-aggregated metrics foregrounded in the Abstract. That skill also rests on well-defined conditions. Only ~60% of the domain lies inside the area of applicability. The training set mixes pit protocols and GPR averages, and these carry method-specific residuals. The 2009–2024 pooling represents a September–October peak-thaw climatology. The predictions also lean on surface-sensitive indices, whose surface–subsurface coupling weakens in some settings. None of this undermines the contribution. Together these factors define the envelope within which the product should be used. Stating that envelope clearly is what will make the dataset most valuable to the modelling community.
Specific comment 1
You already note that ST_B10 pixels along the ~95°, 97° and 99°E swaths were filled with 0 instead of no-data. This kind of transparency is exactly what reviewers should reward. Please take it one step further. Give the actual affected pixel fraction (even a rough percentage) and confirm that these pixels all fall outside the fused AOA mask. Users who ingest this dataset into models need that number. Stating it removes any doubt about product completeness.
Specific comment 2
Since the fused AOA covers only ~60% of the domain, the remaining ~40% is extrapolation with higher expected error. In my experience, users too often treat every pixel of a full raster as equally reliable. This product will also be picked up by model users who are not permafrost specialists. Please state the following prominently, in bold, in both the data readme and the Data Availability section. Outside-AOA pixels should not be read as point estimates, and the AOA mask is provided precisely for this purpose. It would also help to report the outside-AOA fraction by watershed in the supplement, since much of the uncovered area falls in the Qiangtang interior.
Specific comment 3
The Abstract and Conclusions report the random-CV R² (0.62–0.63) and the bin-aggregated slope of 0.99. They do not report the pixel-level spatial-CV performance of the fused product itself (Table 3: R² = 0.368, RMSE = 0.104, MAE = 0.081, bias = −0.015 m³ m⁻³). These two sets of numbers tell very different stories about transferable accuracy. A reader relying on the Abstract alone would clearly overestimate the pixel-level skill of the delivered map. Please report the pixel-level metrics alongside the random-CV metrics. Please also clarify the relation between the bin-aggregated slope and the pixel-level statistics, and include a per-fold breakdown. G1_East (n = 158) is the fold users will worry about. It deserves to be visible rather than hidden in the table.
Specific comment 4
Three things stand out. First, the 0.2 and 0.4 m³ m⁻³ thresholds are empirical values fitted to the spatial-CV residuals of this dataset. A sentence acknowledging their limited transferability to other regions, periods or training configurations would therefore be appropriate. Second, the thresholds and their evaluation come from the same spatial-CV residuals. This makes the comparison of fusion rules circular with respect to threshold choice. Re-running the rules on an independent residual set (e.g., random-CV or LOBO) would settle this. Third, the bias-aware rule is marginally below the simple mean at the pixel level (R² = 0.383 vs 0.368). Users therefore need a short practical note on when to use which layer. One more numerical detail: the scheme creates a visible dip in the probability density around 0.18–0.20 m³ m⁻³. Quantifying it with a finer-grained histogram would pre-empt questions later. So would a comment on what the dip means for the land-surface permafrost models that consume the product.
Specific comment 5
The 123 "test pits" turn out to follow at least three different protocols: ring-sampled pits, drill cores and manual auger samples. Each has its own vertical sampling scheme and depth-averaging convention. All are mixed with a dielectric-weighted GPR depth average in the same training set. The method-specific residuals differ systematically. From the archived sample dataset, the GPR samples are essentially unbiased. The manual auger samples are severely underestimated, and their observed mean exceeds 0.6 m³ m⁻³. Please describe the composition and depth-averaging convention of each protocol. Please also put the per-method VWC distributions and residual statistics in the supplement. This would save the next user of the sample dataset a great deal of detective work.
Specific comment 6
Pooling 2009–2024 samples into a climatology is defensible. However, precipitation varies strongly from year to year in the endorheic basins of the central Qiangtang Plateau. About 17% of the samples (56 of 342) predate the 2014–2024 predictor window. There are also no samples at all in 2016 and 2021. A short discussion of the inter-annual noise and the sample-year distribution would help. So would a note on early- versus late-period residuals in the supplement. Together these would make the climatological interpretation far more robust. Please also state explicitly in the Abstract and Section 5 that the product represents climatological peak-thaw conditions (September–October). It is not intended for seasonal or inter-annual interpretation.
Specific comment 7
The GPR+CRIM calibration against the pits is excellent (RMSE ≈ 0.032 m³ m⁻³). However, microwave sensitivity drops below roughly 1–2 m, while QTP active layers can exceed 3 m. A short caveat would pre-empt over-interpretation. It should note how representative the column average is in thick-ALT areas, and whether calibration error is expected to grow with depth. A related point concerns use. The product is a depth average, whereas land-surface models output layered moisture. Computing active-layer water storage also requires an external active-layer-thickness (ALT) dataset. A sentence or two on these comparison challenges in Sect. 4.5 or 5 would help that community use the dataset correctly.
Specific comment 8
The feature-importance analysis shows that surface-sensitive remote sensing indices dominate the predictive weight (58–69%). This presumably reflects strong surface–subsurface hydrothermal coupling on the plateau. I would appreciate a brief acknowledgement of the situations in which that coupling weakens. These include very deep active layers, coarse-textured surface horizons above shallow supra-permafrost water, or thick organic mats. A note on how predictions should be interpreted in such areas would also help.
Specific comment 9
Sect. 2.8 still contains the placeholder "(citation to be added)" for the ChinaMet petPM dataset. The complete citation must be supplied. The 1 km field is also bilinearly resampled to the 90 m grid. It is worth stating explicitly that this resampling only aligns the grids and adds no information at the 90 m scale.
Specific comment 10
A few small points. The Abstract says "≈60%" AOA coverage, while Table 5 and Fig. 11 give 59.9%. Please use one figure throughout. Unify the terminology for the target variable (ALM vs VWC; "showed as VWC" in the Figure 2 note should read "shown as VWC"). Fix the spacing typos ("slope of0.99", "The0.20–0.33 drop") and change "include" to "including" in the Abstract. Both panels of Figure 11 are annotated with bare raster indices. Please re-label them with map-referenced axes (projected or geographic) consistent with the other maps. Check the river and lake toponyms in Figures 1 and 9 for spelling and spacing. Also harmonise the unit format for volumetric water content (m³ m⁻³) across captions and tables.
Citation: https://doi.org/10.5194/essd-2026-330-RC1 -
AC2: 'Reply on RC1', Erji Du, 09 Oct 2026
reply
Dear Editor and Referee,
We thank the referee for the careful, constructive and thoroughly useful review of our manuscript. We are glad that the referee recognises the value of a depth-integrated active layer moisture (ALM) field for the permafrost-modelling community, and we have addressed the general comment and all ten specific comments. None of them required new analyses; they concern how the product should be interpreted and used, and we have revised the manuscript accordingly. All changes are marked as tracked changes in the revised manuscript, and the affected locations are indicated in each response below.
General Comment
A 90 m, depth-averaged active-layer moisture field for the entire TP permafrost region fills a real need. It also comes with an honest uncertainty story, combining the AOA with per-pixel layers. The authors openly flag the LST artefact rather than hide it. The field dataset, the multi-ensemble framework, the spatial cross-validation and the open data and code practice are all solid. The comments below are mostly about making the product easier to use correctly and harder to misinterpret. None of them requires new analysis.
One scientific concern unifies the specific comments below. The transferable accuracy of the product is modest and strongly conditioned. The manuscript should make those limits as visible as its strengths. Pixel-level spatial cross-validation (R² ≈ 0.37) tells a far more cautious story than the random-CV and bin-aggregated metrics foregrounded in the Abstract. That skill also rests on well-defined conditions. Only ~60% of the domain lies inside the area of applicability. The training set mixes pit protocols and GPR averages, and these carry method-specific residuals. The 2009–2024 pooling represents a September–October peak-thaw climatology. The predictions also lean on surface-sensitive indices, whose surface–subsurface coupling weakens in some settings. None of this undermines the contribution. Together these factors define the envelope within which the product should be used. Stating that envelope clearly is what will make the dataset most valuable to the modelling community.
Response: We agree with this assessment and have made the conditional nature of the product explicit throughout the revised manuscript. (i) The pixel-level spatial-CV metrics of the delivered fused product (R² = 0.368, RMSE = 0.104 m³ m⁻³) are now reported in the Abstract and the Conclusions alongside the random-CV metrics, and Sect. 3.2 adds the fold-level breakdown (Table 3; Supplementary Table S3). (ii) The interpretation of outside-AOA pixels is now stated in bold in the Data Availability section (Sect. 5) and in the dataset readme, together with the watershed-level outside-AOA fractions (Supplementary Table S3). (iii) The composition and depth-averaging convention of the training profiles (Sect. 2.1), the 2009–2024 pooling, and the peak-thaw (September–October) climatological framing are now stated explicitly in Sect. 2.1, the Abstract and Sect. 5. (iv) Sect. 4.1 now acknowledges the settings in which the surface–subsurface coupling weakens and how the predictions should be read there. The envelope within which the product should be used is therefore now defined explicitly rather than left to the reader.
Specific Comments 1
You already note that ST_B10 pixels along the ~95°, 97° and 99°E swaths were filled with 0 instead of no-data, and that kind of transparency is exactly what reviewers should reward. Please take it one step further: give the actual affected pixel fraction (even a rough percentage) and confirm that these pixels all fall outside the fused AOA mask. Users who will ingest this dataset into models need that number, and stating it removes any doubt about product completeness.
Response: We have added the requested number in Sect. 3.6: the zero-filled pixels account for approximately 2.7% of the permafrost-domain pixels (3.8% of the full 90 m modelling grid) and are concentrated mainly in the 93–100°E sector, with the largest contribution between 95°E and 96°E. We have also verified pixel by pixel that none of them falls inside the fused AOA, so the delivered AOA-masked product is entirely unaffected by the artefact, and users can rely on the masked product without further filtering.
Specific Comments 2
Since the fused AOA covers only ~60% of the domain, the remaining ~40% is extrapolation with higher expected error. I have seen too many users treat every pixel of a full raster as equally reliable, and this product will be picked up by model users who are not permafrost specialists. Please state prominently — bold in the data readme and in the Data Availability section — that outside-AOA pixels should not be read as point estimates, and that the AOA mask is provided precisely for this purpose. It would also help to report the outside-AOA fraction by watershed in the supplement, since much of the uncovered area falls in the Qiangtang interior.
Response: (i) Sect. 5 (Data Availability) now carries the requested statement in bold: users are advised to mask the product to the fused AOA (provided as a companion layer) before any pixel-level interpretation or model ingestion, and pixels outside the AOA represent extrapolation with substantially higher expected error and should not be read as point estimates. The same statement has been added to the readme accompanying the archived dataset. (ii) The outside-AOA fraction by watershed is reported in Supplementary Table S3 (Panel D), with a pointer added in Sect. 3.6. One clarification on the geography: the largest outside-AOA fractions occur along the southern and eastern margins (Brahmaputra 83.5%, Ganges 78.1%, Salween 69.7%), whereas the Qiangtang interior basins have the smallest fractions (Dogai Coring 3.3%, Ayakku 6.5%, Yibug Caka 9.8%, Serling Co 17.0%); those interior basins nevertheless contribute a substantial absolute area because they are large.
Specific Comments 3
The Abstract and Conclusions report the random-CV R² (0.62–0.63) and the bin-aggregated slope of 0.99, but not the pixel-level spatial-CV performance of the fused product itself (Table 3: R² = 0.368, RMSE = 0.104, MAE = 0.081, bias = −0.015 m³ m⁻³). These two numbers tell very different stories about transferable accuracy, and a reader relying on the Abstract alone would clearly overestimate the pixel-level skill of the delivered map. Please report the pixel-level metrics alongside the random-CV metrics, clarify the relation between the bin-aggregated slope and the pixel-level statistics, and include a per-fold breakdown — G1_East (n = 158) is the fold users will worry about, and it deserves to be visible rather than hidden in the table.
Response: We have reported the pixel-level spatial-CV metrics of the fused product (R² = 0.368, RMSE = 0.104 m³ m⁻³, MAE = 0.081 m³ m⁻³, bias = −0.015 m³ m⁻³; Table 3) in both the Abstract and the Conclusions, with an explicit note that they should be read as a conservative estimate of the transferable pixel-level skill of the delivered map. Sect. 3.2 now also gives the fold-level breakdown for the fused product: R² = −0.04 (G1_East, n = 158), 0.45 (G2_Northeast, n = 86), −0.02 (G3_CentralLakes, n = 75) and 0.54 (G4_Peripheral, n = 23), confirming that the eastern fold is the principal accuracy bottleneck; complete fold-level statistics are collected in Supplementary Table S3 (Panel A). We also agree that the two families of metrics should not be conflated, and Sect. 3.4 now states the relation explicitly: the bin-aggregated slope (0.99) quantifies the systematic bias of interval-mean values and is insensitive to the scatter of individual pixels, whereas the pooled R² (0.368) quantifies pixel-level agreement, so the near-unity aggregated slope does not imply pixel-level accuracy of comparable magnitude.
Specific Comments 4
Three things stand out. First, the 0.2 and 0.4 m³ m⁻³ thresholds are empirical values fitted to the spatial-CV residuals of this dataset, so a sentence acknowledging their limited transferability to other regions, periods or training configurations would be appropriate. Second, the thresholds and their evaluation come from the same spatial-CV residuals, which makes the comparison of fusion rules circular with respect to threshold choice; re-running the rules on an independent residual set (e.g., random-CV or LOBO) would settle this. And since the bias-aware rule is marginally below the simple mean at the pixel level (R² = 0.383 vs 0.368), users need a short practical note on when to use which layer. One more numerical detail: the scheme creates a visible dip in the probability density around 0.18–0.20 m³ m⁻³; quantifying it with a finer-grained histogram and commenting on what it means for land-surface permafrost models that consume the product would pre-empt questions later.
Response: We address the four points in turn.
(i) Transferability of the thresholds. Sect. 2.9.3 now states that the 0.2 and 0.4 m³ m⁻³ thresholds are anchored to the sign transitions of the spatial-CV residuals (Sect. 3.2) rather than fitted to the data, so the rule contains no tunable parameters and can be re-applied to another dataset without recalibration. We agree, however, that the two values themselves are dataset-specific, since they are read off the residual structure of the present sample set: applying the rule to another region, period or training configuration requires re-diagnosing the residual sign transitions on the new data, and the gain of the rule over the mean-type alternatives should be re-verified in that setting. Sect. 2.9.3 now states this caution explicitly.
(ii) Circularity of the rule comparison. In Sect. 3.4 we have added independent re-implementations with re-derived thresholds on independent residual sets: across the five fusion rules, the pooled R² spread was 0.003 under a random five-fold design and 0.028 under a three-model group-4 design, with pooled RMSE differences ≤ 0.002 m³ m⁻³ in both cases. The ranking of the rules is therefore stable when the thresholds are re-derived from residual sets different from the one used for evaluation.
(iii) Practical note on the layers. We thank the referee for this observation, and we have corrected our description of it. Table 3 shows that the bias-aware layer is not the best performer on the pooled pixel-level metrics: R² = 0.368 versus 0.383 (simple mean) and 0.387 (R²-weighted mean), with RMSE = 0.104 versus 0.103 and 0.102 m³ m⁻³ and MAE = 0.081 versus 0.081 and 0.080 m³ m⁻³. These differences are practically negligible (ΔRMSE ≤ 0.002 m³ m⁻³, an order of magnitude below the 0.032 m³ m⁻³ retrieval uncertainty of the GPR reference data), and they are traded against the systematic bias structure: the bias-aware rule removes the systematic underestimation of the wet tail (≈ −0.14 m³ m⁻³ for ALM > 0.4 m³ m⁻³) and preserves the full value range, whereas the mean-type layers compress it. We therefore retain the bias-aware layer as the formal product and recommend it where faithful representation of the dry and wet tails matters as much as point-wise accuracy (e.g. permafrost and land-surface modelling), while the simple-mean and R²-weighted layers are provided as sensitivity references for users who prefer the marginally lower pooled RMSE. Sect. 3.4 now states this practical guidance explicitly.
(iv) The density dip. Sect. 4.3 now quantifies the dip: only 1.3% of the permafrost-domain pixels fall in the 0.18–0.20 m³ m⁻³ interval, versus 8.6% in 0.16–0.18 and 16.9% in 0.20–0.22; the R²-weighted alternative, which lacks the low-tail adjustment, shows a monotonic distribution across the same intervals (14.2%, 16.1% and 16.8%), confirming that the dip is a structural consequence of the bias-aware rule rather than a data defect. Sect. 4.3 also comments on the implication for land-surface models: because the dip is structural rather than physical, area fractions and distribution-based diagnostics computed within the narrow 0.18–0.20 m³ m⁻³ band inherit the artefact and should be interpreted with caution, whereas the spatial pattern and the tails of the distribution are unaffected.
Specific Comments 5
The 123 “test pits” turn out to be at least three different protocols — ring-sampled pits, drill cores and manual auger samples — each with its own vertical sampling scheme and depth-averaging convention, mixed with a dielectric-weighted GPR depth average in the same training set. The method-specific residuals differ systematically (from the archived sample dataset: GPR samples are essentially unbiased, while the manual auger samples, whose observed mean exceeds 0.6 m³ m⁻³, are severely underestimated). Please describe the composition and depth-averaging convention of each protocol, and put per-method VWC distributions and residual statistics in the supplement. This is the kind of thing that will save the next user of the sample dataset a lot of detective work.
Response: Sect. 2.1 now describes the composition of the 123 test-pit profiles (37 shallow pits, 80 drill-core profiles, 6 manually drilled profiles) and their depth-averaging convention: all three protocols were sampled with cutting rings at 10 cm depth intervals from the ground surface to the thaw front, and profile ALM is the depth-averaged gravimetric VWC of the sampled column. Per-method observed and predicted means, distribution summaries and residual statistics are given in Supplementary Table S3 (Panel B). The manually drilled profiles are the wettest group of the dataset (observed mean 0.62 m³ m⁻³ versus 0.22–0.27 m³ m⁻³ for the other protocols) and are consequently underestimated most strongly (bias = −0.279 m³ m⁻³); this reflects the scarcity of very wet samples in the training set rather than a deficiency of the drilling protocol. All three profile protocols are direct ring/gravimetric measurements and are treated as equally reliable reference observations; among the four protocols only the GPR survey relies on indirect retrieval, so its velocity-interpretation uncertainty is the only measurement-side error analysed explicitly in the paper (Sect. 2.1).
Specific Comments 6
Pooling 2009–2024 samples into a climatology is defensible, but precipitation varies strongly from year to year in the endorheic basins of the central Qiangtang Plateau, and about 17% of the samples (56 of 342) predate the 2014–2024 predictor window, with no samples at all in 2016 and 2021. A short discussion of the inter-annual noise, the sample year distribution, and a note on early- versus late-period residuals in the supplement would make the climatological interpretation bulletproof. Please also state explicitly in the Abstract and Section 5 that the product represents climatological peak-thaw conditions (September–October) and is not intended for seasonal or inter-annual interpretation.
Response: (i) Sect. 2.1 now states explicitly that pooling suppresses year-specific departures and that the product should therefore be read as the long-term climatological peak-thaw-season ALM state, and gives the per-year sample counts (56 samples in 2009–2013 versus 286 in 2014–2024; Supplementary Table S3, Panel C). (ii) The Abstract and Sect. 5 now state that, as a multi-year composite of peak-thaw-season (September–October) observations, the dataset represents the long-term climatological state of ALM rather than the moisture of any specific year, and that it is not intended for seasonal or inter-annual interpretation. (iii) The sample-year distribution is now documented in Supplementary Table S3 (Panel C) and Sect. 4.5 discusses the temporal representativeness of the product. We note that we do not report a separate early- versus late-period residual comparison: the 56 pre-2014 samples are unevenly distributed among the four spatial folds, so such a comparison would largely reflect the fold structure and the basin-specific biases diagnosed in Sect. 3.4 rather than a temporal effect, and we have therefore made the temporal limitation explicit in the text instead.
Specific Comments 7
The GPR+CRIM calibration against the pits is excellent (RMSE ≈ 0.032 m³ m⁻³), but microwave sensitivity drops below roughly 1–2 m while QTP active layers can exceed 3 m. A short caveat on how representative the column average is in thick-ALT areas, and on whether calibration error is expected to grow with depth, would pre-empt over-interpretation. Related to this: the product is a depth average, whereas land-surface models output layered moisture, and computing active-layer water storage requires an external active-layer-thickness (ALT) dataset; a sentence or two on these comparison challenges in Sect. 4.5 or 5 would help that community use the dataset correctly.
Response: Sect. 4.5 now carries both caveats. First, the product holds no vertical information: it is a depth average over the full thawed column, and its use as a proxy for any specific depth interval (e.g. the 0–5 cm surface layer) is not supported by the retrieval. Second, the representativeness of the column average degrades as the active layer thickens: the radar velocity retrieved from the common-midpoint profiles is a bulk value for the whole thawed column and retrieval sensitivity decreases with depth as signal attenuation increases, so in active layers thicker than about 3 m an increasing share of the column average comes from the deep, less well-constrained part of the profile, and the effective uncertainty of the depth-averaged ALM should be expected to increase with active-layer thickness. The same paragraph also states that computing active-layer water storage from the depth-averaged field requires an external active-layer-thickness dataset.
Specific Comments 8
The feature-importance analysis shows that surface-sensitive remote sensing indices dominate (58–69%) the predictive weight, which presumably reflects strong surface–subsurface hydrothermal coupling on the plateau. I would appreciate a brief acknowledgement of the situations in which that coupling weakens — very deep active layers, coarse-textured surface horizons above shallow supra-permafrost water, or thick organic mats — and a note on how predictions should be interpreted in such areas.
Response: We thank the referee for this point. Sect. 4.1 now clarifies that the 58–69% contribution is reported relative to the other candidate predictor groups (terrain, soil physicochemical and climate), because the contributions of all selected predictors sum to 100% within each model; it therefore expresses the share of the predictive weight carried by the surface-sensitive group rather than its absolute explanatory power. Sect. 4.1 also acknowledges the situations in which the coupling weakens—very deep active layers (thicker than roughly 2–3 m), coarse-textured surface horizons overlying shallow supra-permafrost groundwater, and thick organic mats—and states that the depth-averaged ALM should be interpreted with corresponding caution in such areas, because the product is most reliable where the surface state and active-layer moisture co-vary strongly and least reliable where the two are decoupled.
Specific Comments 9
Sect. 2.8 still contains the placeholder “(citation to be added)” for the ChinaMet petPM dataset; the complete citation must be supplied. Since the 1 km field is bilinearly resampled to the 90 m grid, it is also worth stating explicitly that the resampling only aligns the grids and adds no information at the 90 m scale.
Response: The placeholder "(citation to be added)" in Sect. 2.8 has been replaced by the complete citation of the ChinaMet petPM dataset — Hu, Y., Zhang, L., Zhao, Y., and Che, T.: ChinaMet: A multisource integrated high-resolution multi-variable meteorological dataset for China, Zenodo [dataset], https://doi.org/10.5281/zenodo.14580706, 2024 — which has been added to the reference list (Hu et al., 2024). Sect. 2.8 also now states that the ChinaMet annual PET fields are produced from multi-source meteorological forcing (station observations, reanalysis and remote sensing), and that the bilinear resampling of the 1 km fields to the 90 m grid only aligns the grids and adds no information at the 90 m scale; pet_mean mainly controls the regional-scale spatial pattern of ALM and should not be interpreted as providing sub-grid-scale information.
Specific Comments 10
A few small points. The Abstract says "≈60%" AOA coverage, while Table 5 and Fig. 11 give 59.9%. Please use one figure throughout. Unify the terminology for the target variable (ALM vs VWC; "showed as VWC" in the Figure 2 note should read "shown as VWC"). Fix the spacing typos ("slope of0.99", "The0.20–0.33 drop") and change "include" to "including" in the Abstract. Both panels of Figure 11 are annotated with bare raster indices. Please re-label them with map-referenced axes (projected or geographic) consistent with the other maps. Check the river and lake toponyms in Figures 1 and 9 for spelling and spacing. Also harmonise the unit format for volumetric water content (m³ m⁻³) across captions and tables.
Response: We have made the following corrections. (i) The AOA coverage is now given as 59.9% throughout (Abstract, Sect. 4.4 and the Conclusions; Table 5 and Fig. 11 already used 59.9%). (ii) The target variable is now referred to as ALM consistently: the parenthetical alternative abbreviation in Sect. 3.1 has been removed, so the term is defined once in Sect. 1 and used uniformly thereafter. (iii) The spacing typos have been corrected ("slope of 0.99", "The 0.20–0.33 drop") and the Abstract reads "including" rather than "include". (iv) The bare raster indices in both panels of Figure 11 have been replaced by geographic coordinates (°E/°N) in the updated figure accompanying this revision. (v) The river and lake toponyms in Figures 1 and 9 have been re-checked. (vi) The unit format for volumetric water content (m³ m⁻³) is now consistent across the captions and tables. We have also re-checked the revised text for phrasing duplicated during revision: the pooled-metrics sentence in the Conclusions no longer repeats the Abstract verbatim, the watershed-grouping sentence in Sect. 2.9.2 now states explicitly that 12 of the 14 watersheds listed in Table 4 contain training samples, and the corresponding Conclusions sentences no longer repeat the Fig. 5 and Fig. 8 captions.
In summary, both the general comment and all ten specific comments have been addressed by textual revisions that are fully marked in the tracked-changes version of the revised manuscript; none of the changes alters the data products, the numerical results or the core conclusions of the paper. We believe the revised manuscript is now clearer and harder to misinterpret for the model- and permafrost-science communities it targets, and we hope it is now suitable for publication in Earth System Science Data. We thank the referee once more for the time and expertise invested in this review.
Sincerely,
Erji Du, Tonghua Wu, Liyun Dai, Youhua Ran, and co-authors
on behalf of all authors of essd-2026-330
-
AC2: 'Reply on RC1', Erji Du, 09 Oct 2026
reply
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 338 | 132 | 179 | 649 | 83 | 213 | 162 |
- HTML: 338
- PDF: 132
- XML: 179
- Total: 649
- Supplement: 83
- BibTeX: 213
- EndNote: 162
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript presents a spatially continuous, 90 m resolution dataset of peak-thaw season active layer moisture (ALM) across the Qinghai–Tibet Plateau (QTP) permafrost domain. By synthesizing 342 in situ observations (drill/pit core sampling and GPR surveys) with multi-source remote sensing, soil properties, and terrain factors, the authors trained four ensemble learning models and implemented a bias-aware multi-model fusion framework. Rigorous spatial validation and an Area of Applicability (AOA) analysis are provided alongside per-pixel uncertainty estimates. The dataset addresses an important gap in cold-regions science by moving beyond shallow skin-layer satellite soil moisture to characterize the column-integrated active layer state. The data and scripts are archived openly and comply with the standards of Earth System Science Data. The manuscript is well structured, but several minor points regarding physical interpretation, the multi-model fusion scheme, and text formatting require attention prior to publication.
1. The feature importance analysis reveals that surface-sensitive remote sensing indices contribute 58%–69% of the predictive weight across models. While this reflects strong surface–subsurface hydrothermal coupling on the plateau, the authors should briefly acknowledge scenarios where decoupling occurs—such as deep active layers (>2–3 m) with coarse surface layers overlying perched water tables (supra-permafrost groundwater) or thick organic mats, where surface signals might have weaker sensitivity to bottom-layer moisture dynamics.
2. Given that in situ samples were pooled from 2009 to 2024 while remote sensing predictors were compiled as 2014–2024 composites, the resulting product is explicitly a long-term climatological representation of the annual maximum thaw state (September–October). Ensure this distinction is consistently highlighted in the Abstract and User Guidance to avoid users misinterpreting the product as a time-slice snapshot of any single year.
3. Because the four-model intersection AOA covers ~60% of the permafrost domain, approximately 40% of the region represents an environmental extrapolation. The authors have handled this commendably by publishing the companion AOA and expected-RMSE layers. A sentence should be added to the Data Availability section reminding users to apply the AOA mask when reliable empirical bounds are required.
4. Line 254: Remove the placeholder "...ChinaMet 0.01° (approximately 1 km) gridded dataset of annual Penman-Monteith PET (petPM; citation to be added) for 2009-2024" and supply the missing citation.
5. Line 322: Correct the typo in sample size: "G4_Peripheral (Tarim + Hexi + Salween R., n=2?)".
6. Figure 11: The axes for both panels currently display raw grid row/column counts. Please replace these with projected coordinates (km) or geographic coordinates (°E/°N) to match the other maps in the paper.
7. Figures 1 & 9: Ensure consistent typography for river and lake labels across maps (e.g., standardizing spaces in “Bangong Co”, “Yibug Caka”, and checking “Brhmaputra R.” for spelling).