SETP_GLI: An annual 10&ndash;30 m glacial lake inventory for the southeastern Tibetan Plateau from 1990 to 2025

Li, Hao; Dou, Jie; Kusky, Timothy; Dong, Shun; Shi, Zihao; Li, Jie; Xiang, Xinjian; Ding, Fange

doi:10.5194/essd-2026-452

Preprints

https://doi.org/10.5194/essd-2026-452

Preprints

09 Jul 2026

| 09 Jul 2026

Status: this preprint is currently under review for the journal ESSD.

SETP_GLI: An annual 10–30 m glacial lake inventory for the southeastern Tibetan Plateau from 1990 to 2025

Hao Li, Jie Dou, Timothy Kusky, Shun Dong, Zihao Shi, Jie Li, Xinjian Xiang, and Fange Ding

Abstract. Glacial lakes in the southeastern Tibetan Plateau (SETP) have expanded, increasing the potential for cascading hazards associated with glacial lake outburst floods (GLOFs). However, long-term, annual monitoring data that include micro glacial lakes remain relatively limited for this region. To address this gap, this study integrated Landsat series and Sentinel-2 imagery and used the GLA-RCNN deep learning framework with an embedded Convolutional Block Attention Module to construct and release an annual glacial lake inventory (SETP_GLI). The dataset comprises 36 annual vector layers from 1990 to 2025, recording the annual evolution of regional glacial lake numbers and areas. The use of 10 m resolution imagery and model optimization improved the detection of micro glacial lakes (<0.01 km²). The inventory provides annual vector boundaries and standardized physical attributes—including longitude, latitude, area, perimeter, and mean elevation, together with area uncertainty metrics derived from mixed-pixel theory. Quality assessments indicated that the extraction framework is robust against interference from mountain shadows and turbid water. For model performance, the overall F1 scores for typical years remained above 0.82 (with a maximum of 0.895); cross-validation with existing public databases (Hi-MAG and Glacial lake inventory of high-mountain Asia) showed that the matched polygon-level Intersection over Union (IoU) ranged from 0.54 to 0.80, with spatial agreement increasing with improvements in historical image quality. Spatiotemporal analysis revealed a persistent expansion trend, with the annual area growth rate rising from 3.65 ± 1.12 km² a⁻¹ (1990–2012) to 5.95 ± 2.44 km² a⁻¹ (2016–2025). The dataset is archived at the National Tibetan Plateau Data Center (TPDC) (https://doi.org/10.11888/Cryos.tpdc.303491), with processing code released openly. SETP_GLI serves as a baseline dataset for cryospheric response analysis, hydrological modeling, and GLOF risk assessment.

Received: 10 Jun 2026 – Discussion started: 09 Jul 2026

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

Download & links

Hao Li, Jie Dou, Timothy Kusky, Shun Dong, Zihao Shi, Jie Li, Xinjian Xiang, and Fange Ding

Status: final response (author comments only)

RC1:
'Comment on essd-2026-452', Anonymous Referee #1, 03 Aug 2026
General comments
Li and Dou et al. present a dataset of glacial lakes on the southeastern Tibetan Plateau derived using a GLA-RCNN deep learning framework they developed and applied to Landsat series and Sentinel-2 imagery. The dataset has the potential to provide a useful contribution and is presented in a suitable format with associated metadata. Issues requiring addressing are detailed below and mainly relate to: (1) the erroneous use/ misunderstanding about the origin and date of the ‘ALOS PALSAR RTC DEM’; (2) lack of details about the model training and notable manual adjustments performed to the outputs at several stages; (3) overstated improvements of the developed model with respect to other models, as there are negligible (<0.01) differences in e.g. F1 scores; (4) the stepped increase in glacial lake area and count resulting from using imagery of 30 m and 10 m (post 2016) resolution, which was identified, but is not addressed; (5) the paper does not define what qualifies as a ‘glacial lake’, as other water bodies including in non-glacierised catchments (e.g. TRK_03213), appear to be included in the dataset; (6) there are inconsistencies in the timeseries for individual lakes, potentially resulting from the availability/quality of imagery in a given year, causing lakes to be missed. I would support publication if the authors address these issues.

Specific comments;
A range of uncommon terminology is used to describe glacial lakes including ‘micro’, ‘tiny’, ‘small’. This should be made consistent throughout.

Statistics should be reported for the number and percentage of lakes that were manually adjusted, and the amount by which outlines were adjusted. For example, the change in lake area (area/percentage) following adjustment. Three operators made manual corrections, so there should also be an inter-operator comparison to test agreement.

This F1 score seems to be inconsistent with those reported later in Table 2.

Figure 1. ‘Typ-glacial Lakes’. These appear to be Google Earth (or similar) screenshots/views of lakes, but the images are not clear or labelled. Remove these or replace with satellite nadir views. Show the dataset produced in this paper on the map, rather than another inventory.

‘ALOS PALSAR RTC DEM’. It appears two different DEM products were used due to a misunderstanding of their data date/origin. ‘SRTM DEM was used for slope derivation for the Landsat-period topographic auxiliary layer, and ALOS PALSAR RTC DEM was used for slope derivation for the Landsat 8 and Sentinel-2 topographic auxiliary layer’. The ALOS PALSAR RTC DEM with a specified date range of 2013–2025 in Table 1, is a resampled SRTM in most regions that was created by ASF for processing ALOS PALSAR data. It is not a new DEM product at 12.5 m resolution. Change the analysis to use a single DEM product. The specified date range for SRTM is also incorrect, it should be 2000, not 1990–2012.

‘Because Sentinel-2 does not provide a 10 m SWIR band, the input channel number was inferred automatically in the model configuration rather than forced to a fixed band combination for all sensors’. Please clarify this.

‘a dodging algorithm was applied to color-balance the imagery for each year’. Provide a description and citation of this algorithm. Clarify how the improvement for the ‘temporal consistency of feature extraction’ was evaluated.

Clarify the bit depth of imagery supplied to the model after the normalization.

L144-145. Resampling the DEM to match the imagery resolution does not lead to ‘pixel-level registration’. Coregistration offsets could still be present. Rephrase to state they were resampled to the same grid.

Elaborate on what ‘high frequency noise’ is expected to be present in the glacial lake environments.

Figure 2. Remove Radiometric Calibration and Atmospheric Correction as these were already applied as part of the surface reflectance product used. ‘STREM DEM’ spelling mistake. What is the multidimensional feature dataset? Technical validation section: change these figures to simplify as the details/text is not legible. Perhaps simplify to single panels or an illustrative figure.

Information regarding the imagery used for training is lacking, including the dates and geographic coverage, and if all the imagery is from the same date as the published inventory used as reference.

Clarify how many lakes and chips were used in the training and validation datasets.

‘The learning rate was selected during model training, and the total number of training epochs was set to 30.’ Please elaborate on how the number of epochs was set, and why/if it was fixed, rather than allowing the model to reach an optimum during training.

‘elongated river valleys are frequently misidentified as glacial lakes due to spectra similarities’. Please elaborate on this and provide some example in supplementary information. Substantial rivers should be morphologically and spectrally dissimilar to glacial lakes in most cases.

And L253. There appears to be notable manual adjustment of the dataset in these sections over several stages. Please quantify how many lakes were modified in each step and the size of those changes (e.g. percentage change).

L266-269. 200 sample points is small for such a large dataset. It is not clear how ‘typical error-prone regions’ were identified to perform the stratification, or what was measured/compared during the ‘manual visual cross-validation’. The sample size should be increased, and expanded to span the temporal extent of the dataset.

The mean elevation was assigned to lakes using a dated (2000) DEM. This is likely to introduce notable errors where glaciers have retreated to become lakes, since the elevation values will still reflect the glacier surface. A newer DEM should be used, or use a method to more robustly assign and quality check elevations.

This statement is not supported by the results as the model was not tested without CBAM to quantify the improvements in shadow areas.

L309 Table 2. The model improvement between the Mask R-CNN and GLA-RCNN is negligible. (<0.01 difference for F1 scores). U-Net is also similar. This means the stated significant improvement over other models is not supported by the results ‘This performance not only significantly surpasses that of mainstream U-Net and DeepLab V3 but also quantitatively validates the contribution of the CBAM attention mechanism in filtering out shadow noise interference and precisely restoring the topological structure of contiguous water bodies.’

Figure 8f. There are linear misclassification edges on both glacial lakes shown. Discuss the potential origins and check other locations for issues.

L406-507. ‘this study strictly limits the observation window to September–December of each year.’ This is not consistent with earlier text suggesting the observation window was extended if suitable imagery was not available. It is not clear what classifications were completed on single images vs image composites spanning multiple acquisitions. Please clarify this in a supplementary table to show the images used.

F1 scores should also be added to Table 4.

The identified step means lakes pre and post 2016 cannot be compared. Applying a uniform 0.01 km2 threshold filter to all data does not then allows the datasets to be compared across the full timeseries. This does not address the model inference of individual lakes, where the boundaries will be biased by different image resolutions, making them more precise in the Sentinel-2 data and less in the Landsat series. There are studies that have shown this e.g. https://doi.org/10.1016/j.rse.2018.08.020. Training separate models for the different datasets is likely to have reinforced this issue. A better approach could be to train a single model on the entire dataset resampled to e.g. 30 m resolution, and then inferencing on imagery also at 30 m resolution. This could support creating a consistent dataset through time, which could then be supplemented with the higher-resolution detections from Sentinel-2. Related to this point, there are inconsistencies in the timeseries for individual lakes, likely resulting from the available imagery in a given year, causing lakes to be missed. For example, lake, iD TRK_03326, which is missing in years 1991, 1993, 1994, 1997, 2010. Or lake ID TRK_03327. The dataset should identify persistent lakes that are missed in some years if no suitable imagery was available for classification.

‘This result indicates a sustained regional cryospheric and hydrological response to climate warming.’ No climate data are presented to support this statement.

Table 6 requires restructuring. ‘This study’ and ‘Key improvements’ are separated by two other datasets, making the comparisons confusing.

Technical corrections:
L13. Define ‘GLA-RCNN’
L145. Change the DEM and correct references to ‘ALOS DEM’ or similar.
L151. ‘sensor-specific’ not ‘adaptive’.
L222. Clarify ‘full manual verification’. Were all lake polygons in the full dataset checked?
L330 . 41,916.3 m². Check whether the reported precision and number of significant figures are appropriate.
L474. Change ‘apparent numerical growth’ to ‘increase in lake count’ or similar.
L477. Replace/define ‘mathematical level’.
L493. ‘minute glacial lakes’. Make terminology consistent and define area thresholds.
L595. Remove ‘target’.
Citation: https://doi.org/10.5194/essd-2026-452-RC1
RC2:
'Comment on essd-2026-452', Anonymous Referee #2, 08 Sep 2026
This manuscript presents a valuable contribution to glacial lake monitoring in a critical region, with a well-designed methodological framework and a comprehensive dataset. However, several issues require attention before publication. Major comments are as follows.
GLA-RCNN Architecture Description is unclear. Should provide a detailed architecture diagram showing where CBAM modules are inserted (after which residual blocks). Specify the reduction ratio used in channel attention and kernel sizes in spatial attention.

Training Data Generation. While the text notes that ground-truth labels were manually produced using Wang et al. (2020) as a reference, the specific methodology for the 1991 training node remains ambiguous, given that Wang et al. only provides inventories for 1990 and 2018. The authors should explicitly clarify how contemporaneous imagery was utilized for 1991 manual digitization to avoid temporal mismatch. Furthermore, to ensure data reliability, the manuscript must provide essential operational details: who performed the digitization, what quality control protocols were implemented, and what inter-operator consistency metrics were achieved.

Sensor Transition Handling. Lines 186-191 mention training separate models for different resolutions but don't explain how predictions are harmonized for the full time series. Explain whether separate models were applied to each period (1990-2012, 2013-2015, 2016-2025) or a single model with resolution-aware inference was used. How model performance differences across resolutions were addressed in the final product.

Validation Limitations. Lines 258-262 state that the Hi-MAG and Wang et al. (2020) databases were used for accuracy assessment, but Lines 96-98 indicate these same datasets assisted in label generation. Using the same data to inform training and conduct validation introduces a fundamental circularity bias. This not only artificially inflates spatial agreement metrics but also penalizes the model's high-resolution innovations, as new micro-lakes missed by the older 30m reference datasets will register as false positives. Please ensure a strict separation between training and validation sets. If the dual-use of these datasets is unavoidable, the authors must explicitly acknowledge this limitation in the text and thoroughly discuss how this circularity biases the reported accuracy metrics.

Ground Truth Quality for Small Lakes. Lines 371-373: Google Earth imagery used as validation for <0.01 km² lakes. But google Earth imagery has varying dates and resolutions in the SETP. The author may provide specific dates of Google Earth imagery used and quantify the resolution and quality, and discuss potential misalignment between Google Earth imagery and Landsat/Sentinel-2 acquisition dates.

Uncertainty Quantification. Lines 274-284: Area uncertainty calculated using Hanshaw and Bookhagen (2014) boundary pixel method. This method doesn't account for georeferencing errors (especially in early Landsat), seasonal water level fluctuations and shadow-induced boundary shifts. May add a discussion of additional uncertainty sources.

Temporal Consistency Issues. Lines 102-119 mentioned the primary window for image selection are Sept-Nov, backup August, then Dec-Feb. This introduces seasonal variability that could obscure long-term trends. Lakes in December may be partially frozen; August lakes may be at peak melt. Can quantify the proportion of lakes from each month and analyze whether there's a systematic bias in lake area from different months. Lines 226-233 say sensor-adaptive thresholds of 0.001 km² (10m), 0.002025 km² (15m), and 0.0054 km² (30m). This means the minimum detectable lake size changes over time. Lines 473-485: There's a gap between Landsat and Sentinel-2 coverage. How was 2014-2015 handled? Was Landsat 8 used exclusively? What about 2016 transition year? The author can provide detailed per-year information on primary sensor used, the number of scenes and cloud cover statistics and any data gaps and how they were filled.

Minor comments.
Lines 287-298: GLA-RCNN compared with Mask R-CNN, U-Net, and DeepLab V3. Were the comparison models optimized for this task? Should specify training details for comparison models (learning rate, epochs, data augmentation). Were hyperparameters tuned for each model? Should ensure a fair comparison with equal training resources.

Lines 165-170: "CBAM modules were embedded into each fused feature layer (P2–P5)." This is non-standard. How were the CBAM outputs integrated with the FPN features? The authors should explain precisely how CBAM outputs were integrated with FPN features and provide a schematic diagram of the modified architecture.

Lines 133-134: "input channel number was inferred automatically in the model configuration." Given that Sentinel-2 (10m bands) and Landsat (30m bands) feature different band configurations, the authors must clarify how the model handles these variable input dimensions and discuss any potential spectral inconsistencies introduced.

Lines 174-179: The loss function description is very technical. Suggest to provide a brief intuitive explanation of what each loss term does to improve readability.

Lines 521-544 repeats the Introduction. Suggest to focus Discussion on the implications of the new dataset, not reiterating its features.

Figure 2, correct ‘STREM’ to ‘SRTM’

Figure 3, ensure all text can be seen clearly in all panels.

Figure 10, what is raw and MMU trend. Explain in figure caption or main text.
Citation: https://doi.org/10.5194/essd-2026-452-RC2

Hao Li, Jie Dou, Timothy Kusky, Shun Dong, Zihao Shi, Jie Li, Xinjian Xiang, and Fange Ding

Data sets

Annual glacial lake inventory dataset for the southeastern Tibetan Plateau from 1990 to 2025 H. Li et al. https://doi.org/10.11888/Cryos.tpdc.303491

Model code and software

GLA-RCNN-SETP-GLI: GLA-RCNN code for SETP_GLI v1.0.1 AI-Geohazard https://doi.org/10.5281/zenodo.20555222

Hao Li, Jie Dou, Timothy Kusky, Shun Dong, Zihao Shi, Jie Li, Xinjian Xiang, and Fange Ding

Viewed

Total article views: 547 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
407	87	53	547	38	61

HTML: 407
PDF: 87
XML: 53
Total: 547
BibTeX: 38
EndNote: 61

Views and downloads (calculated since 09 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	176	35	9	220
Aug 2026	133	21	6	160
Sep 2026	98	31	38	167

Cumulative views and downloads (calculated since 09 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	176	35	9	220
Aug 2026	133	21	6	160
Sep 2026	98	31	38	167

Viewed (geographical distribution)

Total article views: 457 (including HTML, PDF, and XML) Thereof 457 with geography defined and 0 with unknown origin.

Country	#	Views	%

Cited

Latest update: 22 Sep 2026

Short summary

Climate change is melting glaciers in the Tibetan Plateau, creating lakes that can burst and cause floods. To track this, we used artificial intelligence to analyze thirty-six years of satellite images from 1990 to 2025. We discovered these mountain lakes are expanding quickly, with growth accelerating over the last decade. This new dataset will help predict future flood risks, manage water resources, and protect downstream communities.


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

SETP_GLI: An annual 10–30 m glacial lake inventory for the southeastern Tibetan Plateau from 1990 to 2025

Data sets

Model code and software

Viewed

Viewed (geographical distribution)

Cited

1 citations as recorded by crossref.