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
- AC2:
  'Reply on RC1', hao li, 02 Oct 2026
  RESPONSE TO REVIEWER 1
  SETP_GLI: An annual 10–30 m glacial lake inventory for the southeastern Tibetan Plateau
  
  We thank the reviewer for the careful and constructive assessment of the dataset, DEM sources, model training and validation, model comparison, cross-sensor consistency, glacial-lake definition, and temporal completeness. We have revised the manuscript and dataset documentation to correct the DEM source and date descriptions, clarify the training and manual-adjustment procedures, moderate the model-comparison claims, establish a common 30 m Landsat 7/8/9 baseline for 2013-2025, define the operational criteria for glacial lakes, and clarify target-year image compositing and missing annual observations. The response below first addresses these five general comments and then provides detailed responses to the specific comments.
  General comment 1 - DEM source and date
  Reviewer comment
  The origin and date of the ALOS PALSAR RTC DEM were incorrectly described, and the application periods were confused with the source-data dates.
  Response
  Thank you for identifying this issue. We corrected the DEM description by distinguishing the SRTM-derived source dates from the periods of application. NASADEM HGT v001 is reported with its 2000 source date, and the ancillary DEM distributed with the ASF ALOS PALSAR RTC product is described as an SRTMGL1-derived terrain-correction layer with a 30 m source resolution and a 12.5 m output grid. The application periods are now stated separately from the source-data dates.
  Changes made in the manuscript
  Table 1 and Sections 3.1, 4.1, and 4.5, together with the Data Records and References, were revised to correct the DEM source names, source-data dates, output grids, and periods of application.
  
  General comment 2 - Training details and manual adjustments
  Reviewer comment
  Details about model training and the notable manual adjustments applied to the outputs at several stages were insufficient.
  Response
  We added details on the training imagery, geographic coverage, image preparation, validation data, model stopping criterion, and manual correction procedure. The current inventory contains 7,660 manually adjusted polygons out of 174,777 (4.38%), with period-specific proportions reported in the response to Comment 2. These values refer to the current inventory. After the planned 1990-2012 target-year pixel-level supplementation and re-extraction, the correction statistics will be recalculated; improved coverage is expected to reduce the need for manual corrections.
  Changes made in the manuscript
  Sections 4.3 and 5.2, the Data Records, and the relevant training and validation descriptions were expanded to report the training workflow and the reproducible manual-adjustment statistics.
  
  General comment 3 - Model performance claims
  Reviewer comment
  The improvements of the developed model over other models appeared overstated because the differences in metrics such as F1 score were negligible.
  Response
  We moderated the statements about performance differences relative to U-Net, DeepLab V3, and Mask R-CNN, and added the CBAM ablation results. The revised text describes the measured differences without claiming unsupported significant superiority, while retaining the evidence for the model's performance in shadow-affected scenes.
  Changes made in the manuscript
  The model-comparison text, tables, and CBAM discussion were revised to report the observed metric differences and the ablation results without overstating their significance.
  
  General comment 4 - Cross-sensor and temporal consistency
  Reviewer comment
  The stepped increase in glacial lake area and count associated with the use of 30 m and post-2016 10 m imagery was identified but not addressed, and individual-lake time series contained gaps related to image availability or quality.
  Response
  To reduce the resolution-related step in the long-term record, we established a common 30 m Landsat 7/8/9 baseline for 2013-2025 and retained the 10 m Sentinel-2 layers as higher-resolution companion observations. We now state explicitly that a common area threshold cannot remove all boundary differences between sensors. We also clarified target-year image selection and pixel-level cloud-gap supplementation. The same target-year supplementation and re-extraction will be extended to 1990-2012 to improve continuity for variable glacial lakes. Missing annual records are treated as unavailable or uncertain observations rather than confirmed lake disappearance.
  Changes made in the manuscript
  Sections 3.2, 4.1, 4.3, 5.3, 8.2, and 8.4, together with Supplementary Table S1 and the dataset documentation, were revised to describe the common Landsat baseline, target-year compositing, cross-sensor limitations, and the interpretation of missing annual records.
  
  General comment 5 - Definition of a glacial lake
  Reviewer comment
  The paper did not define what qualifies as a glacial lake, and other water bodies in non-glacierised catchments appeared to be included in the dataset.
  Response
  Thank you for identifying this ambiguity. We added operational criteria based on present or former glacial association, including glacier-fed, moraine-dammed, recently deglaciated, and glacially eroded basins, and excluded artificial reservoirs, river reaches, seasonal ponds, and water bodies without evidence of glacial association. The annual layers were re-screened using these criteria. TRK_03213 was retained because its basin shows evidence of a former glacial geomorphic setting, and the revised inventory was used to update the relevant statistics, figures, and conclusions.
  Changes made in the manuscript
  The Introduction and References were revised to define glacial lakes operationally, specify included and excluded settings, and document the re-screening of the inventory.
  
  Comment 1 - Terminology for small glacial lakes
  Reviewer comment
  A range of uncommon terminology is used to describe glacial lakes including 'micro', 'tiny', and 'small'. This should be made consistent throughout.
  Response
  Thank you for this suggestion. We now use "small glacial lakes" consistently throughout the manuscript and define them operationally as lakes with an area <0.01 km². The definition is introduced in the Introduction with references to regional glacial-lake inventories (Nie et al., 2017; Wang et al., 2020; Zhang et al., 2022). The terms "micro", "tiny", "minute", and related target-based expressions have been replaced accordingly.
  Changes made in the manuscript
  Abstract, Introduction, Methods, Results, Discussion, Conclusions, figure captions, and Table 7: standardized the terminology to "small glacial lakes" and added the <0.01 km² operational definition.
  
  Comment 2 - Manual correction statistics
  Reviewer comment
  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.
  Response
  Thank you for this comment. We added the statistics of manually adjusted polygons to Section 5.2. Across all 36 annual layers, 7,660 of 174,777 polygons (4.38%) were manually adjusted. The proportions were 8.82% (5,464/61,953) for 1990–2012, 3.78% (807/21,357) for 2013–2015, and 1.52% (1,389/91,467) for 2016–2025. Thus, the proportion of manually adjusted polygons decreased progressively across the three sensor periods. The released annual layers now also include a binary ManualAdj field (1 = manually adjusted; 0 = not marked as manually adjusted). These statistics describe the current native-resolution inventory. The planned pixel-level supplementation and re-extraction of the 1990–2012 record is expected to reduce the need for manual corrections, and the counts and percentages will be recalculated after reprocessing and quality review.
  Changes made in the manuscript
  Sections 5.2 and 6 (Table 6): added the total and period-specific numbers and percentages of manually adjusted polygons and documented the ManualAdj attribute supplied with the released annual layers.
  
  Comment 3 - F1-score consistency
  Reviewer comment
  This F1 score seems to be inconsistent with those reported later in Table 2.
  Response
  Thank you for identifying this inconsistency. The F1 scores in the Abstract refer to validation of the final annual inventories against manually delineated reference data, whereas Table 2 reports model performance on independent image-chip test and validation sets. We corrected the erroneous annual maximum of 0.895 and clarified both evaluation levels. The annual-inventory F1 scores were 0.830, 0.887, and 0.875 for 1991, 2013, and 2016, respectively, while the GLA-RCNN F1 scores in Table 2 were 0.9167 and 0.9236 for the test and validation sets, respectively. Section 5.2 was also revised to report all three annual values explicitly.
  Changes made in the manuscript
  Abstract and Section 5.2: distinguished annual-inventory validation from image-chip model evaluation, corrected the maximum annual F1 score from 0.895 to 0.887, and reported the annual F1 scores for all three representative years.
  
  Comment 4 - Figure 1
  Reviewer comment
  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.
  Response
  Thank you for this comment. We revised Figure 1 by removing the Google Earth views of typical glacial lakes and replacing the external reference inventory with the 2025 SETP_GLI layer produced in this study. The revised map now directly shows the spatial distribution of the published dataset within the study area, and the figure caption has been updated accordingly.
  Changes made in the manuscript
  Figure 1 and its caption: removed the Google Earth lake views and replaced the external reference inventory with the 2025 SETP_GLI layer.
  
  Comment 5 - DEM source and date
  Reviewer comment
  
  'ALOS PALSAR RTC DEM'. It appears two different DEM products were used due to a misunderstanding of their data date/origin. The ALOS PALSAR RTC DEM specified as 2013-2025 is a terrain-correction ancillary layer that is based on resampled SRTM in most regions, rather than a new 12.5 m DEM. 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.
  
  Response
  Thank you for pointing this out. We corrected the DEM descriptions in Table 1 and the relevant Methods text. Both terrain inputs are now treated as static SRTM-derived representations of the 2000 acquisition rather than as DEMs from two different time periods. NASADEM HGT v001 is identified as an SRTM-derived DEM with a source date of 2000 and a 1 arc-second (~30 m) grid. The ancillary *.dem.tif distributed with the ASF ALOS PALSAR RTC product is described as an SRTMGL1-derived DEM with a 30 m source resolution and a 12.5 m output grid. In Table 1, the source date is reported in the Source-data date column, while the periods of application (1990–2012 and 2013–2025) are stated in the Key Usage column. The 12.5 m RTC layer is therefore not interpreted as an independent later DEM; its existing output grid is retained only for the corresponding processing period. This correction resolves the source-date ambiguity, but the two processing-period inputs remain distinct DEM-derived products rather than a single DEM product.
  Changes made in the manuscript
  Table 1, Sections 3.1/4.1, and References: corrected the DEM product names, source dates, source resolution, output-grid spacing, periods of application, and citation.
  
  Comment 6 - Sensor-specific input channels
  Reviewer comment
  
  '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.
  
  Response
  Thank you for this comment. We clarified the actual band combinations and the meaning of automatic channel inference in Section 4.1. The 30 m Landsat 7/8/9 input stacks comprised five harmonized optical bands (blue, green, red, NIR, and SWIR), NDWI, and slope, resulting in seven channels. Because a 10 m SWIR band is not available for Sentinel-2, the Sentinel-2 input comprised four optical bands (blue, green, red, and NIR), NDWI, and slope, resulting in six channels. The "in_channels: auto" setting reads the channel number from the prepared image chips before model initialization and configures the first convolutional layer. It does not select bands automatically, and the channel number remains fixed within each trained model.
  Changes made in the manuscript
  Section 4.1: specified the harmonized Landsat 7/8/9 and Sentinel-2 band combinations and channel counts, and clarified the function of the "in_channels: auto" setting.
  
  Comment 7 - Dodging color balancing
  Reviewer comment
  
  '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.
  
  Response
  Thank you for this comment. We expanded Section 4.1 to describe the dodging color-balancing algorithm. The method derives local image statistics within adaptive windows and uses a gamma function to adjust pixel values toward a spatially varying target color surface (Zhou, 2015). Because an annual mosaic may contain scenes acquired in different months, this processing reduces brightness and color discontinuities at scene boundaries and provides more consistent within-year inputs for feature extraction. We did not conduct a separate before-and-after experiment for this preprocessing step; the manuscript therefore describes its harmonization purpose and cites prior evidence (Ulfa et al., 2023) without claiming a separately quantified improvement in this dataset.
  Changes made in the manuscript
  Section 4.1 and References: added the dodging algorithm description and citations, clarified its purpose for annual mosaics composed of scenes acquired in different months, and cited Ulfa et al. (2023) only to support its beneficial effect on subsequent image recognition.
  
  Comment 8 - Bit depth after normalization
  Reviewer comment
  
  Clarify the bit depth of imagery supplied to the model after the normalization.
  
  Response
  Thank you for this comment. We clarified in Section 4.1 that, after Min-Max normalization, the image data were supplied to the model as 32-bit floating-point tensors (float32) with values in the [0, 1] range.
  Changes made in the manuscript
  Section 4.1: specified that the normalized [0, 1] image data were supplied to the model as 32-bit floating-point (float32) tensors.
  
  Comment 9 - DEM resampling to image grids
  Reviewer comment
  
  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.
  
  Response
  Thank you for this clarification. We revised Section 4.1 to remove the statement that DEM resampling achieved "pixel-level registration." The text now states that bilinear interpolation was used to resample the ancillary DEM to the same pixel grids as the Landsat 8 (15 m) and Sentinel-2 (10 m) imagery. This wording does not imply that possible coregistration offsets were eliminated.
  Changes made in the manuscript
  Section 4.1: replaced "pixel-level registration" with a statement that the ancillary DEM was resampled to the same pixel grids as the corresponding optical imagery.
  
  Comment 10 - High-frequency noise
  Reviewer comment
  
  Elaborate on what 'high frequency noise' is expected to be present in the glacial lake environments.
  
  Response
  Thank you for this comment. We revised Section 4.2 to specify that high-frequency noise consists of small-scale, rapidly varying textures produced by fragmented snow and ice, exposed rocks and moraine debris, sharp terrain-shadow boundaries, and mixed land-water pixels along lake margins.
  Changes made in the manuscript
  Section 4.2: specified the principal environmental sources of high-frequency noise in glacial lake environments.
  
  Comment 11 - Figure 2 workflow
  Reviewer comment
  
  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.
  
  Response
  Thank you for this comment. We redrew Figure 2 to match the actual workflow. The radiometric-calibration and atmospheric-correction steps were removed because the inputs are surface-reflectance products. The erroneous 'STREM DEM' label was replaced with the concise source labels 'NASADEM (30 m)' and 'ASF RTC ancillary DEM (12.5 m grid)'. The undefined 'multidimensional feature dataset' was replaced by sensor-specific feature stacks comprising the selected optical bands, NDWI, and slope. The technical-validation portion was simplified into clearly labelled illustrative panels, and the figure text was enlarged for full-column readability.
  Changes made in the manuscript
  Figure 2 and its caption: removed preprocessing steps already completed in the supplied surface-reflectance products, corrected and shortened the DEM labels, defined the sensor-specific feature stacks, and simplified the validation and release workflow.
  
  Comment 12 - Training imagery dates and geographic coverage
  Reviewer comment
  
  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.
  
  Response
  Thank you for this comment. We expanded Section 4.3 to clarify that all training imagery was selected from within the SETP study area. The 30 m Landsat 5, 15 m pan-sharpened Landsat 8, and 10 m Sentinel-2 training datasets used the annual mosaics corresponding to the 1991, 2013, and 2016 SETP_GLI layers. Their acquisition dates matched the corresponding published annual mosaics and were primarily between September and November. Where late-season coverage was insufficient, within-year fallback imagery documented in Supplementary Table S1 was used. We also clarified that the external published inventory was used only as a spatial reference and that the ground-truth lake boundaries were manually delineated on the contemporaneous imagery.
  Changes made in the manuscript
  Section 4.3: added the geographic coverage, primary acquisition window, documented fallback imagery, and relationship between the training mosaics and the external reference inventory.
  
  Comment 13 - Training lake-polygon and chip counts
  Reviewer comment
  
  Clarify how many lakes and chips were used in the training and validation datasets.
  
  Response
  Thank you for this comment. Section 4.3 now reports the numbers of lake polygons and source image chips used for all three training datasets. The 30 m, 15 m, and 10 m source annotations contained 1,619, 6,536, and 9,278 manually delineated glacial lake polygons, respectively. The corresponding source-chip totals were 8,082, 42,733, and 76,672, with training/validation/test splits of 6,465/808/809, 34,186/4,273/4,274, and 61,337/7,667/7,668, respectively. The source-chip IDs were first partitioned at an 8:1:1 ratio using a fixed random seed of 42. Random 90-degree rotations and horizontal or vertical flips were then applied only when loading the training subset; the validation and test subsets were loaded without augmentation. Thus, augmented views derived from a source chip could not enter the validation or test subsets. The current split was performed at the source-chip level rather than using lake-object-level or spatial-block holdouts.
  Changes made in the manuscript
  Section 4.3: reported the annotated lake-polygon totals, source-chip totals, source-chip split, training-only augmentation procedure, and training/validation/test chip counts for the 30 m, 15 m, and 10 m datasets.
  
  Comment 14 - Learning rate and training epochs
  Reviewer comment
  
  '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.
  
  Response
  Thank you for this comment. We corrected and clarified the training description in Section 4.3. AdamW used an initial learning rate of 1 × 10−4. Thirty epochs was the common upper limit used to maintain a comparable optimization budget, rather than an automatic stopping point. Validation IoU was evaluated after every epoch, and the checkpoint with the highest validation IoU within the 30-epoch window was retained for inference; the selected model was therefore not necessarily the final-epoch checkpoint.
  Changes made in the manuscript
  Section 4.3: specified the initial learning rate, the common 30-epoch upper limit, validation-IoU monitoring, and retention of the best validation checkpoint.
  
  Comment 15 - River-related false positives
  Reviewer comment
  
  ‘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.
  
  Response
  Thank you for this comment. We agree that substantial rivers are generally distinguishable from glacial lakes when their complete morphology and upstream–downstream continuity are visible. However, GLA-RCNN was applied to 256 × 256 image chips, and instance classification was further based on local region proposals; river channels were therefore sometimes represented only by partial segments. Narrow channels, wide reaches, river bends, and fragmented braided-channel water surfaces could locally appear as isolated water bodies. In addition, NDWI enhances open water in both rivers and lakes and does not distinguish their geomorphic types. Consequently, river segments with strong water responses could be classified as glacial lakes where channel continuity was unavailable within the local input. We revised Sections 4.4 and 5.1, replaced ‘elongated river valleys’ with ‘river segments’, and added representative narrow- and wide-river false-positive examples in the new main-text Figure 6.
  Changes made in the manuscript
  Sections 4.4 and 5.1; new main-text Figure 6: clarified the causes of river-related false positives and added narrow- and wide-river examples showing the full-scene RGB imagery, 256 × 256 model chips, NDWI, and initial model outputs.
  
  Comment 16 - Quantification of manual corrections by correction type
  Reviewer comment
  
  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).
  
  Response
  Thank you for this comment. We clarified that the four items in Section 4.4 are overlapping correction types applied as required within a single manual-verification stage rather than sequential processing stages; an individual polygon could receive more than one correction type. Each corrected polygon was therefore counted once. Across all 36 annual layers, 7,660 of 174,777 polygons (4.38%) were manually adjusted. The proportions were 8.82% (5,464/61,953) for 1990–2012, 3.78% (807/21,357) for 2013–2015, and 1.52% (1,389/91,467) for 2016–2025. The retained production archive contains only the final annual geometries and does not include correction-type logs or paired pre- and post-correction boundaries. We therefore report the reproducible overall and period-specific counts and percentages, but cannot provide reliable correction-type counts or area-change magnitudes. These statistics describe the current native-resolution inventory and will be recalculated following the planned 1990–2012 reprocessing described in our response to Comment 24.
  Changes made in the manuscript
  Sections 4.4 and 5.2: clarified that the four correction types were applied within one manual-verification stage and could overlap, defined the polygon-level counting rule, and reported the overall and period-specific numbers and percentages of manually adjusted polygons.
  
  Comment 17 - Validation sample size and sampling design
  Reviewer comment
  
  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.
  
  Response
  Thank you for this comment. We clarified that the 200 sample points referred specifically to the supplementary Google Earth visual inspection, rather than to the quantitative polygon-based validation. The quantitative assessment used 1,498 manual-reference polygons for 1991, 6,531 for 2013, and 9,273 for 2016 (17,302 in total), representing the 30, 15, and 10 m sensor-processing regimes. The Google Earth cases were used to compare water-body presence, glacial-lake eligibility, and shoreline correspondence under challenging conditions; precision, recall, F1, and IoU were calculated from polygon overlays rather than from these qualitative cases.
  Changes made in the manuscript
  Sections 4.5 and 5.2: distinguished the 200-point Google Earth visual inspection from the quantitative polygon-based validation and reported the 17,302 manual-reference polygons used for geometric assessment.
  
  Comment 18 - DEM-derived reference elevation
  Reviewer comment
  
  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.
  
  Response
  Thank you for this comment. We acknowledge that glacier retreat and terrain changes since DEM acquisition can introduce errors in DEM-derived elevation attributes and affect lake delineation through the slope input. A temporally consistent DEM series covering the full study area and the entire 1990-2025 period was not available for this study, so static DEM-derived terrain information was used. We have discussed these limitations in Section 8.4 and clarified that the elevation attributes represent DEM-derived reference elevations rather than contemporaneous lake-surface elevations.
  Changes made in the manuscript
  Table 1 and Sections 3.1, 4.5, 8.4, and Data Records: clarified the DEM-derived elevation processing, including resampling the 12.5 m output grid to 10 m, and added the limitation that no temporally consistent DEM series was available for the full 1990-2025 study period.
  
  Comment 19 - CBAM ablation and shadow-area attribution
  Reviewer comment
  
  This statement is not supported by the results as the model was not tested without CBAM to quantify the improvements in shadow areas.
  
  Response
  Thank you for this comment. We conducted a controlled 2 × 2 ablation experiment on the 30 m dataset, with CBAM independently enabled or disabled in the backbone and FPN while keeping the data split, training procedure, and evaluation metrics unchanged. On the independent test set, the complete configuration achieved an F1 score of 0.9167 and an IoU of 0.8462, compared with 0.9074 and 0.8305 for the no-CBAM baseline, corresponding to increases of 0.0093 and 0.0157, respectively. The FPN-only variant achieved an F1 score of 0.9127 and an IoU of 0.8420, whereas the backbone-only variant achieved 0.9104 and 0.8355. On the validation set, the complete configuration achieved the highest F1 score (0.9236), while the FPN-only variant achieved the highest IoU (0.8599). Thus, both CBAM placements contributed to overall performance, with a larger stand-alone gain from FPN attention and the best independent-test performance from their joint configuration. Because each variant was evaluated in a single run, we do not describe these differences as statistically significant. Figure 5 provides a qualitative comparison in shadow-affected scenes, where the complete GLA-RCNN more closely reproduces the ground-truth lake masks than the comparison models. We therefore use the ablation as quantitative evidence of the modest overall CBAM contribution, whereas Figure 5 provides only qualitative evidence that the complete model can better reproduce lake boundaries in selected shadow-affected scenes; the ablation is not presented as a shadow-specific quantitative test.
  Changes made in the manuscript
  Sections 4.3, 5.1, 8.1, and Conclusions, Figure 5, Tables 3 and 7: updated the controlled CBAM ablation with results consistent with the Table 2 evaluation, reported the modest overall gains from both CBAM placements, identified the larger stand-alone gain from FPN attention and the best independent-test performance of the joint configuration, and retained the contribution to shadow-area recognition based on the qualitative comparison in Figure 5.
  
  Comment 20 - Table 2 model-comparison claims
  Reviewer comment
  
  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.’
  
  Response
  Thank you for this comment. We agree that the previous wording overstated the differences in Table 2. GLA-RCNN achieved the highest F1 score (0.9167), but the differences relative to Mask R-CNN and U-Net were only 0.0093 and 0.0144, respectively; the difference relative to DeepLab V3 was 0.0691. We therefore removed 'significantly surpasses' and now report these differences objectively. We also separated the interpretation of Table 2, which compares complete models, from the controlled CBAM ablation in Table 3. Under the same data split and evaluation procedure, the complete configuration in Table 3 reproduces the GLA-RCNN test F1 and IoU in Table 2 (0.9167 and 0.8462), while the no-CBAM baseline has the same reported test metrics as the Mask R-CNN row (0.9074 and 0.8305). The measured contribution of the complete CBAM configuration is therefore an F1 increase of 0.0093 and an IoU increase of 0.0157. Neither comparison is presented as evidence of statistical significance. We interpret this as a controlled comparison under a common optimization budget, rather than as an exhaustive comparison of individually optimized architectures.
  Changes made in the manuscript
  Sections 4.3 and 5.1 and Tables 2 and 3: retained the objective F1-score comparison among complete models, removed the unsupported claim of significant superiority, aligned the controlled ablation with the Table 2 test-set evaluation, and reported the measured CBAM contribution without claiming statistical significance.
  
  Comment 21 - Linear edges in Figure 8f
  Reviewer comment
  
  Figure 8f. There are linear misclassification edges on both glacial lakes shown. Discuss the potential origins and check other locations for issues.
  
  Response
  Thank you for pointing this out. The two examples in revised Fig. 9f (Fig. 8f in the submitted manuscript) were derived from 30 m Landsat imagery. Small protrusions and indentations along moraine-dammed and rocky lake margins are represented by only a few mixed shoreline pixels and are therefore generalized in the extracted outlines; comparison with high-resolution Google Earth imagery makes these differences more apparent. We also inspected the annual lake layers and found no comparable pattern in the inspected annual layers. This explanation has been added to Section 5.2.
  Changes made in the manuscript
  Section 5.2: added a concise explanation of the locally linear shoreline segments in revised Fig. 9f and reported the inspection of the annual lake layers.
  
  Comment 22 - Observation window and annual image mosaics
  Reviewer comment
  
  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.
  
  Response
  Thank you for this comment. We removed the inconsistent fixed-window statement. Section 3.2 now states that imagery was selected primarily from September to November and that the search was extended only to other suitable observations within the target year where coverage was insufficient. For the reprocessed 2013–2025 Landsat baseline, annual scenes were aligned to a common 30 m grid and composited pixel by pixel: clear pixels from the priority observation were retained, and later suitable observations filled only cloud- or invalid pixels at the same locations. Annual classifications were based on observations from the target year. Multiple observations within that year were combined to fill cloud- or invalid-value pixels where appropriate. When reliable target-year coverage was unavailable, the record was treated as unavailable or uncertain rather than being directly replaced with imagery from an adjacent year. Sentinel-2 layers were retained separately as 10 m high-resolution companion observations. Each annual classification therefore used a spatial mosaic assembled from multiple acquisitions, and the scene identifiers and dates are documented in Supplementary Table S1 and the source_img field. We will also apply target-year pixel-level supplementation and re-extraction to the 1990–2012 Landsat record to improve the continuity of observations of glacial lakes with variable extents.
  Changes made in the manuscript
  Sections 3.2, 4.3, and 5.3, together with Supplementary Table S1, specified the preferred September–November acquisition window, the 30 m Landsat 7/8/9 annual baseline, pixel-level cloud-gap supplementation, the retained 10 m Sentinel-2 companion layers, and the scene/date records used for each annual mosaic and cross-dataset interpretation.
  
  Comment 23 - F1 scores in Table 5 (Table 4 in the submitted manuscript)
  Reviewer comment
  
  F1 scores should also be added to Table 4.
  
  Response
  Thank you for this suggestion. We added the Overall F1 score to Table 5 for each comparison year, calculated as the harmonic mean of overall precision and overall recall. The resulting F1 scores are 0.4392 for 1990, 0.6376 for 2013, and 0.6711 for 2018. The corresponding metric descriptions in Sections 4.4 and 5.3 were also updated.
  Changes made in the manuscript
  Sections 4.4 and 5.3 and Table 5: added the Overall F1 metric and its values for the 1990, 2013, and 2018 cross-dataset comparisons.
  
  Comment 24 - Cross-sensor consistency and missing annual lake records
  Reviewer comment
  
  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.
  
  Response
  We sincerely thank the reviewer for this important comment. We agree that the different native resolutions of Landsat and Sentinel-2 affect individual lake boundaries, especially for small lakes and complex shorelines. The revised processing now provides a common 30 m Landsat baseline for 2013–2025 using Landsat 7, 8, and 9, while retaining the 10 m Sentinel-2 layers as a higher-resolution companion. This reduces the dependence of the temporal baseline on the former 15 m pan-sharpened Landsat phase, but it does not make the 30 m and 10 m boundaries geometrically equivalent.
  For the reprocessed 2013–2025 annual Landsat mosaics, the scenes were aligned to the same 30 m grid and combined pixel by pixel. Valid pixels from the priority observation were retained, and a subsequent observation supplied a pixel only where the higher-priority pixel was cloud-obscured or invalid; valid pixels were never overwritten. This is a cloud-gap compositing step rather than whole-scene replacement or numerical interpolation of lake boundaries. It reduces local omissions caused by clouds and Landsat 7 scan-line gaps while preserving the higher-resolution Sentinel-2 product for small-lake and shoreline detail.
  We will also extend pixel-level supplementation and re-extraction to the 1990–2012 Landsat record, using suitable observations from each target year to fill cloud-obscured or invalid pixels while retaining valid observations. This is intended to improve the continuity of observations of glacial lakes with variable extents by reducing gaps caused by incomplete image coverage. Improved image coverage is expected to reduce the need for manual corrections; the resulting correction counts and percentages will be recalculated after reprocessing and quality review.
  We have revised the Methods and Discussion to clarify that the 30 m Landsat record is the preferred temporal baseline for regional and long-term analysis, whereas Sentinel-2 is a complementary high-resolution product. The common ≥0.01 km² threshold remains a screening rule that reduces, but does not remove, sensor-dependent boundary differences. Subtle year-to-year changes in individual small lakes should therefore be interpreted cautiously.
  The revised manuscript no longer treats the former 2015–2016 transition as evidence of complete cross-sensor equivalence. The remaining numerical differences are discussed as observation and boundary uncertainties, and the 30 m Landsat baseline is used to reduce the resolution-related discontinuity in the temporal record.
  The planned extension to 1990–2012 will also allow the early-year gaps highlighted by the reviewer to be reassessed using additional suitable target-year observations, where available. A polygon is still not inferred when all available observations are obscured by cloud, snow or ice, terrain shadow, or insufficient lake size. The absence of a polygon in a given annual layer therefore cannot be interpreted as confirmed physical disappearance; the current release does not provide a lake-year observation-status flag, so the cause of each gap must be checked against the annual imagery. Missing annual boundaries were not inferred directly from adjacent-year imagery; such years were treated as unavailable or uncertain observations.
  Changes made in the manuscript
  Methods Sections 3.1, 3.2, 4.1 and 4.3, Discussion Sections 8.2 and 8.4, Data Records, Usage Notes, and Conclusions: added the 30 m Landsat 7/8/9 temporal baseline, described pixel-level cloud-gap compositing, retained Sentinel-2 as a 10 m high-resolution companion, and clarified that remaining gaps and cross-sensor boundary differences require cautious interpretation. Target-year gaps are treated as unavailable or uncertain observations rather than being filled from adjacent years.
  
  Comment 25 - Unsupported climate attribution
  Reviewer comment
  
  ‘This result indicates a sustained regional cryospheric and hydrological response to climate warming.’ No climate data are presented to support this statement.
  
  Response
  Thank you for this comment. We agree that the previous statement attributed the observed glacial-lake changes to climate warming without presenting a climate-data analysis. We therefore deleted this sentence from Section 7. The revised Results section now reports only the observed changes in lake number and area and the limitations of the cross-sensor comparison, without inferring a climatic cause.
  Changes made in the manuscript
  Section 7: removed the unsupported statement attributing the observed glacial-lake changes to climate warming.
  
  Comment 26 - Table 7 column order (Table 6 in the submitted manuscript)
  Reviewer comment
  
  Table 6 requires restructuring. ‘This study’ and ‘Key improvements’ are separated by two other datasets, making the comparisons confusing.
  
  Response
  Thank you for this comment. We agree that the previous column order in Table 6 of the submitted manuscript separated ‘This Study’ from ‘Key Improvements’ by the Wang et al. (2020) and Chen et al. (2021) columns, making the comparison difficult to follow. We restructured Table 7 to use the order ‘Comparison Dimension | Wang et al. (2020) | Chen et al. (2021) | This Study | Key Improvements’. The information for this study is now directly adjacent to its corresponding improvements, while all comparison dimensions and values remain unchanged.
  Changes made in the manuscript
  Table 7: reordered the comparison columns so that ‘This Study’ is immediately adjacent to ‘Key Improvements’; all table values were retained.
  
  Technical corrections Technical correction 1 - Definition of GLA-RCNN
  Reviewer comment
  
  L13. Define ‘GLA-RCNN’.
  
  Response
  Thank you. We defined GLA-RCNN in Section 4.2, where the model is introduced, as the Glacial Lake Attention-enhanced Region-based Convolutional Neural Network (GLA-RCNN). The Abstract now uses only the acronym.
  Changes made in the manuscript
  Section 4.2: expanded GLA-RCNN at its first occurrence in the main text; removed the expanded form from the Abstract.
  
  Technical correction 2 - DEM terminology and references
  Reviewer comment
  
  L145. Change the DEM and correct references to ‘ALOS DEM’ or similar.
  
  Response
  Thank you. We corrected the DEM terminology and corresponding references throughout the manuscript. The second topographic layer is now described as the SRTMGL1-derived ancillary DEM distributed with the ASF ALOS PALSAR RTC product, with a 30 m source resolution and a 12.5 m output grid, rather than as an independent 12.5 m ALOS DEM. The unrelated ALOS PRISM DEM reference was removed, and the ASF RTC product citation was added.
  Changes made in the manuscript
  Table 1 and Sections 3.1, 4.1, 4.5, Data Records, and References: standardized the DEM terminology, source resolution, output-grid spacing, provider, and citation.
  
  Technical correction 3 - Sensor-specific terminology
  Reviewer comment
  
  L151. ‘sensor-specific’ not ‘adaptive’.
  
  Response
  Thank you. We revised the wording at this location to ‘sensor-specific model inputs’, because the input channel stack is defined separately for each sensor.
  Changes made in the manuscript
  Section 4.1: changed ‘adaptive’ to ‘sensor-specific’ for the multi-sensor model inputs.
  
  Technical correction 4 - Scope of manual verification
  Reviewer comment
  
  L222. Clarify ‘full manual verification’. Were all lake polygons in the full dataset checked?
  
  Response
  Thank you. We removed the term ‘full manual verification’ because it could imply independent polygon-by-polygon checking of the entire inventory. Section 4.4 now describes this step as ‘manual verification and visual interpretation’: the 36 annual extraction layers were reviewed against the corresponding source imagery, and polygons requiring correction were manually adjusted.
  Changes made in the manuscript
  Section 4.4: removed ‘full manual verification’ and described the annual-layer visual review and correction procedure.
  
  Technical correction 5 - Numerical precision
  Reviewer comment
  
  L330. 41,916.3 m². Check whether the reported precision and number of significant figures are appropriate.
  
  Response
  Thank you. We rounded all three RMSE values to three significant figures: 3.97 × 10⁴ m² for 1991, 9.54 × 10³ m² for 2013, and 3.42 × 10³ m² for 2016.
  Changes made in the manuscript
  Section 5.2: rounded the three RMSE values to three significant figures.
  
  Technical correction 6 - Lake-count terminology
  Reviewer comment
  
  L474. Change ‘apparent numerical growth’ to ‘increase in lake count’ or similar.
  
  Response
  Thank you. We replaced the ambiguous expression ‘apparent numerical growth rate’ with ‘observed annual increase in lake count’ and made the corresponding wording consistent in the Conclusions.
  Changes made in the manuscript
  Sections 7 and 9: revised the lake-count terminology in the Results and Conclusions.
  
  Technical correction 7 - Mathematical-level terminology
  Reviewer comment
  
  L477. Replace/define ‘mathematical level’.
  
  Response
  Thank you. We replaced the undefined phrase ‘at the mathematical level’ with ‘reduced boundary-related area uncertainty’, which directly identifies the reported effect.
  Changes made in the manuscript
  Section 7: replaced ‘at the mathematical level’ with ‘reduced boundary-related area uncertainty’.
  
  Technical correction 8 - Small-lake terminology
  Reviewer comment
  
  L493. ‘minute glacial lakes’. Make terminology consistent and define area thresholds.
  
  Response
  Thank you. As detailed in our response to Comment 1, ‘minute glacial lakes’ was replaced with ‘small glacial lakes’. This term is now used consistently throughout the manuscript and is operationally defined as glacial lakes with an area <0.01 km².
  Changes made in the manuscript
  Introduction and Section 7: standardized the term as ‘small glacial lakes’ and linked it to the <0.01 km² operational threshold.
  
  Technical correction 9 - Removal of ‘target’
  Reviewer comment
  
  L595. Remove ‘target’.
  
  Response
  Thank you. We removed ‘target’ from this sentence. The revised wording states that ‘the recall reached 78.16%’.
  Changes made in the manuscript
  Conclusions: changed ‘the target recall reached 77.69%’ to ‘the recall reached 78.16%’.
  
  Citation: https://doi.org/10.5194/essd-2026-452-AC2
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
- AC1: 'Reply on RC2', hao li, 02 Oct 2026
  
  RESPONSE TO REVIEWER 2
  SETP_GLI: An annual 10–30 m glacial lake inventory for the southeastern Tibetan Plateau from 1990 to 2025
  
  We thank the reviewer for the careful assessment of the GLA-RCNN architecture, training data, validation design, sensor transition, uncertainty treatment, and presentation. We have made targeted textual revisions in the manuscript based on the current dataset and processing workflow. The revised text clarifies the model configuration, contemporaneous label generation, the 30 m Landsat temporal baseline, pixel-level cloud-gap compositing, the retained 10 m Sentinel-2 companion layers, validation limitations, image-selection records, and uncertainty sources. Figure-file updates are identified in the relevant change summaries. The response also distinguishes the completed 2013–2025 Landsat 7/8/9 processing from the planned target-year pixel-level supplementation and re-extraction of the 1990–2012 Landsat record. This work is expected to improve continuity for variable glacial lakes and reduce the need for manual corrections; the corresponding statistics will be recalculated after reprocessing and quality review.
  Reviewer comment
  GLA-RCNN Architecture Description is unclear. Please provide a detailed architecture diagram showing where CBAM modules are inserted, specify the reduction ratio used in channel attention, and give the spatial-attention kernel size.
  Response
  We clarified the architecture in Section 4.2. CBAM is inserted after ResNet-152 Conv4_x and Conv5_x, and after the FPN outputs P2–P5 and the pooled P6 level. The channel-attention reduction ratio is r = 16 and the spatial-attention convolution uses a 7 × 7 kernel. The attention maps sequentially reweight each feature map by elementwise multiplication. The architecture schematic will be updated accordingly in the final figure set.
  Changes made in the manuscript
  ·         Section 4.2: added the CBAM insertion points, reduction ratio, spatial kernel size, and multiplicative feature-reweighting description. Figure 3: schematic clarification is reserved for the final figure replacement.
  
  Reviewer comment
  The methodology for the 1991 training node is ambiguous because Wang et al. (2020) provides inventories for 1990 and 2018. Clarify how contemporaneous imagery was used for 1991 digitization, who performed the digitization, the quality-control protocol, and inter-operator consistency metrics.
  Response
  Thank you for this comment. The 1991 training labels were manually delineated from contemporaneous 1991 imagery within the study area. The public inventory of Wang et al. (2020) was used to locate candidate lakes, rather than to supply their 1991 boundaries. One operator examined the candidate lakes individually against the 1991 imagery and delineated their boundaries according to the visible lake extent at that time. Thus, the labels were not copied directly from the 1990 or 2018 inventory. The resulting labels were then reviewed by two other operators to check lake identification and boundary correspondence with the same imagery. Quality control therefore consisted of individual image-based delineation followed by review by two additional operators. A separate quantitative inter-operator agreement statistic was not recorded, and we do not claim a quantified level of agreement.
  Changes made in the manuscript
  ·         Section 4.3: clarified the source and timing of the 1991 imagery, the role of Wang et al. (2020), and the three-person cross-check procedure. Section 8.4: noted that inter-operator agreement was qualitatively reviewed but not quantified.
  
  Reviewer comment
  Explain whether separate models were used for 1990–2012, 2013–2015, and 2016–2025, how predictions were harmonized, and how performance differences across resolutions were addressed.
  Response
  Thank you for this comment. The three resolutions did not correspond to three independently designed models. A common GLA-RCNN architecture was used, with adjustments to the sensor-specific input configuration, including the available spectral bands and input-channel number. Resolution-dependent performance was evaluated separately for the 30, 15, and 10 m inputs, as reported in Table 4.
  To improve long-term temporal consistency, we reprocessed the 2013–2025 observations using Landsat 7, 8, and 9 imagery on a common 30 m grid, extending the 30 m Landsat record beyond 2012. We will also apply target-year pixel-level supplementation and re-extraction to the 1990–2012 Landsat record to improve the continuity of observations of glacial lakes with variable extents. The 15 m Landsat results for 2013–2015 and the 10 m Sentinel-2 results for 2016–2025 were retained as complementary higher-resolution datasets rather than directly concatenated with the 30 m record. These overlapping products preserve detailed information on small lakes and complex shorelines and support cross-checking of lake presence and boundary delineation.
  The revised framework therefore distinguishes the 30 m temporal baseline from the higher-resolution companion products. Maintaining a common mapping resolution reduces resolution-driven discontinuities, although differences in spectral response, acquisition conditions, and boundary detectability remain. The common ≥0.01 km² analysis threshold provides additional screening but is not interpreted as eliminating these residual differences.
  Changes made in the manuscript
  ·         Section 4.3: clarified the distinction between the 30 m Landsat temporal baseline, the retained 10 m Sentinel-2 companion layers, the original 15 m validation benchmark, and the pixel-level compositing rule; native-grid outputs are not treated as geometrically equivalent boundaries.
  
  Reviewer comment
  Hi-MAG and Wang et al. (2020) assisted label generation and were also used for accuracy assessment. This dual use can introduce circularity bias and penalize newly detected small lakes absent from older 30 m references.
  Response
  We agree that the internal held-out chip metrics and the external inventory comparisons answer different questions. During label preparation, Hi-MAG and Wang et al. (2020) were used only to locate and cross-check candidate lake positions; the final boundaries were delineated from contemporaneous imagery rather than copied from the reference polygons. The labels also included observable lakes absent from the public inventories, including lakes that were not necessarily micro-lakes. The 8:1:1 chip split provides held-out image-chip evaluation, but the public inventories also informed candidate selection, so strict independence cannot be claimed for the external comparisons. Shared reference information may increase agreement for common lakes, whereas valid lakes absent from older 30 m inventories may be counted as unmatched detections and reduce apparent precision. Because no independent high-resolution inventory covering the full SETP and the complete study period was available, we retain these comparisons as consistency assessments and explicitly acknowledge this limitation in the manuscript. Accordingly, these external comparisons are described as cross-inventory consistency assessments, not as independent estimates of absolute mapping accuracy.
  Changes made in the manuscript
  ·         Sections 4.5 and 8.4: added an explicit limitation concerning reference overlap, circularity, and the resolution-related penalty applied to newly detected small lakes.
  
  Reviewer comment
  Google Earth imagery has heterogeneous dates and resolutions. Provide dates, resolution, quality information, and discuss temporal misalignment with Landsat and Sentinel-2 acquisitions.
  Response
  Thank you for this comment. We agree that the Google Earth imagery used in the small-lake checks has heterogeneous acquisition dates and effective spatial resolutions across the SETP. Its primary purpose was to verify whether lakes smaller than 0.01 km² were detected and whether their general spatial locations and shoreline shapes were plausible, rather than to provide temporally exact, pixel-level ground truth. Exact nominal spatial-resolution metadata were not consistently available for all Google Earth scenes, so unsupported resolution values were not assigned; the imagery is treated as a qualitative high-resolution reference rather than temporally exact pixel-level ground truth.
  Changes made in the manuscript
  ·         Section 5.2: added the temporal and spatial heterogeneity caveat for Google Earth reference imagery and its implications for small-lake checks. Figure 9/10: metadata and legibility updates are reserved for final figure replacement.
  
  Reviewer comment
  The boundary-pixel method does not account for georeferencing errors, seasonal water-level fluctuations, or shadow-induced boundary shifts, especially in early Landsat scenes. Additional uncertainty sources should be discussed.
  Response
  Thank you for this comment. The boundary-pixel method of Hanshaw and Bookhagen (2014) was used specifically to quantify the area uncertainty caused by mixed pixels and the pixel-scale irregularity of lake boundaries. It does not represent a complete uncertainty budget.
  We agree that additional sources of uncertainty, including residual georeferencing errors in early Landsat imagery, seasonal water-level fluctuations, snow and ice cover, terrain-shadow effects, and debris-covered shorelines, were not independently quantified. We have therefore clarified the scope of the boundary-pixel estimate and expanded Section 8.4 to discuss these additional uncertainty sources and their potential effects on lake boundaries and areas. The reported values are therefore termed boundary-related area uncertainty, not total mapping uncertainty.
  Changes made in the manuscript
  ·         Section 8.4: expanded the limitations of the boundary-pixel uncertainty estimate to include georegistration, seasonal, shadow, snow/ice, and debris-related effects.
  
  Reviewer comment
  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.
  Response
  Thank you for this detailed comment. We revised the image-selection and temporal-consistency procedures and clarified the sensor-transition treatment.
  The preferred observation window is September–November, when seasonal snow and cloud effects are generally reduced and lake boundaries are comparatively stable. Other observations within the same target year, including adjacent-month acquisitions, were used only when suitable September–November coverage was unavailable. For the reprocessed 2013–2025 Landsat baseline, clear pixels from the priority observation were retained, and subsequent suitable observations were used only to fill cloud- or invalid-value gaps at the same locations; valid pixels were never overwritten. This pixel-level compositing reduces seasonal discontinuities caused by image availability, although it cannot completely remove effects from seasonal water-level changes, freezing, or snow cover. We expanded the supplementary annual image records to report acquisition dates and months, sensors, scene or tile information, cloud or data-gap status, and the source of local gap filling, together with the monthly observation distribution needed to assess potential seasonal bias in mapped lake area. A separate numerical sensitivity test quantifying area bias between preferred- and fallback-month observations was not performed in this revision; this remains a limitation when interpreting individual lake areas. No adjacent-year imagery was used as a direct substitute for a target-year boundary. If suitable target-year coverage remained unavailable, the annual observation was treated as unavailable or uncertain. This approach reduces the possibility that cross-year filling could obscure genuine interannual changes.
  We also clarified that the 0.001, 0.002025, and 0.0054 km² thresholds are sensor-specific minimum mapping units for 10, 15, and 30 m imagery, respectively. They are practical screening thresholds and do not imply equivalent detection capability across the full time series. The uniform ≥0.01 km² threshold is applied only for aggregate trend screening and reduces, but does not eliminate, cross-sensor differences.
  To improve temporal comparability, we reprocessed the 2013–2025 annual baseline using Landsat 7, Landsat 8, and Landsat 9 imagery on a common 30 m grid. Thus, the 2014–2015 and 2016 baseline layers were generated using the same 30 m Landsat processing framework, with pixel-level supplementation where necessary; Landsat 8 was not treated as the sole source for all years. The 15 m Landsat 8 results for 2013–2015 and the 10 m Sentinel-2 results for 2016–2025 were retained as higher-resolution companion products. Sentinel-2 therefore supplements the 30 m baseline after 2016 rather than replacing it in the primary long-term record.
  We will also extend pixel-level supplementation and re-extraction to the 1990–2012 Landsat record, using suitable observations from each target year to fill cloud-obscured or invalid pixels while retaining valid observations. This is intended to improve the continuity of observations of glacial lakes with variable extents by reducing gaps caused by incomplete image coverage. Improved image coverage is expected to reduce the need for manual corrections; the resulting correction counts and percentages will be recalculated after reprocessing and quality review. The manual-correction statistics currently reported refer to the existing native-resolution inventory.
  Changes made in the manuscript
  ·         Sections 3.2 and 4.3: clarified the preferred September–November acquisition window, adjacent-month gap handling, annual multi-acquisition mosaics, pixel-level Landsat cloud-gap supplementation, and the common 30 m Landsat 7/8/9 baseline for 2013–2025. Section 5.3: clarified the observation window used when interpreting cross-dataset consistency metrics. Section 8.2: clarified the residual seasonal and cross-sensor limitations. Supplementary Table S1: expanded the annual scene/date and source-image records used for classification, including the information needed to assess acquisition-month distribution and gap filling.
  
  Reviewer comment
  Were Mask R-CNN, U-Net, and DeepLab V3 optimized for this task? Specify learning rate, epochs, augmentation, hyperparameter tuning, and whether the comparison used equal training resources.
  Response
  The comparison models used the same 1991 source-chip split (seed 42; 8:1:1), 256 × 256 chips, normalization, AdamW, learning rate 1 × 10⁻⁴, weight decay 1 × 10⁻⁴, a maximum of 30 epochs, and best-validation-checkpoint selection. Random 90-degree rotations and horizontal or vertical flips were applied only to the training subset; validation and test subsets were loaded without augmentation. No model received additional data or a larger epoch budget, and no model-specific hyperparameter search was performed. Detection and semantic-segmentation models used their memory-compatible batch sizes (2 and 8, respectively, in the comparison configuration), which is now acknowledged as a practical implementation difference. We therefore interpret this experiment as a controlled baseline comparison under a common optimization budget, rather than as an exhaustive comparison of individually optimized architectures.
  Changes made in the manuscript
  ·         Section 4.3: added the common comparison-model training settings, checkpoint rule, and the batch-size qualification.
  
  Reviewer comment
  How were the CBAM outputs integrated with the FPN features? Provide a precise explanation and schematic.
  Response
  For each FPN output, CBAM first generated channel-attention and spatial-attention maps. The feature map was multiplied element by element by the channel-attention map and then by the spatial-attention map before being passed to the subsequent FPN or prediction operation. The revised text also identifies P2–P5 and the pooled P6 output. The schematic will be updated for readability in the final figure set.
  Changes made in the manuscript
  ·         Section 4.2: added the sequential multiplicative FPN integration description and the pooled output. Figure 3: schematic update is reserved for final figure replacement.
  
  Reviewer comment
  Clarify how variable input dimensions are handled between Sentinel-2 and Landsat and discuss potential spectral inconsistencies.
  Response
  The input configuration remains sensor-specific. The 30 m Landsat 7/8/9 inputs contain seven channels (five harmonized optical bands plus NDWI and slope), whereas the Sentinel-2 inputs contain six (four optical bands plus NDWI and slope) because a 10 m SWIR band is unavailable. Bands are normalized before stacking with NDWI and slope. The prepared image chips determine the channel count through the in_channels: auto option, and that count remains fixed during each training and inference run. The 30 m Landsat and 10 m Sentinel-2 feature spaces are therefore not assumed to be interchangeable.
  Changes made in the manuscript
  ·         Sections 4.1 and 4.2: clarified automatic channel configuration, the six- versus seven-channel inputs, normalization, and the resulting spectral non-equivalence.
  
  Reviewer comment
  The loss description is technically dense. Add a brief intuitive explanation of each loss term.
  Response
  We added a short explanation: classification term distinguishes lake instances from background, box-regression terms align candidate extents, and mask loss refines pixel-level boundaries.
  Changes made in the manuscript
  ·         Section 4.2: added the intuitive explanation immediately after the mathematical loss definition.
  
  Reviewer comment
  The Discussion repeats the Introduction. Focus it on implications of the new dataset rather than repeating its features.
  Response
  We condensed the repeated introductory description and refocused the Discussion on the interpretation of sensor-dependent trends, uncertainty, validation scope, applications, and limitations.
  Changes made in the manuscript
  ·         Section 8.1: revised the opening discussion to emphasize dataset implications and interpretation rather than repeat the Introduction.
  
  Reviewer comment
  Correct ‘STREM’ to ‘SRTM’ in Figure 2.
  Response
  The text uses the correct SRTM terminology. The figure label will be corrected in the final figure file.
  Changes made in the manuscript
  ·         Figure 2: spelling correction reserved for the final figure replacement; related body text remains consistent with SRTM.
  
  Reviewer comment
  Ensure all text can be read clearly in every panel of Figure 3.
  Response
  We will enlarge and simplify the panel text and labels in the final architecture schematic. The manuscript text now supplies the exact module locations and parameters independently of the figure.
  Changes made in the manuscript
  ·         Section 4.2: architecture details added in text. Figure 3: legibility update reserved for final figure replacement.
  
  Reviewer comment
  What do ‘raw’ and ‘MMU’ trends represent? Explain this in the caption or main text.
  Response
  We clarified the terminology. The full-inventory trend is the annual inventory after the sensor-specific mapping thresholds. The MMU trend is the subset retained after applying the uniform ≥0.01 km² analysis threshold. The latter reduces, but does not eliminate, cross-sensor differences.
  Changes made in the manuscript
  ·         Annual-trend caption and Section 7: defined the full-inventory and MMU series and qualified their use for cross-period comparison.
  
  Citation: https://doi.org/10.5194/essd-2026-452-AC1

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: 741 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
508	132	101	741	130	100

HTML: 508
PDF: 132
XML: 101
Total: 741
BibTeX: 130
EndNote: 100

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	123	44	58	225
Oct 2026	76	32	28	136

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	123	44	58	225
Oct 2026	76	32	28	136

Viewed (geographical distribution)

Total article views: 651 (including HTML, PDF, and XML) Thereof 651 with geography defined and 0 with unknown origin.

Country	#	Views	%

Cited

Latest update: 14 Oct 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.