The Greenland ice-marginal lake annual inventory series is a follow-on from the 2017 Greenland ice-marginal lake inventory, largely adopting the same classification approach, data structure and formatting . provided the first Greenland-wide ice-marginal lake inventory as a static dataset, building upon regional multi-temporal efforts, such as the southwest inventory classified from Landsat imagery .

The dataset presented consists of a series of annual inventories, mapping the presence and extent of ice-marginal lakes across Greenland (Fig. ). Ice-marginal lakes are defined as water bodies ≥0.05 km2 (based on satellite image spatial resolution), which are immediately adjacent to (and therefore in contact with) the Greenland Ice Sheet and/or the PGICs of Greenland. The annual inventory series spans the entirety of Greenland, including all terrestrial regions. Thus far, there are 8 annual inventories, covering the Sentinel satellite era from 2016 to 2023, where one inventory represents one year.

The inventory series data are distributed as polygon vector features in an open GeoPackage format (.gpkg), with coordinates provided in the WGS NSIDC Sea Ice Polar Stereographic North (EPSG:3413) projected coordinate system. File names follow the convention defined in the original 2017 Greenland ice-marginal lake inventory : <inventory-year>-<funder>-<project-acronym>-IML-f<version-number>.<file-extension>.

Each GeoPackage file contains metadata regarding the lake description, physical measurements, lake surface temperature, method/s of classification, verification and possible editing (Table ). A key piece of metadata to highlight is the lake identification number (“lake_id”, Table ), which are assigned to each classified ice-marginal lake, often consisting of multiple polygon features and/or classifications. These unique identifications are compatible across inventory years, therefore supporting time-series analysis and comparison across inventories.

Information is included regarding whether the adjacent ice margin is either the ice sheet or PGIC (“margin”, Table ). This margin information originates from the MEaSUREs (NASA Making Earth System Data Records for Use in Research Environments) GIMP (Greenland Ice Mapping Project) 15 m ice mask, where the provided GeoTiff files were merged, transformed into a vector format and then assigned as either an ice sheet or PGIC margin . In addition, each ice-marginal lake is assigned a region – north-west (NW), north (NO), north-east (NE), central-east (CE), south-east (SE), south-west (SW), and central-west (CW) (“region”, Table ). These regions and their corresponding names are based on ice sheet catchment regions from , which are used to also extend to the terrestrial periphery beyond the ice sheet. By doing so, regional trends can be identified from ice-marginal lakes with a PGIC margin as well as the ice sheet .

Lake names (“lake_name”, Table ) are assigned in instances where a name is available, with preference to West Greenlandic (Kalaallisut) placenames followed by Old Greenlandic and alternative foreign placenames. Placenames are provided by Oqaasileriffik (the Language Secretariat of Greenland) placename database (https://nunataqqi.gl/, last access: 12 November 2025), filtering placenames to those associated with lake features. The placename database is distributed with QGreenland v3.0 .

A readme file is included with the dataset that outlines the data file contents and terms of use. An additional data file is included which is a point vector GeoPackage file representing all identified lakes across the inventory series (presented in Fig. ). This includes manually identified lake locations that are not captured in the inventory series using the automated classification approaches.

Ice-marginal lakes (≧0.05 km2) were identified using three established classification methods, from Synthetic Aperture Radar (SAR) and multi-spectral imagery, and Digital Elevation Models (DEMs). This approach is based on the production of the 2017 Greenland ice-marginal lake inventory, which is summarised in Fig.  . The main progression (and therefore difference) is that the processing pipeline is now unified and operates through Google Earth Engine to conduct the heavy image processing, and filtering/post-processing conducted with open-source spatial packages in Python, namely geopandas and rasterio . The Python pipeline is deployable as a package called GrIML, which is accompanied by thorough documentation and guidelines (https://griml.readthedocs.io, last access: 12 November 2025) . This ensures a high level of reproducibility and transparency that adheres to the FAIR (Findability, Accessibility, Interoperability, and Reusability) principles .

A summer water surface temperature estimate was provided for each classified lake across inventory years. Water surface temperature estimates were derived from the Landsat 8 and Landsat 9 OLI/TIRS surface temperature data, which is a Collection 2, Level-2 science product that is part of a large Landsat re-processing effort (Table ). Surface temperature estimates were generated from descending, day-lit Landsat 8/9 acquisitions with thermal infrared band information (100 m spatial resolution) and auxiliary data (i.e. Top Of Atmosphere reflectance and brightness temperature), along with ASTER datasets (global emissivity and normalised difference vegetation index) and MODIS and VIIRS atmospheric auxiliary data (geopotential height, specific humidity and air temperature) .

Due to the lack of in situ lake surface temperature measurements in Greenland, the scheme proposed by was adopted, whereby surface temperature values (LSTland) were corrected to water surface temperature (LSTwater) using the following calibration: 1LSTscaled=LSTland×0.00341802+149.02LSTwater=(0.806×LSTscaled+54.37)-273.15 Where LSTscaled is the applied scale factor for computing temperature in Kelvin (K) units, and LSTwater is the calibrated water surface temperature in °C . This calibration has previously shown strong correlations against in situ measurements (average RMSE =2.8 °C and R2=0.8) from 38 lakes in Poland, highlighting accurate estimates through a simple linear calibration .

A summer average water surface temperature estimate was derived using this approach, calculating an average, maximum and minimum water surface temperature value for each lake extent over each inventory year, along with the standard deviation. Scenes were filtered by a maximum cloud cover of 20 %, with acquisitions limited to the month of August to reduce the probability of ice-covered lake conditions. Lake extents were cropped by a border pixel (i.e. 100 m) to reduce the impact of edge effects. All unrealistic estimates below freezing (i.e. <0 °C) were removed, therefore reducing the influence of icebergs and ice cover on the water temperature estimates.

In total, the dataset identifies 2918 automatically delineated ice-marginal lakes across all inventory years (2016–2023) (Fig. ). Of these lakes, 2054 share a margin with the ice sheet whilst 864 are in contact with PGICs (Fig. ). The SW region holds the most lakes compared to other regions, with 786 classifications (640 classified at the ice sheet margin and 146 at the PGIC margins). A high abundance of PGIC lakes is found in the NO region, with 278 lakes, compared to only 37 PGIC lakes in SE. This reflects the presence of more PGICs in northern Greenland, compared to the greater ice sheet cover in the southeast.

Small fluctuations in the abundance of lakes are evident, fluctuating between 1963 (inventory year 2021) and 1827 (inventory year 2022) lakes (Fig. ). The largest variation in lake abundance at the ice sheet is evident at the SW margin, with a fluctuating range of 48 lakes (8 %) between 567 (2018, 2023) and 615 lakes (2021) (Fig. a). The CW and SE margin experienced the least variability, only varying by 16 lakes and 15 lakes, respectively. Changes in lake abundance at the ice sheet margin do not follow any spatial or temporal trends, with fluctuations unconnected to inventory year or margin region.

Lakes at the margins of periphery ice caps and glaciers vary between 723 (inventory year 2022) and 806 (inventory year 2019), with an average range of 11 lakes at the margins of periphery ice caps and glaciers (compared to an average range of 26 lakes at the ice sheet margin) (Fig. b). The largest fluctuations in lake abundance are seen in the NO and NE regions, fluctuating by 22 (8 %) and 29 lakes (14 %), respectively. This is linked to the higher number of lakes in these regions, which is supported by the smallest fluctuations evident in the regions where fewer lakes are present (i.e. NW, SE, CE and CW).

The largest lake in the inventory is Romer Sø (80°59′54′′ N, 19°09′21′′ W), located in northeast Greenland, with a total area of 126.86 km2 (Fig. ). The average lake size is 1.29 km2, and the median lake size is 0.27 km2 with 2395 lakes between 0.05–1.00 km2 (82 %). Only 59 lakes in the inventory series have a total area above 10 km2 (2 %). When lake surface extent is subcategorised by region then the NE region holds the largest lakes with a median lake area of 0.34 km2 (mean: 1.63 km2). The largest lakes sharing a margin with the ice sheet reside in the NO and NE regions, with a median lake size of 0.43 (mean: 1.78 km2) and 0.40 (mean: 1.43 km2), respectively. The largest PGIC lakes are located in the NW region with a median lake size of 0.34 km2 (mean: 0.52 km2).

The inventory series also holds information on the change in lake area over time, by comparing corresponding lake extents from one of the direct classification methods (i.e. from SAR and/or multi-spectral imagery). Lake areas are first dissolved across the SAR and multi-spectral classifications, and then statistics are derived by comparing lake area through the inventory years. The total (summed) lake area for each region is presented for the purpose of assessing changes in lake area across the inventory series. In addition, median lake area statistics are presented instead of averages in order to lessen the influence of outliers, especially those originating from large lakes (Fig. ).

The total (summed) lake area is categorised by ice margin type (ice sheet and PGIC) in Fig. , with subsequent time steps revealing absolute area change across each region. No inherent trends in total lake area between 2016 and 2023 are evident across all regions, such as uniform change through time. The lack of a strong trend could reflect the high climatic variability across Greenland given its large latitudinal range and exposure to variable atmospheric and oceanic climatic forcing. It could also be related to variability in lake contact with the ice margin over time, and/or varying iceberg presence which is not considered in the SAR and multi-spectral classification approaches.

Regional inter-comparison provides valuable insights into terrestrial meltwater storage. At the ice sheet margin, the SW region makes up the largest component of total lake area, fluctuating between 327.15 km2 (2018) and 733.89 km2 (2017) (Fig. a). Total lake area across PGIC margins is considerably smaller, with regional lake area varying between 2.19 km2 (CW region, 2022) and 73.45 km2 (NO region, 2019) (Fig. b). Two exceptions are the NE and SW regions, where total lake area is notably higher and likely associated with the higher number of PGIC lakes in these regions.

Change in median lake area over the ice sheet margin is relatively consistent across the inventory series, with the median across all regions fluctuating between 0.14 km2 (SE region, 2018) and 0.38 km2 (CW region, 2020) (Fig. a). Outliers are evident at the CE and NE regions in 2022, where the median lake area reaches peaks of 0.59 and 0.44 km2, respectively. As a result, the CE and NE regions demonstrate most variability in median lake area compared to others. PGIC lakes have comparative median lake areas, fluctuating similarly between 0.13 km2 (SE, 2018) and 0.40 km2 (CW, 2022), with a high outlier in the CE region (2017, 0.67 km2) (Fig. b). PGIC lakes in the SW region have markedly higher median areas (0.24–0.37 km2) compared to other PGIC lake regions and their ice sheet region counterparts.

Overall, the rate of change in lake area could be analysed across 1696 lakes in the inventory series, representing 31 % of all mapped lakes. Of these, 83 lakes showed growth between 2016 and 2023, with an increase in area of greater than 0.05 km2yr-1, while 240 lakes experienced a decline, with a decrease in area of greater than 0.05 km2yr-1. The remaining 1373 lakes exhibited no significant change, with area variations limited to ±0.05 km2yr-1 (Fig. ). The largest rate of area change was observed in the larger lakes. The average area of lakes that expanded or contracted was 6.18 and 7.77 km2, respectively, with a total combined area of 2378.75 km2. In contrast, the average area of lakes that remained stable was much smaller, at 0.41 km2, contributing a total combined area of 560.39 km2. Therefore, while the majority of lakes experienced minimal changes in area, suggesting stability in size, they account for a much smaller proportion of the total ice-marginal lake volume compared to the larger, more dynamic lakes. It is likely that the regional statistics on lake area change are more influenced by shifts in the largest lakes than by trends across all regions and lake sizes.

The inventory series demonstrates changes to lake morphology (and the corresponding change in ice margin morphology), of which four example scenarios are presented in Fig. . A classic terminus basin retreat style is evident across many ice-marginal lake extents, as presented in Fig. a, where terminus retreat/lake expansion is marked in the central section of the glacier outlet, leaving a trailing terminus morphology at the lateral margins. Peripheral terminus retreat is highlighted in Fig. b, where terminus retreat/lake expansion is focused at the lateral margins. There are instances where the presence of a lake affects the boundary conditions of two glacier termini, as demonstrated in Fig. c, where two glaciers terminate into the same common ice-marginal lake. And finally, there are instances displayed in the inventory series where there is margin retreat/lake expansion focused around a discrete zone, such as in Fig. d where a marked embayment has formed at a particular point in the north region of the glacier terminus.

An average surface temperature estimate was derived for each inventory lake from all available Landsat 8/9 scenes acquired in the month of August for each inventory year (see Sect. ). This information is provided in the metadata of the ice-marginal lake inventory series. Examining the average lake surface temperature estimate across all lakes (i.e. the sum of all lake averages divided by the number of lakes), the average lake surface temperature fluctuates between 3.8 °C (2018) and 5.3 °C (2023) (Fig. ). The median lake surface temperature follows a similar trend, fluctuating between 3.0 °C (2018) and 4.4 °C (2023). Fluctuations year on year vary, with instances of lake temperature being lower between annual time steps (e.g. from 4.5 to 3.8 °C from 2017 to 2018), higher (e.g. from 4.8 to 5.3 °C from 2022 to 2023), and remaining consistent (e.g. 4.8 °C for 2021 to 2022). Overall, there is no evident trend in average lake surface temperature across the period. However, there is a notable anomaly in 2018 which could reflect a true climatic event but needs to be investigated further.

Average surface temperature can be examined spatially across each lake in the inventory series (Fig. a). This reveals an apparent latitudinal trend, with lakes across the northern regions (NO, NW, and NE) being cooler, on average, than those in the southern regions (SW and SE). The northern regions have a higher abundance of lakes with an average surface temperature between 0 and 4 °C, whereas lake temperature in the southern regions tends to be between 4 and 10 °C. There are visible exceptions to this spatial trend, such as ice-marginal lakes present in nunatak areas which are generally cooler because of a greater presence of ice surrounding them and/or beneath them.

When assessing lake temperature change from 2016 to 2023, lake size appears to influence the average surface temperature and the rate of temperature change across each inventory year (Fig. b). The smallest lakes (≦0.1 km2) are warmest, on average, with an average temperature of 5.45 °C and the overall average varying between 4.9 and 5.9 °C. The largest lakes (5.0–150.0 km2) are colder, with an average temperature of 3.9 °C, varying between 3.5 and 4.5 °C. Lakes with a smaller surface extent (≦0.5 km2) remained relatively consistent temperatures across each inventory year, with the average fluctuating by a maximum of 0.2 °C between 2016 and 2023 °C (Fig. b). The largest lakes in the inventory series (5.0–150.0 km2) experienced the largest temperature change between 2016 and 2023, cooling by an average of 0.5 °C.

Classification information is provided with the ice-marginal lake inventory series, so that the performance of each classification method can be evaluated (Fig. ). 14 020 of all detections in the inventory series (66 %) were classified using only one of the methods, composed largely from the DEM method (Fig. a). 5156 detections were classified using two methods (24 %), and 2264 detections were classified using all three methods (11 %). It is noted that the number of classification methods is not a measure of certainty but instead should be interpreted as a reflection of lake appearance and its adherence to the criteria of each classification method, as well as satellite data availability.

The SW region was typically where most lakes were classified with all three classification methods; across both the ice sheet margin (Fig. b) and the PGIC margins (Fig. c). This is likely because the classification methods have been extensively applied and developed in the SW region, making them particularly well-suited for use there compared to other regions e.g.,. The DEM method was heavily relied upon in the NO and NE regions where direct classification of open water was challenging as lakes were more likely to be consistently ice/snow covered, and satellite image availability from Sentinel-1 and Sentinel-2 was limited .

All ice-marginal lakes across Greenland were manually verified for each inventory year, including those that were not classified using the automated methods. All manually identified lakes are included in the dataset as point locations alongside centroid locations for those automatically classified, with metadata flags for differentiating them.

Across all inventory years, 4543 ice-marginal lakes were manually identified in total, of which 2915 (64 %) are captured by the automated classification approach. This forms an abundance error estimation of -809 (36 %), which reflects the undercounting of lakes by the automated classification (i.e. a negative bias). However, manually classified lakes include those under the size threshold (i.e. <0.05 km2) adopted in the automated classification approach. The under-estimation of ice-marginal lakes within the inventory series therefore, in part, reflects smaller lakes that are removed from the dataset automatically due to the minimum area filtering. Therefore, the error estimate reported is a conservative lower bound, which reflects the underestimation of ice-marginal lakes due to false negatives (i.e. where lakes are missed by the automated classification method).

Lakes classified with both the SAR backscatter and multi-spectral indices classification approaches were compared to assess variability in footprint size and provide an error estimation of lake extent. Classifications with the sink detection approach were excluded from this analysis, as the sink detection approach is an indirect measurement of lake extent rather than a direct classification of water. The use of independently derived classification methods has previously been adopted and offers a reasonable proxy for estimating error in lake delineations e.g.,.

Across all ice-marginal lake classifications within the inventory series (2016–2023), 3070 lakes were successfully classified with both the SAR backscatter and multi-spectral indices classification approaches. On average, there is a difference of 1.54 km2 between the two classifications, with a median difference of 0.17 km2. When all lake areas are summed and compared, the total lake area error (i.e. the total areal difference between the SAR and multi-spectral classifications) is 774.94 km2, forming a central percentage error estimate of ±5 % (Table ).

The absolute error in classified lake areas propagates according to lake size, with lakes over 5.00 km2 (339 lakes) having a total error of 405.98 km2 compared to a total error of 1.89 km2 across lakes between 0.05 and 0.10 km2 (195 lakes) (Table ). However, the percentage (i.e. relative) error varies marginally according to lake size, with the central percentage error estimate lying between ±6 % (0.05–0.10 and over 5.00 km2) and ±11 % (0.10–0.20 km2) based on the size groupings presented in Table .

Water surface temperature estimates were validated against all known and/or open-access in situ measurements of lake temperature in Greenland (Fig. ). The only continuous/long-term in situ surface measurements (i.e. ≦2 m) are from six lake records in southwest Greenland – Kangerluarsunnguup Tasia (64°07′50′′ N, 51°21′36′′ W) and Qassi-Sø (64°09′14′′ N, 51°18′27′′ W) , Russell Lake (67°13′77′′ N, 50°07′63′′ W) (courtesy of Kristian K. Kjeldsen), and three lakes as part of the Asiaq Greenland Survey hydrological monitoring programme (Qassi-Sø 2024 measurements; Qamanersuaq, 63°47′71′′ N, 50°00′50′′ W; and an unnamed lake referred to as Asiaq station 924, 64°12′99′′ N, 51°36′39′′ W).

Comparison of the 133 coinciding in situ measurements with those estimated using the remote sensing approach adopted here exhibit a strong correlation (r2=0.87), with an RMSE of 1.68 °C, suggesting that the remotely sensed temperature estimates are reliable (Fig. ). This trend appears to be consistent regardless of the time of year. An interesting cluster of data points is evident, originating from measurements taken at Qamanersuaq and Asiaq station 924 which could be related to specific lake characteristics, such as lake depth/morphology or suspended sediment concentration. An error estimation of ±1.2 °C is determined, based on the average difference from data points across all lake sites.

Ideally, a correction factor specifically for calibrating values to Greenland lakes would be adopted. However, in situ validation datasets in Greenland are sparse and the derived correction factor appears to agree well with the limited datasets available. In the future, more in situ observations would strengthen the assessment, with a possibility to derive a Greenland-specific correction scheme. In addition, the influence of ice presence is limited based on the selection of strictly summer scene acquisitions and removing temperatures below 0 °C. In future work, icebergs and ice cover could be explicitly removed before a temperature estimate is derived to reduce the influence of their presence. This is currently beyond the scope of the dataset given the challenges with classifying the presence of ice on lakes across large regions in the Arctic e.g.,.

The inventory series presented here is the first step to quantifying the terrestrial storage of meltwater, and how it changes over time, which would be highly valuable for refining estimations of the future sea level contribution of the Greenland Ice Sheet and surrounding PGICs. The dataset extends the value of a single, time-static inventory by providing a consistent, multi-year record of lake evolution. This temporal dimension enables analyses that capture lake variability and persistence through time. Tentative findings have been outlined, yet further analysis and evaluation against other datasets is needed to investigate causal links. For example, the inventory series could be used to address the drivers of change in lake area with comparison to potential influences such as meltwater flux, sedimentation rates, bedrock type, and GLOF magnitude and frequency e.g.. The inventory series could also be incorporated with mapping efforts of terrestrial lakes (i.e. no contact with the ice margin) to provide a detailed overview of dynamic and stable storage of water at the terrestrial margins of Greenland e.g. Additionally, the inventory series would be a valuable dataset for examining lacustrine terminus retreat dynamics, expanding investigations from a case study basis e.g. to a regional and/or national scale e.g..

The ice-marginal lake inventory series is applicable to climate and cryosphere research, enabling inter-annual comparison of lake change (abundance, extent and surface temperature) over time, similar to inventories for other regions such as Svalbard . Such inventories have been used to characterise ice dam types e.g.,, monitor GLOFs e.g.,, and assess lake conditions in catchments of interest e.g.,. Lake conditions could also provide insights into glacier dynamics in lacustrine settings around Greenland; for example, to investigate submarine melting in lacustrine settings and its impact on glacier retreat e.g.,, and to explore regional variability and evolution of lake surface temperatures and their influence on downstream ecology e.g.,. More widely, the lake changes documented in this inventory series would be valuable to studies of the redistribution of mass on the earth surface, affecting gravity, geodesy and lithospheric elastic response e.g.,.

Beyond scientific research, the inventory series will also be a useful resource in Greenland's assessment of infrastructure, with hydropower being the main sector that could benefit. Given Greenland's commitment to the Paris Agreement strongly suggests the expansion of current hydropower infrastructure, the ice-marginal lake inventory series could be valuable in infrastructure assessments . For example, the inventory series can be used to distinguish glacier-fed lakes from catchment-fed lakes, identify draining lakes, and other characteristics that are useful to discern viable catchment regions.

It is planned to update the ice-marginal lake inventory series annually with new inventory years, using the methodology and data sources outlined here. Further automated classification methods need to be explored and incorporated into the data production pipeline to address the under detection of ice-marginal lakes, as highlighted in the validation of automated classifications against manual detections (see Sect. ). Methodologies such as Forel-Ule color indexing (FUI) have been successfully applied to the detection of lakes in the High Arctic , and more widely in a global context , which could be suitable for applying to ice-marginal lakes in Greenland after testing their accuracy and feasibility. Another avenue to explore is the inclusion of manually delineated lake extents where lakes have not been identified with the automated classification approaches. However, this would add further labour to the manual curation of the inventory series. An alternative would be to look at implementing new automated classification methods with machine learning, using the existing lake classifications as the foundation of a training dataset. Lake classification aided by machine learning has been successfully used for supraglacial lake detection on the Greenland Ice Sheet, so the use of machine learning in ice-marginal lake detection is worth further investigation .

One of the key limitations of this work to be addressed in the future is the reliance on static data products, in particular the static ArcticDEM 2 m mosaic for classification, and the MEaSUREs GIMP static ice margin for filtering. The use of static data products in the inventory series presented here highlights the importance of high-labour, time-consuming manual dataset curation. For the DEM classification, an alternative would be SAR-derived DEMs from the TanDEM-X mission or the ArcticDEM strip data product, which are both time variant, but data coverage is lacking currently and scenes covering all Greenland may not be possible from year to year e.g.,. Another option could be to use coarser spatial resolution DEM products, such as PRODEM (500 m) , however, smaller lakes would not be identifiable. For the ice margin filtering, machine learning ice margin products show promise in being used in future editions of the inventory series, such as AutoTerm (trained with the TeamPicks dataset) . The use of dynamic ice margin datasets in the future could negate the need for generating a classification spatial buffer around the margin data and instead classify ice-marginal lakes directly from their intersection with the ice margin position.

Another opportunity for future iterations of the inventory series could be to include past years, prior to the Sentinel satellite era. However, this is limited by the open availability of SAR and multi-spectral satellite imagery at a high spatial resolution (i.e. 10 m). The addition of valuable metadata could also be explored, including information such as the characteristics and dynamics of each classified ice-marginal lake. The type of damming has been included in other inventories, proving to be useful for assessing present and future lake conditions under a changing climate e.g.,. Incorporating known GLOFs and/or drainage periods for each lake would also provide insight into abrupt changes in terrestrial water storage and be highly valuable information for infrastructure assessments, such as hydropower utilities e.g.,.