Landsat archives consist of optical data from a series of Earth-observing satellite missions. For this study, Landsat tiles covering the NSR were obtained from the United States Geological Survey (USGS). Surface temperatures on lakes were estimated from the thermal infrared (TIR) bands of Landsat 5 TM (Thematic Mapper; 1984–2013), Landsat 7 ETM+ (Enhanced Thematic Mapper Plus; 1999–present), and Landsat 8 OLI/TIRS (Operational Land Imager and the Thermal Infrared Sensor; 2013–present) instruments. Landsat instruments orbit at an altitude of 705 km, are sun-synchronous, and have a 16 d repeat cycle. The thermal band (band 6) of Landsat 5 and Landsat 7 record emitted radiation between the wavelengths of 10.40 and 12.50 µm, while that of Landsat 8 (band 10) records between 10.6 and 11.19 µm. The spatial resolution of the thermal bands for Landsat 5 TM (120 m), Landsat 7 ETM+ (60 m), and Landsat 8 OLI/TIRS (100 m) are resampled with the cubic convolution method and distributed at a spatial resolution of 30 m to match optical bands (USGS, 2022). In addition, other bands, including the quality band (BQA), near-infrared band, red band, and metadata, are also used in the retrieval of LST. About 34 Landsat tile scenes cover the NSR, with each tile containing 5000 × 5000 30 m pixels and overpass times ranging from 18:00 to 20:00 UTC.

Total column water vapour obtained from ERA5 reanalysis data (Copernicus Climate Change Service – C3S) from 1984 to 2021 was used as input in the algorithm to correct for atmospheric effects on Landsat images. Values were derived from the hourly data, with a ∼ 30 km spatial resolution obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF; Hersbach et al., 2020). ERA5 reanalysis data are a dataset generated from a combination of in situ observations and modelling to provide estimates of land, atmospheric, and ocean changes on a global scale. Average ERA5 hourly total column water vapour on single levels was used as an input in the LST retrieval algorithm.

Each lake's name, location, depth, size, elevation, and outline were retrieved from a combination of the HydroLAKES database, CanVec Series, and the Water File – Lakes and Rivers database. The HydroLAKES database is a digital map repository developed in the Global HydroLAB (http://wp.geog.mcgill.ca/hydrolab/, last access: 20 May 2022) from a collection of several databases (e.g. global and regional databases like CanVec Series and Shuttle Radar Topography Mission, SRTM, Water Body Data; Slater et al., 2006). This database provides information on world lakes and their significant properties through high-resolution maps. Over 1 427 688 individual lake vector polygons greater than 10 ha are included in the repository (Messager et al., 2016). The mode of pixel-level lake elevation data obtained from the EarthEnv-DEM90 (digital elevation model) and the USGS-provided GTOPO30 DEM was used to calculate the HydroLAKES elevation data. A geostatistical model developed from surrounding land surface topography was derived to generate average lake depths and volumes (Messager et al., 2016). As part of the Government of Canada initiative (https://open.canada.ca, last access: 28 March 2022), the CanVec Series provides a geometric description and fundamental characteristics of hydrographic phenomena in the form of geospatial vector data. The Water File – Lakes and Rivers (polygons) data (https://www12.statcan.gc.ca, last access: 30 March 2022) maps lakes and rivers under the 2006 census, as created by Statistics Canada under the Government of Canada on 29 August 2013. These data were the primary source of lake names attributed to lake outlines in our dataset.

Landsat-derived LST was generated during both open-water and ice-covered seasons. Retrieved data were evaluated against in situ measurements collected over selected locations within the study area (Fig. 1). In situ measurements from Mackenzie DataStream were used for evaluating LST derived from Landsat. DataStream is an open-access freshwater data platform that provides water-monitoring data collected by governments and communities across Canada. The database for the NWT region was the product of the NWT-wide community-based water quality monitoring (CBM) programme, which is collected during open-water seasons. The CBM programme was implemented in 2012 as part of a partnership between the Department of Environment and Natural Resources (ENR), the Government of the Northwest Territories (GNWT), communities, and regional organisations in NWT to monitor water quality and changes. The surface temperatures on lakes were measured with YSI sondes and EXO2 sondes and interpreted by ENR. Collated surface temperature data used for evaluation from this source were from the years 2014 to 2019. Another primary source was lake temperature data collected by Environment and Climate Change Canada (ECCC) from 1999 to 2003. Temperature loggers were used to measure hourly temperature on lakes during open water periods; however, only temperature collected at the skin surface (depth = 0 m) was used for LST evaluation in this study.

The MODIS (MYD11_L2) surface temperature dataset from 2003 to 2021 was used to evaluate the Landsat-derived LST data generated during both open-water and ice-covered seasons. The dataset was obtained from NASA's Earth Observing System Data and Information System (EOSDIS), mounted on the Terra and Aqua satellites; MODIS records within the spectral ranges of 0.405–14.385 µm across 36 bands. The Aqua product contains nighttime and daytime LST measurements on a spatial resolution of ∼ 1 km derived from the thermal infrared bands. For this study, the daytime LST measurement covering lakes in the NSR was compared against the Landsat-derived LST.

LST retrieval algorithms were applied to the thermal bands in conjunction with other processed output (bands and metadata) from Landsat data to generate the North Slave LST dataset. A quality assurance (QA) band, outlining the surface, atmosphere, and sensor conditions included in the Landsat data, was used to mask clouds and other obstructions. The QA band assesses cloud influence at different confidence levels (high, with 67 %–100 %, medium, with 34 %–66 %, and low, with 0 %–33 %), making it possible for cloud removal. This study categorised high and medium confidence values as cloud pixels, while low confidence values were considered to be cloud-free. LST retrieval algorithms and equations (Eqs. 1–11) were applied to the thermal bands of all tiles from 1984 to 2021. Cloud masks were generated and applied to retrieved LST to eliminate cloud-distorted pixels. LST pixels were extracted using the vector files of lake outlines from the HydroLAKES dataset. A 100 m negative buffer was applied to remove the effect of lake pixel mixing with land surface pixels. Possible erroneous pixels were flagged using z scores which calculate how far a value is from the mean. This was used to access spatial differences and outliers in pixels. Pixels of lakes with z-score values above 3.5 and below -3.5 were flagged. In addition, LST output with equal pixels across the entire lake or group of pixels having the same value to four decimal places were flagged. Further visual quality checks and analysis were applied to flagged LST to clean the generated data and remove the erroneous cloud cover that the masks could not capture. The overall framework for the retrieval and generation of the LST dataset for selected lakes in the NSR is highlighted in Fig. 2.

It is essential to highlight the limitations of data estimates from satellite-based records (Merchant et al., 2017). This provides an awareness of the degree to which a sensor is stable, in addition to the observations obtained from them. Additionally, these reports are necessary to inform the confidence of the data extracted and the structures of their errors through time and space. One significant distortion of the Landsat archives is the failure of the scan line corrector (SLC) of Landsat 7 ETM+ on 31 May 2003. As a result, the measurement from scans could not be corrected, causing all images sensed after that date to lose about 22 % of the data extracted. This limitation, named Landsat 7 ETM+ SLC-off issue, is more prominent at the edges of images than in the centre. Nevertheless, Landsat 7 ETM+ data were still used in the study because the radiometric and geometric corrections are unaffected by this scan line issue.

Indicators used to evaluate the performance of North Slave LST against in situ and MODIS LST were the root mean squared deviation (RMSD), mean bias deviation (MBD), and R squared (R2). The MBD assesses systematic differences and evaluates the underprediction and overprediction between two datasets (Eq. 12). A MBD value of 0 indicates an utterly random error. 12MBD=∑i=1NPi-OiN, where Oi and Pi are the observed and generated values, respectively, while N is the number of points used for evaluation. The index values ranged between 0 and 1, indicating the worst and best possible performance.

The root mean squared deviation (RMSD) measures the total difference between the two datasets without distinguishing between the over- or underprediction of models and/or algorithms (Eq. 13). No deviation in the values results in a RMSD value of 0. 13RMSD=∑i=1NPi-Oi2N

The accuracy of generated North Slave LST was examined by evaluating the Landsat-derived LST to corresponding in situ data (Fig. 5). Dates from measured in situ surface water temperature data (DataStream and ECCC) and derived North Slave LST data were matched. In addition, the widely used daily MODIS LST was compared with the generated dataset. Ground-based observations were compared against equivalent pixels within which measurements were taken, and the North Slave LST data were plotted against corresponding in situ surface temperature measurements (Fig. 5). A good correlation was observed between North Slave LST data and in situ surface water temperature, with an R2 value of 0.89 for the regression line. The North Slave LST was slightly higher than in situ records, with a MBD of 0.12 ∘C and RMSD of 1.71 ∘C.

Statistical parameters, including average yearly LST for the open-water season, total average LST, and variance, were calculated for available ECCC and DataStream in situ data and compared against North Slave LST, which is highlighted in Table 2. Absolute differences calculated for the parameters ranged from 0.1 to 1 ∘C. The highest absolute difference for the average LST of open water between the two datasets was 1 ∘C calculated for the year 2000 of the ECCC data. The variance was 0.6 and 0.3 ∘C for the ECCC and DataStream data, respectively. Differences between the total LST average were the lowest, with 0.3 ∘C from 1999 to 2003 and 0.1 ∘C from 2014 to 2019.

Deviations between North Slave LST and measured surface water temperature could be due to differences between image acquisition times and the time of in situ measurements. The Landsat capture times of the NSR were between 18:00 and 20:00 UTC, corresponding to 12:00 and 14:00 LT (local time). However, the time of in situ observations was variable and did not necessarily correspond to the time of satellite image acquisition. Further variations in North Slave LST can also be attributed to the differences in sample collection and spatial resolution. North Slave LST is essentially the mean of ∼ 60–120 m2 area instead of a single in situ location. Possible errors reported by other studies for the differences in measured and Landsat values include georeferencing and radiometric and memory effects (Chander and Markham, 2003; Markham et al., 2014; Sentlinger et al., 2008; USGS, 2022; Young et al., 2017).

MODIS LST was first compared against available water surface temperature measurements from DataStream (Fig. 6a) and North Slave LST for days when records were available for all three data sources. The aim was to compare the deviation of North Slave LST and water surface temperature to that of MODIS LST and water surface temperature. A relatively low coefficient of determination was observed for MODIS LST (R2=0.5) compared to the North Slave LST (R2=0.94) when evaluated against measured water surface temperature. RMSD values were also higher for MODIS LST (4.63 ∘C) than North Slave LST (1.55 ∘C), with a MBD of 2.35 and -0.12 ∘C for MODIS LST and North Slave LST, respectively.

North Slave LST was further compared against MODIS LST from 2003 to 2021 (ice-covered and open-water areas separately) for larger study lakes (30 km2) to avoid pixel mixing with land (Fig. 6b). Results showed a RMSD of 2.56 ∘C and MBD of 1.45 ∘C for ice-covered LST, suggesting an overestimation of North Slave LST during this period. On the other hand, an underestimation against MODIS LST was observed (MBD = -1.14 ∘C) for open-water LST with a relatively higher RMSD of 3.39 ∘C. This underestimation was expected as MODIS LST overestimates when compared against in situ data (Fig. 6a). Even though a prior comparison of MODIS LST to surface water temperature demonstrated a lower coefficient of determination, North Slave LST was still further compared against MODIS LST in this study. However, the decision to use MODIS for comparative analysis was due to the unavailability of continuous historical measurements of lake surface temperature. Additionally, MODIS LST provided an added outlook on the capability of North Slave LST to highlight historical trends despite the low temporal resolution by demonstrating a good correlation between the LST values (R=0.93).

Figure 7a and b demonstrate the yearly and monthly RMSD values derived from the comparison between North Slave LST and MODIS LST. Yearly RMSD shows a generally decreasing RMSD from earlier to later years, which may be attributed to Landsat's sensor change in recent years. LST values derived from 2013 onwards were extracted from Landsat 8 OLI/TIRS, which is known to have improved signal-to-noise ratio and calibration, higher 12 bit radiometric resolution, and narrower spectral bands compared to previous sensors (Irons et al., 2012; Roy et al., 2014). Most importantly, Landsat 8 OLI/TIRS has a radiometric uncertainty of 3 %, compared to Landsat 7 ETM+ (5 %), and a reduced band saturation (Markham et al., 2014). Monthly RMSD comparing MODIS LST to North Slave LST showed that RMSDs were lowest in spring and highest in winter. LST in spring months (March–May) had the least deviation (RMSD = 1.9–2.9 ∘C) when compared against MODIS data.

LST was derived from the thermal radiation of the lakes' uppermost layer; hence; the skin temperature is given. A total of 673 223 gridded data files and 536 tabular data were included in the generated North Slave LST dataset (https://doi.org/10.5683/SP3/J4GMC2) for the 535 lakes studied across the NSR (Attiah et al., 2022). The yearly and monthly distribution of the dataset within and between lakes varied temporally and is highlighted in Fig. 8. Overall, the yearly distribution of the North Slave LST dataset was greater in recent years, with the period between 2014 and 2021 having the majority of the data and yearly percentages ranging from 4.15 % to 5 % of the total dataset. The larger number of data files in recent years was due to LST retrieval from a combination of Landsat 7 and Landsat 8 compared to the single-sensor retrieval (Landsat 5) for earlier years. As a result, the highest yearly percentage of the North Slave LST dataset was for 2014 (5 %), and the lowest yearly percentage was for 1988 (1.2 %). Unavailable data for the various years were predominantly due to insufficient usable Landsat data for winter months.

The monthly distribution of the North Slave LST dataset showed the month of May with the highest percentage (13.9 %) and December (1.3 %) with the lowest. Generally, colder months (October–April) had fewer data (42.7 %) compared to relatively warmer months (May–September; 57.3 %). Data are unevenly distributed across months and years due to differences in overpass times and influences like cloud cover and other atmospheric impacts on data retrieval.

While study lakes are widely distributed across the NSR, a large number (144 out of 535) were within 150 km distance of Yellowknife. The average yearly number of images for each lake in the study region is demonstrated in Fig. 9. The average yearly minimum number of images was 20, reaching a maximum of 45. Lakes with a relatively smaller number of images were mainly distributed around Yellowknife. Smaller-sized lakes generally had fewer images than relatively larger ones, and this can be primarily attributed to clouds covering the entirety of small lakes. Most lakes (152 out of 535) had between 40 and 45 images, and 71 % of the total lakes in the dataset had more than 30 images per year. Lakes with fewer pixels are more likely to be entirely cloud covered and lose relatively more surface area due to the lake buffer.

The North Slave LST dataset includes LST for 500 lakes with known names and 35 without names. NetCDF (Network Common Data Form) and tabular data are the data types in this dataset. NetCDF is a file format for storing multidimensional data and, hence, capturing the spatial coverage and dimension of LST for study lakes in this dataset. To facilitate easy data queries, each NetCDF filename includes the name of the lake, date, longitude, latitude, minimum LST, maximum LST, mean LST, number of pixels, and the percentage area of the LST pixels coverage of the lake for a given day. The naming convention for lakes and their explanation is summarised in Table 3. The dataset was grouped based on the first alphabet of the lake name, then the name of the lake, and finally into yearly sub-groups.

The NetCDF files in our dataset have a two-dimensional variable, lst, which shows the spatial distribution of lake surface temperature. In addition, the one-dimensional x and y axes show the lake's extent and the number of pixels. The spatial reference for the data is the World Geodetic System 1984, EPSG:4326, with a 30 m resolution.

The second type of data included in the North Slave LST dataset is the tabular data containing LST statistics on individual study lakes for a given day. Derived attributes include the minimum, maximum, median, mean, number of ice cover pixels, number of open-water pixels, percentage of lake captured, and other lake properties. Table 4 highlights the column and field names from the tabular data and what they represent. These tabular data are generated for each lake and are included in the dataset. Each filename consists of the lake name followed by longitude and latitude, for easy query, based on location (e.g. AcastaLake_-115.564_65.3783). Additionally, monthly means were calculated for each lake and combined into one file in the dataset.

The seasonal spatial distribution of mean LST from 1984 to 2021 is shown in Fig. 10 to highlight the spatial variation in LST for different seasons. The distribution of average LST was computed for winter (December–January), spring (March–May), summer (June–August), and autumn (September–November) for all study lakes. LST on lakes in the NSR is generally negative in winter (-26 to -18 ∘C) and spring (-17 to -3 ∘C). This is because lakes are ice covered during these two seasons, constituting negative LST values. Autumn was characterised by both positive and negative LST values (-8 to 3 ∘C). Lakes start to freeze in autumn, and the freezing rate is influenced by several factors, resulting in differences in the open-water duration, which affect the average temperature. The average LST for summer ranged from 6 to 22 ∘C. Winter had the lowest LST ranges with a variability of 8 ∘C, while LST variability for summer was twice (16 ∘C) that of winter. This is expected, as temperatures on lakes during this season are influenced by several factors, including lake size, elevation, depth, latitude, longitude, and volume (O'Reilly et al., 2015; Xie et al., 2022), in addition to air temperature. Seasonal LST spatial distribution provides insight into the climate patterns of the NSR region.

The spatial distribution of the mean annual LST across the NSR for 2021 is shown in Fig. 11a, which highlights the remarkable spatial differences between lakes at higher versus lower elevations, with lower-elevation lakes generally demonstrating higher LST. Based on the mean annual LST values in 2021, the LST category was divided into five different ranges, as shown in the map (-12 to -9 ∘C, -9 to -6 ∘C, -6 to -3 ∘C, -3 to 0 ∘C, and 0 to 3 ∘C). Figure 11b shows most lakes (28 %), with a mean of -3 to 0 ∘C. The lake distribution of mean LST was 8 %, 22 %, 27 %, 28 %, and 15 % from colder to warmer LST categories, respectively. The percentage of the total area covered by lakes in relation to mean LST was 34 %, 27 %, 18 %, 19 %, and 2 %, respectively (Fig. 11b). Although the number of lakes with LST ranging from -12 to -9 ∘C was the least (8 % of lakes), the percentage of the total area covered by lakes with this LST range was the largest (34 % of lakes). The total area covered for all lakes with mean LST from 0 to 3 ∘C was only 2 %. This suggests that several of the lakes with warmer temperatures were smaller sized. Generally, relatively warmer lakes were also distributed around Yellowknife and the southwestern part of the region.

Lakes in several studies are treated as homogenous entities; however, there is spatial variability in the surface temperature of a given lake based on several factors, including the difference in morphometry or the biological, physical, and anthropogenic activities occurring on the lake at a given time (Crosman and Horel, 2009; Huang et al., 2017; Yang et al., 2020). The North Slave LST dataset generated in this study can highlight the spatial variability within a given lake. As expected, the high spatial resolution and multidate LST generated show lakes' surface temperature heterogeneity. The phenomenon is demonstrated with LST distribution on 9 July 2021 for a few selected lakes within our study given as examples (Fig. 12).

Lakes may demonstrate significant surface temperature variations for various reasons, including wind redistribution, depth, and biological and anthropogenic activities. Warmer LSTs are generally at the shallower coastal regions of lakes; however, internal LST variations differ. An example is Duncan Lake (Fig. 12), which demonstrated warmer temperatures in the northern part of the lake than in the south. Maximum and minimum LST on lakes also differ, with some lakes having wider variations (e.g. Duncan Lake; 23–14 ∘C) and others having less variation (e.g. Frame Lake; 28–24 ∘C). Lakes' physical differences, location, and elevation may contribute to the different ranges of surface temperature distribution on individual lakes.