Seasonal ice cover is one of the important attributes of
lakes in middle- and high-latitude regions. The annual freeze-up and breakup
dates as well as the duration of ice cover (i.e., lake ice phenology) are
sensitive to the weather and climate; hence, they can be used as an indicator
of climate variability and change. In addition to optical, active microwave,
and raw passive microwave data that can provide daily observations, the
Calibrated Enhanced-Resolution Brightness Temperature (CETB) dataset
available from the National Snow and Ice Data Center (NSIDC) provides an
alternate source of passive microwave brightness temperature (
Climate change is one of the major challenges facing humanity. New technologies and methods are urgently needed to monitor and quantify the rapid changes in climate at the regional and global scale. Lakes are closely tied to climate conditions and are characterized by many important parameters for long-term monitoring of climate change, including the coverage and duration of lake ice. For lakes located at middle and high latitudes, the spatial and temporal coverage of lake ice and key phenological events provide important information about changes in weather and climate. Lake ice phenology describes the seasonal evolution of ice cover, including the freeze-up and breakup dates, and the ice cover duration (Duguay et al., 2015; Sharma et al., 2016; Smejkalova et al., 2016). The presence or absence of lake ice affects lake–atmosphere interactions, thereby affecting hydrological and ecological processes in lakes (Duguay et al., 2006, 2015; Mishra et al., 2011; Hampton et al., 2017; Knoll et al., 2019). The coverage and duration of lake ice also affect human activities, such as transportation, fishing, and winter recreational activities (Brown and Duguay, 2010; Prowse et al., 2011; Du et al., 2017; Sharma et al., 2019). Changing climate conditions in the cold season will alter the temporal and spatial characteristics of mass (such as precipitation and suspended particles) and energy (such as solar radiation and atmospheric heat) input into the lake, thereby affecting the freeze–thaw processes of lake ice (Mishra et al., 2011). On the other hand, changes in the timing of freeze-up and breakup will cause sudden changes in lake surface properties (such as albedo and roughness) and affect the exchange between lakes and the atmosphere.
Lake ice (as well as river ice) phenology is one of the most detailed
climate data records (CDRs) with the longest historical coverage. For example,
the Global Lake and River Ice Phenology Database (GLRIPD) from the National
Snow and Ice Data Center (NSIDC) contains ice phenology records for 865 sites, including 24 438 ice-on records and 33 370 ice-off records, where the
earliest record can date back to 1443 (Benson et al., 2000). GLRIPD provides
a valuable data source for the study of historical changes in river and lake
ice in the Northern Hemisphere. Based on this dataset, several
investigations have reported that the ice phenology of rivers and lakes in
the Northern Hemisphere has been changing with trends towards later
freeze-up and earlier breakup in different periods of the past century. For
example, Magnuson et al. (2000) found that the freeze-up dates of river and
lake ice had a delaying trend at a rate of 5.8 d per century, and the
breakup dates had an advancing trend at a rate of 6.5 d per century from
1846 to 1995; moreover, the change rates of the freeze-up and breakup dates
accelerated after 1950. Similarly, Benson et al. (2012) found that the
freeze-up dates were occurring later at a rate of 0.03–0.16 d yr
Satellite remote sensing has the advantage of observing the Earth's surface
over large areas at a fixed time interval and has been widely used to
monitor lake ice in recent decades. Most spaceborne satellites can observe
the status of lake ice based on the different signals returned by ice and
water. According to the wavelength used, remote sensing of lake ice can be
divided into two categories: optical remote sensing and microwave remote
sensing. Optical sensors with medium resolution (ca. 250–1000 m) usually
provide a short revisit period (daily) and, therefore, have been widely used
to estimate lake ice phenology (Arp et al., 2013; Kropáček et al.,
2013; Smejkalova et al., 2016; Weber et al., 2016). For example, the
Advanced Very High Resolution Radiometer (AVHRR) reflectance data have been
used to extend existing in situ observations in Canada (Latifovic and
Pouliot, 2007). The Moderate Resolution Imaging Spectroradiometer (MODIS)
land surface temperature product (1 km) and snow cover product (500 m) have
also been used to determine lake ice status and extract lake ice phenology
(Nonaka et al., 2007; Hall et al., 2010; Pour et al., 2012; Weber et al.,
2016; Kropáček et al., 2013; Cai et al., 2019; Wu et al., 2021).
However, optical sensors are easily affected by cloud cover and illumination
conditions, which limits their ability for lake ice monitoring under cloudy
weather and at high latitudes (Maslanik and Barry, 1987; Helfrich et al.,
2007; Kang et al., 2012). Microwave remote sensing allows for acquisitions
regardless of cloud cover and dust in the atmosphere and is not affected by
illumination conditions (Engram et al., 2018; Geldsetzer et al., 2010).
Active microwave provides capabilities for the monitoring of lake ice based on
the difference in backscatter between lake ice and open water. For example,
the European Remote Sensing Satellite (ERS)-1/2 synthetic aperture radar
(SAR) (Jeffries et al., 1994; Morris et al., 1995; Duguay and Lafleur, 2003)
and Radarsat-1/2 SAR (Duguay et al., 2002; Geldsetzer et al., 2010) have
been successfully used to obtain lake ice status. However, existing active
microwave technology is limited by the narrow swath width, relatively low
temporal resolution (especially at lower latitudes), and short historical
coverage, making it difficult to monitor lake ice daily at a large scale
(Latifovic and Pouliot, 2007; Chaouch et al., 2014) as required by the
Global Climate Observing System (GCOS; Belward et al., 2016). Passive
microwave sensors can capture lake ice status based on the difference in
microwave radiation emitted from lake ice and open water. Microwave
radiometers mounted on existing polar-orbiting satellite platforms can
provide daily observations across the Northern Hemisphere and, therefore, can
be used to monitor lake ice phenology. For example, Du et al. (2017) used
Advanced Microwave Scanning Radiometer for Earth Observing System and
Advanced Microwave Scanning Radiometer 2 (AMSR-E and AMSR2, respectively) daily
The latest Calibrated Enhanced-Resolution Brightness Temperature (CETB)
dataset released by NSIDC provides passive microwave
This study uses SMMR and SSM/I–SSMIS data from the new CETB dataset to generate a lake ice phenology data record of the Northern Hemisphere from 1979 to 2019. However, two problems need to be solved: (1) how to extract lake ice phenology events from SMMR and SSM/I–SSMIS data; and (2) how to select lake ice phenology results from multiple satellites with overlapping years. The workflow behind the production of the dataset of annual ice dates and durations for 56 lakes is described. Then, the accuracy of the derived ice dates is compared against existing in situ observations and satellite-based lake ice datasets. Finally, the spatial characteristics of the lake ice phenology in the Northern Hemisphere are analyzed.
To make sure that there is at least one complete passive microwave pixel in
the center of the lake, lakes with a large enough area or a nearly circular
shape were selected. As a result, 56 lakes were selected as the study lakes.
Lake boundaries from the European Space Agency (ESA) Lakes Climate Change
Initiative (Lakes_cci) project
(
Location of the 56 study lakes (the names of the lakes are listed in Table 1). © OpenStreetMap contributors 2021. Distributed under the Open Data Commons Open Database License (ODbL) v1.0.
The location and physical characteristics of the 56 study lakes as well as the number of pixels used for the lake ice phenology extraction.
The CETB dataset consists of gridded enhanced-resolution
GLRIPD contains in situ observations of ice-on and ice-off dates for 865 lakes and rivers of the Northern Hemisphere. In this database, ice on is
defined as the freeze-up end date, and ice off represents the date of
breakup end (Benson et al., 2000). The GLRIPD data were obtained from NSIDC (
Example of determining the lake ice status for a pixel and
extracting the ice phenology for a lake. Panel
The daily AMSR-E/AMSR2 lake ice phenology product (Du et al., 2017) was also
used for comparison with the lake ice phenology results derived from the CETB
dataset. The product was produced based on 36.5 GHz H-polarized
The Great Lakes Environmental Research Laboratory (GLERL) from the National
Oceanic and Atmospheric Administration (NOAA) (
Moving
Percentage of effective observations and the sequence for different satellites.
For each pixel in a lake, the algorithm was applied for the
Lake ice phenology for Great Bear Lake extracted from multiple
satellites, showing the
For the remaining pixels after filtering, the number of daily lake ice/lake water pixels were counted. Afterwards, thresholds of 5 % and 95 % of the total lake pixels were set to extract the lake ice phenology (Fig. 2c, d). In fall or winter, when the number of lake ice pixels is larger than 5 % of the total pixels, the lake is considered to start to form ice cover (freeze-up start), and if the number of lake ice pixels is larger than 95 %, the lake is considered to be completely frozen (freeze-up end). Similarly, in spring or summer, when the number of lake ice pixels is less than 95 %, the lake ice is considered to start to break up (breakup start), and if the number of lake ice pixels is less than 5 %, the lake is considered to be completely ice-free (breakup end). The ice duration can then be calculated: the complete freezing duration represents the period from freeze-up end to breakup start; the ice cover duration represents the period from freeze-up start to breakup end.
Differences in absolute average uncertainty among years and lakes:
Some lakes in warmer regions did not form ice cover in certain years. For
these lakes with intermittent ice cover, if no ice cover was detected
throughout the year, the ice cover duration was recorded as zero; if lake ice
formed but did not completely cover the lake surface, the freeze-up start
and breakup end dates were extracted, while the complete freezing duration
was recorded as zero. To avoid the impact of short-term weather events on the
As there are overlapping time periods for certain satellites, some years may have multiple lake ice results. Using Great Bear Lake as an example, 103-year lake ice phenology results can be obtained from all 10 satellites. However, these data were not evenly distributed within the 41 years. The lake ice phenology can be obtained simultaneously from four satellites from 2011 to 2019, while only data from one satellite can be used from 1979 to 1992 (Fig. 3).
For each satellite, only data covering a complete year (from September to
August of the following year) were retained for lake ice phenology
extraction, except for Nimbus 7 (with no overlapping data from other
satellites) which acquired data from 25 October 1978 until 20 August 1987.
Therefore, the freeze-up dates of 1987 could not be obtained for some lakes.
In addition, an abnormal
Although all of the satellites have a nominal daily temporal resolution, they
have different frequencies of missing retrievals due to the polar orbit
pattern and different sensor swath width. As there are no alternative
data for Nimbus 7 and F08, the lake ice phenology for some lakes prior to
1992 (especially before 1987, due to the low temporal resolution and poor
data quality of SMMR data) were not complete. After the launch of F11, there
were at least two satellites that could retrieve
Ideally, we can obtain four ice dates (freeze-up start, freeze-up end, breakup start, and breakup end dates) in a year. However, due to data quality limitations, we failed to obtain effective observations for some lakes in some years. Using F08 as an example, we should have obtained 896 records (four ice dates for each lake in each year) for all of the 56 lakes from 1988 to 1991; however, mainly due to the continuous missing data in the winter of 1987, we finally obtained 842 ice dates. Based on the number of observations that were actually available compared with those that were expected, we calculated the percentage of effective observations (which was 93.97 % for F08; Table 2), as the basis for selecting annual ice phenology dates from overlapping results. With more missing data and a larger footprint, SMMR from the Nimbus satellite had the lowest percentage of effective observations among all of the 10 satellites. Unfortunately, there were no alternative data to improve the poor data results caused by the data quality. The six SSM/I satellites can all attain an effective observation of more than 90 %. Unexpectedly, we did not obtain the best lake ice results from the three SSMIS satellites, and their effective observation percentage values were all lower than 90 %.
We ranked the percentage of effective observations of the different satellites in a priority list with respect to obtaining lake ice phenology (Table 2). For years with multiple lake ice records extracted from more than one satellite, we prioritized the results from the satellite with the highest percentage of effective observations. Therefore, if a lake had complete lake ice results from all of the satellites, we would use the results from Nimbus 7 for 1979–1987, F08 for 1988–1991, F10 for 1992–1995, F13 for 1996–1997 and 2009, F14 for 1998–2008, and F15 for 2010–2019.
Variations in the number of lake ice and open water pixels for
Comparisons of lake ice phenology results from different satellites. FUS, FUE, BUS, and BUE represent freeze-up start, freeze-up end, breakup start, and breakup end date, respectively.
MAE: mean absolute error.
Finally, a data record of annual ice phenology for the 56 lakes from 1979 to 2019 was obtained. To increase the completeness of the lake ice phenology records, breakup dates for two lakes (Nettilling and Amadjuak) in 1987 were obtained from early F08 data (the two lakes had ice cover until late August, while Nimbus 7 data ended on 20 August 1987, and F08 started from 9 July 1987). Apart from the automatically extracted lake ice dates, 376 lake ice dates (i.e., 2.08 % of all of the records) were manually extracted to increase the ice phenology records (by setting a looser threshold in the extraction of lake ice phenology dates). Among these manually extracted lake ice dates, 263 dates were extracted from Nimbus 7 (i.e., 24.24 % of all of the Nimbus 7 records).
There are two main error sources for lake ice phenology derived from passive
microwave data: (1) the periodically missing data caused by the polar orbit
operation mode of passive microwave satellites and (2) errors associated
with the extraction process of lake ice phenology. Although the nominal
temporal resolution of passive microwave data was 1 d, there were
periodic gaps in the
Comparison of ice-on and ice-off dates from GLRIPD records and SMMR and SSM/I–SSMIS data. Bold numbers correspond to
MAE: mean absolute error.
Comparison of annual maximum ice cover of the Great Lakes from GLERL
records and SMMR and SSM/I–SSMIS data. Bold numbers correspond to
MAE: mean absolute error.
The errors in the extraction of lake ice phenology were mainly caused by
mixed pixels. Although the grid spacing for the enhanced-resolution
In addition, for each pixel, we only determined whether the pixel was
covered by lake ice or not. However, the area for a single passive microwave
pixel is 9.77 km
Furthermore, during application of the MTT algorithm, multiple smoothing
approaches were applied to the original
Therefore, we adopted an automatic algorithm to extract lake ice phenology dates from all of the platforms for all of the lakes and then prioritized the lake ice phenology results extracted from the satellite with a higher percentage of effective observations, which was in order to ensure the comparability of lake ice phenology results among lakes and the consistency of the time series. Despite the uncertainties and inevitable errors caused by the periodically missing data and mixed pixels, the lake ice phenology derived from SMMR and SSM/I–SSMIS can provide reliable information about the differences among lakes and the variations in time series.
For all of the satellite pairs with overlapping years, we calculated the bias
and mean absolute error (MAE) of the overlapping results. Different
satellites can obtain lake ice dates with a bias of
Box plot of the
Average, median, minimum, maximum, extreme difference (Max–Min), and standard deviation (SD) values of lake ice phenology in the Northern Hemisphere. FUS, FUE, BUS, BUE, CFD, and ICD represent the freeze-up start date, freeze-up end date, breakup start date, breakup end date, complete freezing duration, and ice cover duration, respectively.
The lake ice dates recorded in the GLRIPD were compared with the results
extracted from SMMR and SSM/I–SSMIS data. As the ice-on and ice-off
dates in the GLRIPD represent the first date of complete ice cover and the
last date of the breakup process, respectively, we used the freeze-up end and
breakup end results to make the comparison. For Lake Superior and Ladoga,
which had incomplete freeze-up end records, freeze-up start dates were used
instead. Only the lake ice records with an overlapping time of more than 10 years were selected for the comparison. As a result, ice-on dates for five lakes (8 sites) and ice-off dates for seven lakes (10 sites) were compared, and
the correlation coefficient (
The AMSR-E/AMSR2 lake ice phenology product (Du et al., 2017) was also used to
extract the ice dates for the study lakes from 2003 to 2015 (except for 2012
due to the data gap between AMSR-E and AMSR2 from 4 October 2011 to 23 July 2012), and the r, bias, and MAE values compared with the ice dates from
SMMR and SSM/I–SSMIS data were calculated for each lake (Fig. 6). The
freeze-up start and breakup end dates for 55 lakes (except for Large Aral
Sea, which has no records after 2003) and the freeze-up end and breakup start dates
from 49 lakes (except for Large Aral Sea and some lakes without complete ice
cover in winter) were compared. The breakup dates had an overall higher
consistency between the results from the two datasets than the freeze-up
dates (Fig. 6a). Among them, the freeze-up start dates for 49 lakes, the
freeze-up end dates for 45 lakes, the breakup start dates for 46 lakes, and
the breakup end dates for 53 lakes were significantly consistent. The
biases for all pairs of freeze-up start, freeze-up end, breakup start, and
breakup end dates were 2,
Apart from lake ice phenology dates, the annual maximum ice cover was extracted for seven intermittently ice-covered lakes with more than 1000 pixels (Superior, Huron, Michigan, Erie, Ontario, Caspian Sea, and Ladoga) from SMMR and SSM/I–SSMIS data and compared to that from the five Great Lakes contained in the GLERL historical ice cover records. Ice cover maximums from SMMR and SSM/I–SSMIS data and GLERL were significantly consistent for these lakes (Table 5). Among them, the ice cover values extracted for Erie showed the lowest bias and MAE, due to the fact that this lake experiences an extensive ice coverage in winter (Fig. 7d). The remaining four lakes all showed a negative bias, indicating that the ice cover values extracted from SMMR and SSM/I–SSMIS data were usually smaller than the actual situation. This is not only because a buffer of 6.25 km was used to exclude pixels near the lake shore, which happens to be the place where lake ice forms first, but short-term ice cover, which was common on these lakes, was also difficult to detect with the MTT algorithm. This is also why the lake was sometimes 100 % ice covered but only partial coverage was detected by SMMR and SSM/I–SSMIS data. Using 2014 as an example, we recalculated the daily ice cover from the GLERL ice charts over the same areas as the SMMR and SSM/I–SSMIS data to make a comparison of daily ice cover changes from the two datasets (Fig. 8). High consistency in the daily ice cover percentage can be seen for Erie, Huron, and Superior; however, for Michigan and Ontario, a similar change pattern but lower ice cover was obtained from SMMR and SSM/I–SSMIS data compared with the GLERL records because the short-term ice cover on Michigan and Ontario in February and March was not detected by the MTT algorithm after the smoothing approaches were applied (Fig. 8). Although it is difficult to capture the short-term changes in ice cover, the algorithm showed good performance with respect to obtaining ice cover for lakes with long-term ice cover.
Comparisons of the annual maximum ice cover (percentage of lake area) of
the Great Lakes from the GLERL records and SMMR and SSM/I–SSMIS data:
Comparisons of the daily ice cover (percentage of lake area) of the Great
Lakes from the GLERL records and SMMR and SSM/I–SSMIS data in 2014:
Among the 56 study lakes, 45 lakes experienced annual ice cover during their entire lake ice phenology records, and the remaining 11 lakes had no ice detected for 1 year or more. Note that, as we only selected the pixels 6.25 km away from the lake shore to extract the lake ice phenology, it is possible that some lakes had ice cover near the lake shore, but we did not obtain the information. The average date and days of the respective ice dates and durations for each lake during their entire lake ice phenology records were calculated (relative to 1 September) and are shown in Fig. 9. In order to avoid bias in the statistics for the lakes with no ice detected in some years, only the 45 lakes with annual ice cover were included. A total of 2 of the 45 lakes (Caspian Sea and Ladoga) did not have annual complete ice cover in winter; hence, only their freeze-up start dates, breakup end dates, and ice cover durations were calculated (Fig. 9). In addition, the statistical results of ice phenology for the 45 lakes are shown in Table 6, including their average, medium, minimum (earliest/shortest), and maximum (latest/longest) date/days as well as the extreme difference (maximum–minimum) and standard deviation. Here, we can see that the period of ice formation is from the end of October to January of the next year, and the period of ice breakup is from mid-March to late July. Overall, the differences among freeze-up dates are smaller than those for breakup dates (with smaller extreme differences and standard deviations; Table 6), but the spatial characteristics of breakup dates are more consistent with latitude (Fig. 9). Different lakes experienced a complete freezing duration ranging from 58 to 268 d (average 153 d) and an ice cover duration ranging from 62 to 275 d (average 161 d). Lakes in low-latitude Eurasia had the shortest ice cover durations, whereas lakes in northern Canada had the longest ice cover durations (Fig. 9).
Average date/days of ice phenology for the study lakes for their
entire lake ice phenology records extracted from SMMR and SSM/I–SSMIS
data:
The annual lake ice phenology records for the 56 study lakes are available
at
This study used passive microwave SMMR and SSM/I–SSMIS data available on a 3.125 km grid from the CETB dataset to extract the ice phenology for 56 lakes in the Northern Hemisphere from 1979 to 2019. An automatic threshold algorithm based on the MTT method was applied to determine the lake ice status for each pixel, and a buffer of 6.25 km was set to exclude pixels with high land contamination. Then, ice phenology dates were extracted for each lake using the thresholds of 5 % and 95 % of the total pixels. To keep the lake ice phenology consistent and comparable among different years and different lakes, for the overlapping lake ice phenology results extracted from multiple satellites, we prioritized the results from the satellite with the highest percentage of effective observations.
The main error sources of the lake ice phenology extracted from SMMR and SSM/I–SSMIS data were attributed to the periodically missing data at middle
and low latitudes caused by the polar orbit operation mode and the mixed
pixels caused by the original coarse spatial resolution of the satellite
acquisitions. Ice phenology results for the lakes at low latitudes and/or
with small areas tend to be characterized by a larger uncertainty. The
switch between sensors and satellites over the time series also resulted in
certain differences in the lake ice phenology results, with a bias ranging
from
The lake ice phenology results extracted from SMMR and SSM/I–SSMIS data
were compared to the ice dates from in situ observations and an
AMSR-E/AMSR2-derived lake ice phenology product. Compared to in situ
observations, the ice-on dates for 4 out of 8 sites, and the ice-off dates
for 8 out of 10 sites were significantly consistent. The differences between
the in situ observations and the lake ice phenology extracted in this study
were mainly due to the different fields of view of human observers versus
satellite instruments. As for the AMSR-E/AMSR2 product, the ice dates showed strong
agreement, with biases ranging from
From 1979 to 2019, the average complete freezing duration and ice cover duration for all lakes forming an annual ice cover were 153 d (58–268 d for individual lake) and 161 d (62–275 d). Lakes in the low latitudes of Central Asia had the shortest ice durations, whereas lakes in northern Canada had the longest ice durations.
The new dataset consists of lake ice phenology records derived from SMMR and SSM/I–SSMIS data from 1979 to 2019. Lake ice phenology (freeze-up, breakup, and ice cover duration) is a robust indicator of climate change. The analysis-ready dataset is available to the science/user community to investigate, among several possible topics, the following: (1) trends and interannual variability in lake ice phenology across the Northern Hemisphere in response to climate change and atmospheric teleconnections patterns; and (2) the impact of changing ice phenology on local/regional weather, climate and hydrology, aquatic ecosystems, and cultural and socioeconomic activities.
Finally, regular updates of the lake ice phenology data record are planned with future releases of the CETB dataset. Work is also underway on further improving the lake ice phenology retrieval algorithm and product as well as on the possibly of adding more lakes in a future release.
YC and CRD conceived the idea. YC designed the code and carried out the data processing. YC prepared the manuscript with contributions from all co-authors. CQK supervised the study.
The contact author has declared that none of the authors has any competing interests
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The authors wish to thank all of the data providers for their data. We are also grateful to the editors and reviewers for their constructive comments which helped to improve this paper.
This research has been supported by the National Natural Science Foundation of China (grant nos. 41830105 and 42011530120), the China Scholarship Council (grant no. 201906190109), and the Natural Sciences and Engineering Research Council of Canada (grant no. RGPIN-2017-05049).
This paper was edited by Hanqin Tian and reviewed by Tamlin Pavelsky, J. Uusikivi, and one anonymous referee.