Contributors: B. Calmettes (CLS), L. Zawadzki (CLS), J.-F. Crétaux (LEGOS), L. Carrea (University of Reading), C.J. Merchant (University of Reading)
Issued by: University of Reading / L Carrea, CJ Merchant
Date: 31/05/2020
Ref: C3S_312b_Lot4_D2.LK.2-v2.0_202005_Product_Quality_Assessment_Report_LSWT_v1.0
Official reference number service contract: 2018/C3S_312b_Lot4_EODC/SC2
History of modifications
List of datasets covered by this document
Related documents
Acronyms
General definitions
L2P – Geophysical variables derived from Level 1 source data on the same grid (typically the satellite swath projection). Ancillary data and metadata added following GHRSST Data Specification.
L3U – L2 data granules remapped to a regular latitude/longitude grid without combining observations from multiple source files. L3U files will typically be “sparse” corresponding to a single satellite orbit.
L3C – LSWT observations from a single instrument combined into a space-time grid. A typical L3C file may contain all the observations from a single instrument in a 24-hour period.
L3S - Level 3 super-collated: this is a designation of satellite data processing level. “Level 3” indicates that the satellite data is a geophysical quantity (retrieval) that has been averaged where data are available to a regular grid in time and space. “Super-collated” indicates that (for those periods where more than one satellite data stream delivering the geophysical quantity has been available) the data from more than one satellite have been gridded together into a single grid-cell estimate, where relevant.
Uncertainty: The dispersion of values that might reasonably be attributed to a measurand given a measured value, quantified by the standard deviation.
Bias: Estimate of a systematic error
Scope of the document
This document is the Product Quality Assessment Report (PQAR). This document summarises the results from the product assessment based on the Product Quality Assurance Document (V2) [D4] for the C3S Lake Surface Water Temperature product.
Executive summary
The C3S Lake production system (C3S ECV LK) provides an operational service, generating lake surface water temperature (LSWT) and lake water level (LWL) climate datasets for a wide variety of users within the climate change community. The present document covers the LSWT system.
This document presents the results of the quality assessments undertaken for the product including the generation of the matchups at L2 with an updated in situ measurement database, the validation of the L2 temperatures, the validation of the GloboLakes1/C3S LSWT v4.0 product including the temperature and its uncertainty. An assessment of which lakes return more usefully complete data series is also made, since this is not a priori known before processing. The validation of the LSWT CDR v2.0 has been carried out with a larger in situ database with respect to the validation of LSWT CDR v1.0.
This document concerns the second contractual version of LSWT products for C3S, corresponding to “v4.0” processing. It will be updated as necessary for the future TCDR and CDR.
Product validation methodology
Validated products
The Lake Surface Water Temperature (LSWT) is the measure of the water temperature at the surface of the lake (skin temperature). The C3S Lakes product comprises a long-term climate data record (CDR), which includes a brokered dataset, the GloboLakes dataset (from 1995 to 2016) and the C3S extension (from 2017 to 2019). The time series have been computed from sensors on multiple satellites and lake-specific consistency adjustments between sensors have been applied using the MetOpA AVHRR instrument as a reference – MetOpA AVHRR has the best combination of length of record and data density for this purpose. The same algorithm has been used to retrieve the LSWT from all sensors in order to obtain consistent time series for each of the 1000 target lakes. The target list was defined within the GloboLakes project and can be found at http://www.laketemp.net/home_GL/.
The time periods used for each satellite/instrument are provided in Table 1. Not all lakes include LSWT from all sensors in the series because of differing density and geometry of observation. The products are observational, so gaps in time and space are common for all the lakes due to cloud cover and limited swath of the instruments.
Table 1: Time periods for the satellite/instrument used to generate the LSWT product
Satellite | Instrument | Time Period |
ERS-2 | ATSR-2 | 1995 – 2003 |
Envisat | AATSR | 2002 – 2012 |
MetOpA | AVHRR | 2007 – 2019 |
MetOpB | AVHRR | 2017 – 2019 |
A detailed description of the product generation is provided in the Algorithm Theoretical Basis Document (ATBD) [D3] with further information on the product given in the Product User Guide and Specifications (PUGS) [D5].
Validating datasets
A match-up dataset was constructed from the in situ temperature data collected through the ARCLake project, the GloboLakes project and the EU Surface Temperature for All Corners of Earth (EUSTACE) project, and expanded every year for the validation of the C3S product. For this version, the dataset consists of 183 observation locations covering 64 lakes, of which 29 locations over 8 lakes have been added within the C3S evolution. For some of the existing sites low temporal frequency data have been replaced with high temporal frequency temperature data. This has improved the number of matches. Details of the in situ observation locations with their sources are given in Table 2, which reports all locations for the target lakes where there are matches.
Table 2: List of the sources of the in situ data
Source | Lake name (number of locations) |
NDBC – National Data Buoy Centre (USA) | Superior (3), Huron (2), Michigan (2), Erie (1), Ontario (1) |
FOC – Fisheries and Oceans Canada (Canada) | Superior (1), Huron (4), Great Slave (2), Erie (2), Winnipeg (3), Ontario (4), Woods (1), Saint Claire (1), Nipissing (1), Simcoe (1) |
Michigan Technological University (USA) | Superior (2), Michigan (1) |
University of Minnesota (USA) | Superior (2), |
Northern University of Michigan (USA) | Superior (2), |
Superior Watershed Partnership (USA) | Superior (1) |
U.S. Army Corps of Engineers (USA) | Superior (1) |
Technical University of Kenya (Kenya) | Victoria (1) |
GLERL – Great Lakes Environmental Research Lab (USA) | Huron (3), Michigan (2) |
University of Wisconsin-Milwaukee (USA) | Michigan (2) |
Northwestern Michigan College (USA) | Michigan (1) |
University of Michigan CIGLR (USA) | Michigan (2) |
Limno Tech (USA) | Michigan (3), Erie (4) |
Illinois-Indiana Sea Grant and Purdue Civil Engineering (USA) | Michigan (2) |
Leibniz Institute for Freshwater Ecology and Inland Fisheries (Germany) | Tanganyika (1) |
Pierre Denis Plisnier | Tanganyika (4) |
Irkutsk State University (Russia) | Baikal (1) |
Regional Science Consortium (USA) | Erie (1) |
UGLOS – Upper Great Lakes Observing System (USA) | Erie (2), Douglas (1) |
LEGOS – Laboratoire d'Etudes en Géophysique et Océanographie Spatiales (France) | Issykkul (1) |
SLU – Swedish University of Agricultural Science (Sweden) | Vanern (6), Vattern (2), Malaren (9), Hjalmaren (1), Siljan (1), Bolmen (2), Roxen (1) |
Uppsala University (Sweden) | Vanern (1), Erken(1) |
Junsheng Li (China) | Taihu (1) |
KU Leven (Belgium) | Kivu (1) |
SYKE – Finnish Environment Institute (Finland) | Saimaa (2), Paijanne (3), Orivesi and Pyhaselka (2), Pielinen (4), Oulujarvi (1), Lokka (1), Hoytiainen (1), Vanajavesi (1), Pyhajarvi(1), Lappajarvi (1), Vuohijarvi (1) |
Vermont EPSCOR – Established Program to Stimulate Competitive Research (USA) | Champlain (1) |
SUNY Plattsburgh Center for Earth and Environmental Science (USA) | Champlain (1) |
Nipissing University (Canada) | Nipissing (2) |
NIWA (New Zealand) | Taupo (3) |
GLEON – Global Lake Ecological Observatory Network | Tanganyika (3), Balaton (1) |
BLI – Balaton Limnological Institute (Hungary) | Balaton (6) |
KDKVI – Central Transdanubian (Regional) Inspectorate for Environmental Protection, Nature Conservation and Water Management (Hungary) | Balaton (3) |
UMR CARRTEL – Centre Alpin de Recerche sur le Réseaux Trophique des Ecosystèmes Limniques (France) | Geneva (1) |
UC-Davis Tahoe Environmental Research Center (USA) | Tahoe (1) |
Utrecht University (Nederlands) | Garda (1) |
Italian National Research Council (Italy) | Garda (8), Trasimeno (3), Bolsena (1), Bracciano (1) |
Estonian University of Life Sciences (Estonia) | Vorstjarv (4) |
Environmental Protection Agency (Ireland) | Corrib (2), Derg (1) |
Martin Dokulil (Austria) | Neusiedl (1) |
Israel Oceanographic and Limnological Research (Israel) | Sea of Galilee (2) |
National Institute for Environmental Studies (Japan) | Kasumigaura (5) |
Universidad del Valle de Guatemala (Guatemala) | Atilian (1) |
Universitá degli Studi di Perugia (Italy) | Trasimeno (1) |
Centre for Ecology and Hydrology – Edinburgh (UK) | Lomond (1), Leven (1) |
University of Latvia and Latvian Environmental Geology and Meteorology Centre (Latvia) | Razna (1) |
University of Wisconsin-Madison (USA) | Mendota (1) |
NTL LTER – North Temperate Lakes Long-Term Ecological Research (USA) | Mendota (1), Trout (1) |
Uppsala University (Sweden) | Erken (1) |
The Ohio State University (USA) | Douglas (1) |
Table 3 lists the 58 lakes (where matches have been found) together with their maximum distance from land [Carrea et al. 2015], which is an indication of each lake’s size that is meaningful for LSWT remote sensing. Figure 1 shows the distance to land for Loch Lomond in Scotland. The highest resolution of the instruments used for the retrieval of the LSWT is 1 km. If, as here, the lake has a maximum distance to land of 1.5 km, the LSWT retrieval is likely to be available for that part of the lake from time to time, when a combination of factors occurs: the satellite image locations line up so that some pixels are nominally fully water pixels, which requires the satellite view zenith angle (which affects the on-the-ground resolution) to be such that the half-pixel size is smaller than the distance to coast; these pixels are cloud free; and image geolocation errors (which can be of order 1 pixel uncertainty) are small enough so that the nominally water-filled pixels are truly water-filled, meaning that the water detection tests are passed.
Figure 1: Distance to land in km for Loch Lomond in Scotland where each dot represents a 1/360 degree cell.
Table 3: List of the GloboLakes lakes with in situ data and their max distance to land
Lake ID | Lake | Country | N locations | Max distance to land (km) |
2 | Superior | Canada/USA | 12 | 73.5 |
3 | Victoria | Tanzania | 1 | 84.1 |
5 | Huron | Canada/USA | 9 | 73.3 |
6 | Michigan | USA | 15 | 63.8 |
7 | Tanganyika | Tanzania | 8 | 34.1 |
8 | Baikal | Russia | 1 | 33.7 |
11 | Great Slave | Canada | 2 | 44.6 |
12 | Erie | Canada | 10 | 45.6 |
13 | Winnipeg | Canada | 3 | 40.1 |
15 | Ontario | Canada | 5 | 36.1 |
25 | Issykkul | Kyrgyzstan | 1 | 26.9 |
29 | Vanern | Sweden | 7 | 20.3 |
44 | Woods | Canada | 1 | 11.8 |
66 | Taihu | China | 1 | 16 |
67 | Kivu | Zaire | 1 | 13 |
95 | Vattern | Sweden | 2 | 9.9 |
111 | Saimaa | Finland | 2 | 3.4 |
146 | Saint Claire | Canada | 1 | 13 |
157 | Paijanne | Finland | 3 | 3.8 |
163 | Malaren | Sweden | 9 | 2.7 |
165 | Champlain | USA | 2 | 5.8 |
187 | Orivesi and Pyhaselka | Finland | 2 | 4.4 |
195 | Pielinen | Finland | 4 | 4.1 |
198 | Nipissing | Canada | 3 | 9 |
202 | Oulujarvi | Finland | 1 | 6 |
236 | Simcoe | Canada | 1 | 8.4 |
295 | Taupo | New Zealand | 3 | 9.6 |
310 | Balaton | Hungary | 10 | 6 |
327 | Geneva | Switzerland | 1 | 6.2 |
376 | Lokka | Finland | 1 | 4.4 |
380 | Tahoe | USA | 1 | 8.2 |
387 | Hjalmaren | Sweden | 1 | 4.7 |
505 | Garda | Italy | 9 | 5.2 |
654 | Siljan | Sweden | 1 | 5.4 |
657 | Hoytiainen | Finland | 1 | 3.2 |
679 | Vorstjarv | Estonia | 4 | 6.2 |
1028 | Bolmen | Sweden | 2 | 2.7 |
1057 | Corrib | Ireland | 2 | 2.6 |
1115 | Neusiedl | Austria | 1 | 3.6 |
1196 | Sea of Galilee | Israel | 2 | 5.6 |
1201 | Vanajavesi | Finland | 1 | 3.4 |
1204 | Kasumigaura | Japan | 5 | 3.7 |
1240 | Pyhajarvi | Finland | 1 | 3.9 |
1246 | Lappajarvi | Finland | 1 | 3.5 |
1479 | Atilian | Guatemala | 1 | 4 |
1519 | Derg | Ireland | 1 | 1.9 |
1529 | Trasimeno | Italy | 3 | 4.3 |
1596 | Bolsena | Italy | 1 | 5.2 |
1893 | Roxen | Sweden | 1 | 2.9 |
2054 | Vuohijarvi | Finland | 1 | 1.9 |
2516 | Lomond | United Kingdom | 1 | 1.5 |
3307 | Bracciano | Italy | 1 | 3.9 |
3379 | Razna | Latvia | 1 | 3.2 |
4503 | Mendota | USA | 2 | 2.5 |
6785 | Erken | Sweden | 1 | 1.5 |
12262 | Leven | United Kingdom | 2 | 1.5 |
12471 | Trout | USA | 2 | 1.4 |
13377 | Douglas | USA | 2 | 1.5 |
A good portion of the lakes that have been used for the validation are small, which, given the previous discussion, is the most challenging situation for the LSWT retrieval.
Note also that some of the locations of in situ measurements are situated close to the coast. The nearest water-filled pixels may not overlap with the in situ measurement occasion in this circumstance, thus increasing the uncertainty in the comparison from spatial representativity.
The plot of the geographical distribution of the lakes with in situ measurements is shown in Figure 2 where the new acquired sites are reported in red. The lakes cover most of the latitudes and the continents except for South America. Most of the lakes are located in the Northern Hemisphere at relatively high latitudes.
Figure 2: Geographical distribution of the lakes with in situ measurements where the red lakes have sites with new measurements.
As the in situ data are from a variety of sources, with different formats, considerable effort has been put in to consolidate each new source of data to a standard format for use in validation. A quality control procedure for checking the in situ data is also necessary, since they are not always credible. This is partly automated and partly by manual inspection. The quality control procedure was initiated within the ARCLake2 project and updated within GloboLakes and C3S. Moreover, the in situ data have a range of characteristics:
- the measurements have been taken at different depths up to 1m;
- the temporal sampling of the measurements ranges from 15 minutes to few times a year;
- the temporal availability of the in situ measurements varies from few months up to covering all the satellite period;
- for some locations the measurements are averages while for others they have been taken at the reported time;
- none of the in situ measurements, which have been collected, are accompanied by an uncertainty estimation.
While part of the data are available online, the majority has been collected through personal communications and in a proportion of cases we are not licensed to redistribute the data.
Description of product validation methodology
Overview
The quality assessment of the Lake Surface Water Temperature product consists of the comparison of the dataset with independent in situ data. The satellite – in situ matches are created at L2, i.e. in the original satellite coordinates. The output products are gridded (L3S), so the L3 grid cell corresponding to L2 match is identified and the L3 product is thus directly validated. The validation of the L3S product is performed using conventional and robust statistics.
A separate dimension of product quality is the data density, which varies greatly between lakes and is also assessed. The metrics are: the number of target lakes that yield no useful data; and a map of the number of days when LSWT with quality levels=4,5 has been retrieved at the centre of each lake (defined as the position of the max distance to land).
Generation of the L2 matchup database and L2 validation
The satellite – in situ observation matches have been generated at L2, namely in satellite coordinates.
A per-sensor matchup is created and it contains space and time coincident satellite and in situ data. It also provides the reference and time of the in situ location and the associated LSWTs, quality level and uncertainty from the L2 LSWT product. The matchup is created for satellite observations based on the following criteria:
- Spatially within 3km from the location of the in situ measurement and
- Temporally within 3 hours for the in situ measurements where the measurement time was available. For some of the lakes only daily mean or unknown time of the measurements was on the record, therefore the day was matched.
Validation of the L3S GloboLakes/C3S LSWT v4.0
The differences between the L3S LSWT and in situ data are analysed using both standard and robust statistics. Robust statistics are less influenced by outliers in the distribution of differences. Time series of the absolute temperatures together with their difference are generated differentiating the quality levels. "Violin" plots, where the distribution of the difference is shown, are produced for each quality level, together with a scatter plot of the difference per quality level (where some jitter has been randomly added to reduce overlapping). Finally, the difference is plotted as a function of the in situ temperature and a line fit is performed for each quality level. The robust statistics is also investigated per quality level for each year and for each lake.
Validation of the LSWT uncertainty
The validation of the L3S GloboLakes/C3S LSWT v4.0 product is carried out comparing the satellite minus in situ temperature difference with the combination of the satellite uncertainty (present in the products) and an estimate of the in situ uncertainty (which is relatively poorly known). In an ideal case, the standard deviation of the differences between the satellite SST and a reference SST would equal the combined measurement uncertainty plus the uncertainty attribu to representativity effects.
Number of GloboLakes lakes with LSWT
An assessment of the lakes with no retrieved LSWT is reported together with an estimation of the number of observations per lake, which is performed counting the number of days with observations at the lake centre.
Validation results
This section provides the results of the LSWT product validation for the GloboLakes/C3S CDR.
Validation of the GloboLakes/C3S L3S LSWT v4.0
The matchup is attempted per sensor over the 183 locations on 64 lakes. For 52 locations only daily mean in situ data were available. Subsequently, the correspondent L3 cell is found and stored in a L3 database.
The total number of matches is 98781. The number of matches varies per year and, since the AVHRR sensors have a larger swath than the ATSR sensors (ATSRs swath is 500 km and AVHRRs swath is ~2900km), after 2007 the number of matches clearly increases as it is shown in Figure 3. We can notice another clear increase in 2017 when the AVHRR on MetOpA is used together with the AVHRR on MetOpB. In 2019 the number of matches is lower than the previous year because the GloboLakes/C3S LSWT time series ends in August 2019. The number of matches depends also on the availability of the in situ data since a different number of locations is available every year as shown in Figure 4. The number of locations where in situ measurements have been taken has doubled since 1995; however, a portion of the measurement frequency is daily.
Figure 3: Number of matches at L3 per year
Figure 4: Number of locations with matches at L3 per year
Table 4 reports the robust statistics and the traditional statistics per quality level for the matches across all the locations where in situ measurements were available as reported in Table 3.
The agreement between satellite and in situ measurements varies according to the quality levels in a way that is expected. In Table 4, the number of matches per quality levels are listed together with the median and the robust standard deviation of the satellite minus in situ temperature difference and the traditional metrics, the mean and the standard deviation. The difference between the median and the mean is almost neglectable for quality level 5 and it increases as the quality levels get lower, showing a symmetry in the distributions for high quality levels.
Table 4: Robust and traditional statistics of the satellite minus in situ difference per quality level
QL | N | Median | RSD | Mean | SD |
5 | 44630 | -0.120 | 0.578 | -0.133 | 1.072 |
4 | 17270 | -0.260 | 0.860 | -0.326 | 1.307 |
3 | 19268 | -0.240 | 1.008 | -0.387 | 1.472 |
2 | 11827 | -0.490 | 1.379 | -0.738 | 1.738 |
1 | 8786 | -4.135 | 5.789 | -5.029 | 5.311 |
The best agreement is for quality levels 4 and 5, which reflect a higher degree of confidence in the validity of the satellite estimate. As noted in the Product User Guide [D5], our recommendation to users is to use the highest quality level in preference, unless they have specifically verified for a given lake that lower quality levels are fit for their application. Quality level 3 data comparison with the in situ data shows an agreement that may be acceptable to some users; however, they have to be used with care. Quality level 1 data should never be used, and they are classified as “bad data”.
A contribution to the difference on average is the expected skin effect. Infrared radiometers are sensitive to radiation emitted between the air-surface interface and 20mm below the interface, while the in situ measurements considered here are taken at a distance up to 1m from the air-surface interface. During the night, the surface of the water is generally cooler than the subsurface of 0.2 K [Saunders, 1967], [Embury et al, 2012]. However, during the day, thermal stratification due to solar heating contributes to the difference in temperature between the radiometric lake surface and the in situ measurement depth (up to 1 m). At present, we have not assessed the degree of near-surface stratification to be expected in different lakes, which depends on fetch, weather conditions, depth of in-situ measurement, and any local vertical mixing perturbations introduced by the presence of the in-situ measurement system. In summary, a contribution to the satellite minus in situ temperature difference is the expected skin effect of 0.2 K, but it is difficult to infer a precise contribution of satellite LSWT biases to the remaining residual, in the face of in-situ errors and unquantified geophysical effects (near-surface stratification other than the skin effect, plus horizontal variability).
The distributions of the satellite minus in situ temperature differences per quality level are reported in Figure 5. In the upper part of the figures, the “violin” plots display the distributions of the differences per quality level. In the lower part of the figures, all the differences are plotted as strip plot where some horizontal jitter has been randomly assigned in order to reduce overlapping. The distributions become more stretched and less symmetric with longer tail towards negative differences as the quality levels decreases.
Figure 5: Distribution of the satellite minus in situ temperature difference per quality level. Top panel: distributions as “violin” plots where the width indicates the density of data for a given difference. Lower panel: same data showing individual points towards the extremes of the distributions; the horizonal spreads are randomly added jitter to reduce overlap of data points.
The median and the robust standard deviation per quality level (except quality level 1 which are bad data) per year for all the lakes is shown in Figure 6 and Figure 7 together with the number of matches. For high quality levels the median and the standard deviation are consistent throughout the years when different instruments have been adopted and a different number of matches is available. They deteriorate as the quality goes lower especially for the ATSRs sensors which have been employed exclusively up to 2007. The number of matches for quality level 5 is consistently the highest.
Figure 6: Satellite minus in situ temperature difference median per year (upper plot) and number of matches (lower plot) per quality level
Figure 7: Satellite minus in situ temperature difference robust standard deviation per year (upper plot) and number of matches (lower plot) per quality level
The median and robust standard deviation have been inspected also for each lake. Figure 8 and Figure 9 show the plots together with the correspondent number of matches. Higher numbers of matches are for lakes where data were available for longer periods but also where hourly/subhourly measurements were available and for sites far from the lake shore The median and robust standard deviation are consistently better for quality level 5 throughout the lakes, while for quality level 4 the apparent greater variation is related to a very low number of matches. For example, for lake ID 8, the numbers of matches are only 7, 7, 9 and 25 for quality levels 2,3,4 and 5 respectively.
Figure 8: Satellite minus in situ temperature difference median per lake (upper plot) and number of matches (lower plot) per quality level. The lake ID is given on the horizontal axis.
Figure 9: Satellite minus in situ temperature difference robust standard deviation per lake (upper plot) and number of matches (lower plot) per quality level
For Lake Superior, where many sites are available, the robust statistics of the difference for all the matches of quality level 3,4,5 have been plotted per sites in Figure 10 and Figure 11, showing consistency for close sites and higher variability close to the lake shore.
Figure 10: Satellite minus in situ temperature difference median for all the sites on Lake Superior for quality level 3,4,5
Figure 11: Satellite minus in situ temperature difference robust standard deviation for all the sites on Lake Superior for quality level 3,4,5
The satellite minus in situ temperature differences across all data has been plotted against the in situ temperature in Figure 12. The best fit trend in difference against in situ temperature is also shown as a straight line. An offset in this line indicates overall relative bias. A slope in this line indicates that the satellite minus in situ difference is different for lower vs. higher temperature relative to the in situ.
Figure 12: Satellite minus in situ temperature difference against in situ temperature per quality level, where the thicker black line is the best fit.
For quality level 5, the slope is -0.017 K K-1, which means that over the 25 K range of lake temperatures in the data the satellite is warmer relative to in situ observations by 0.425 K for the lowest temperatures compared to the warmest temperatures. The slopes for QL 4 and 3 are even smaller (-0.007 and 0.001 K/K respectively).
The time series of the satellite and the in situ temperature together with their difference have been inspected and they are reported here for two “difficult” validation cases. The first is a small lake (lake Erken in Sweden). The second is a site where only daily averages were available.
The location where the in situ measurements have been collected on Lake Erken in Sweden is shown in Figure 13 (red dot). Figure 14 and Figure 15 show the satellite observations and the in situ measurements in 1997 when only ATSR2 was utilised and in 2008 when observations from AATSR and AVHRR-A were used. For both the years the satellite observations follow remarkably well the in situ measurements, which were temporal high frequency measurements.
Figure 13: Location of the in situ measurements (red dot) on lake Erken in Sweden where each dot represents a 1/120 degree resolution cell.
Figure 14: Satellite observations (dots), in situ matches (white dots), in situ measurements (black line), satellite minus in situ T difference for quality levels 4,5 (green line) and climatology (golden line with climatological variability as the yellow band) for lake Erken in Sweden in 1997.
Figure 15: Satellite observations (dots), in situ matches (white dots), in situ measurements (black line), satellite minus in situ temperature difference for quality levels 4,5 (green line) and climatology (golden line) for lake Erken in Sweden in 2018.
Figure 16 shows the satellite observations and the in situ measurements in 2009 for the Sea of Galilee in Israel, where only daily measurements were available. The plot shows an agreement for the few matches on a lake where we assume that probably not many temperature variations take place.
Figure 16: Satellite observations (dots), in situ matches (white dots), in situ measurements (black line), AVHRR-A minus in situ T difference for quality levels 4,5 (green line) and climatology (golden line) for the Sea of Galilee in Israel in 2009.
Figure 17 and Figure 18 show the satellite observations and the in situ measurements in 2018 and 2019 for lake Superior at the sites 01 and 02 shown in Figure 19. The plots show again a similar sharp increase in LSWT in the beginning of August 2018 for both the locations while in 2019 the increase of LSWT was sharper for location 01 than or location 02, highlighting the temperature spatial structure. The high number of matches is due to the fact that two instruments were employed for the satellite measurements.
Figure 17: Satellite observations (dots), in situ matches (white dots), in situ measurements (black line), AVHRR-A minus in situ T difference for quality levels 4,5 (green line) and climatology (golden line) for site 01 on lake Superior in Canada/USA in 2018 and 2019.
Figure 18: Satellite observations (dots), in situ matches (white dots), in situ measurements (black line), AVHRR-A minus in situ T difference for quality levels 4,5 (green line) and climatology (golden line) for site 02 on lake Superior in Canada/USA in 2018 and 2019.
Figure 19: Location of the in situ measurement sites on Lake Superior in Canada/USA.
Validation of the GloboLakes/C3S L3S uncertainty LSWT v4.0
The LSWT uncertainty has been validated comparing the difference satellite minus in situ temperatures and the correspondent LSWT and in situ uncertainties. The following quantity is calculated for each match:
\[ \Delta= \frac{ T_{LSWT}-T_{INSITU}}{\sqrt{\sigma_{LSWT}^2+\sigma_{INSITU}^2+\sigma_{repr}^2}} \]Where T indicates temperature, for LSWT and in situ as indicated in the subscripts. σ means the standard deviation from measurement uncertainty (for LSWT and in situ) and from real differences because of point-to-pixel representativity effects. The in situ measurements uncertainty is not known for the data we have, and we assume it to be σINSITU=0.2 K, a value based from deployment of similar measurement technologies to the ocean. The representativity effect is presently unquantified and we set it to 0 K; neglecting representativity has the tendency to make the LSWT tendency look underestimated. σLSWT2 is context sensitive and is provided in the products, and so varies from match to match.
The distribution of Δ should be a Gaussian distribution with mean equal to 0 and standard deviation equal to 1 when all standard deviations are well estimated. Figure 20 shows the histograms of the uncertainties per quality level where also the fitted Gaussian and the target Gaussian distributions are shown.
For quality level 5, the Gaussian fit has width 3.18, which means that observed differences are more different than expected from the quoted uncertainties. This may be partly because the product uncertainties are underestimated, but could also arise to the degree that lake in situ data (being more diverse) have larger uncertainty than the assumed value (based on experience of ocean observations), and because representativity is neglected. Interpretation of this outcome is therefore currently ambiguous, and research is needed to better understand the in situ uncertainty and representativity effects.
Figure 20. LSWT uncertainty validation per quality level (indicated in legend): histograms of Δ.
Figure 21 presents the corresponding Q-Q (quantile-quantile) plots, which show the values of actual percentiles of the distribution relative to their theoretical values assuming the ideal Gaussian. The points do not lie on a straight line, showing that the differences include more extreme values than would be expected if they truly came from a normal distribution.
Figure 21: LSWT uncertainty validation per quality level: Q-Q plot of Δ.
Missing lakes and number of observations
For 21 target lakes, no LSWT has been obtained, largely due to the fact that they are too small. The lakes are listed in Table 5 together with the estimated maximum distance to land. None of the lakes are feasible because of their small size, except for the lake 1099 in Greenland, which was not successful due to missing initial prior.
Table 5: GloboLakes lakes with no observations in the product
Lake ID | Lake | Country | Max distance to land (km) |
799 | Hawizeh marhes | Iran | 1 |
1099 | Zzzz3 | Greenland | 1.8 |
2371 | Bering | USA | 1.4 |
2525 | Fyordovoye | Russia | 0.5 |
3114 | Loch Ness | United Kingdom | 1.2 |
3575 | Zzzz | Iraq | 1.1 |
6786 | Loch Tay | United Kingdom | 0.9 |
7889 | Lough Melvin | Ireland | 1.3 |
9322 | Sunapee | USA | 1.1 |
11740 | Windermere | United Kingdom | 0.7 |
12943 | Loch Katrine | United Kingdom | 0.8 |
15309 | Rosu | Romania | 1.2 |
16662 | Jijila | Romania | 1.3 |
17329 | Merhei | Romania | 1.4 |
163748 | Bassenthwaite | United Kingdom | 0.6 |
164293 | Ullswater | United Kingdom | 0.6 |
164384 | Derwent Water | United Kingdom | 0.9 |
208447 | Matita | Romania | 1.1 |
208962 | Isac | Romania | 1.3 |
209099 | Puiu | Romania | 1.3 |
An assessment of the number of days between 1995 and 2018, where LSWT are available, has been carried out at the lake centre, defined as the location on the lake most distant from land [Carrea et al, 2015]. Figure 22 shows the number of days with satellite observations in the GloboLakes/C3S LSWT v4.0 dataset of quality levels 2,3,4,5, where the white circles represent the lakes with no data listed in Table 5. The most observed lakes are at mid latitudes in the Northern and Southern Hemisphere. The least observed lake is lake Auburn in the USA with 246 days of observations and the most observed lake is the Salton Sea in USA with 4297 days of observations. The number of days with observations in the product reduces drastically for some lakes if only quality levels 4 and 5 are considered. For example, lake 1904 in Indonesia has only 1 observations of quality level 4,5 and 987 of quality level 2,3. Salton Sea has 3143 observations of quality 4,5 out of the 4297 from all the quality levels.
Figure 22: Map of the number of observations at the lake centre in the GloboLakes/C3S LSWT v4.0 dataset for quality levels 2,3,4,5. A white circle represents no observations. In order to reduce overlapping, the longitude of the lake centre has been shifted where necessary, but the latitude has been maintained.
Application(s) specific assessment
No application(s) specific assessments have been undertaken for the January 2020 version of the C3S lake surface water temperature dataset.
Compliance with user requirements
The requirements for the C3S Lake water levels are described in the Target Requirements and Gap Analysis document [D1].
Table 6: Compliance with user requirements
Property | Target | Achieved |
Spatial coverage | Global | Global: 979 lakes on 4 continents |
Temporal Coverage | > 20 years | > 20 years |
Spatial resolution | 300m | 0.05 degree |
Temporal resolution | Daily | Daily product, effective repeat varies among lakes |
Standard uncertainty | 0.25 K | Varies among lakes |
Stability | 0.01 K yr-1 | No techniques currently available to assess this since stability of comparison data is unknown. |
The user requirements are mostly met. The spatial coverage can be considered as global since the lakes are distributed in all the continents. More lakes could be included to increase the coverage and also to increase the number of smaller lakes which are the most difficult in regards to LSWT retrieval. The spatial resolution is limited by the current satellite resolution for thermal remote sensing which about is 1km. An improvement in the spatial resolution could be brought by regridding into a 0.025 degree regular grid, although at high latitude some gaps may appear with the current regridding algorithm and different techniques may be required. The temporal resolution is limited ultimately by the satellite swath and varies among lakes and among periods (the ATSRs have smaller swath with respect to the AVHRRs). Improvements in the temporal resolution can be brought by the inclusion of other satellites such as SLSTR on Sentinel3. The uncertainty varies among lakes but a big portion of the lakes throughout the years have uncertainties between the target (0.25K) and the threshold (1K) values from GCOS. The validation of the uncertainty can indicate that it is underestimated, although this validation cannot be considered conclusive since the uncertainty of the in situ measurements is unknown.
More details can be found on the in the Target Requirements and Gap Analysis document [D1].
In addition to comments to the user requirements, we strongly recommend the users to utilize the LSWT data together with the quality levels which indicate the confidence in the result. We recommend quality levels 4 and 5 while 3 and 2 can be use with care after a thorough inspection.
Acknowledgements
Thanks to Iestyn R. Woolway who has established the contacts to set up the in situ database. Thanks to all the institutions listed in Table 2 that have provided us with in-situ data, and in particular to:
- Gil Bohrer, The Ohio State University, Columbus, USA
- Jean-Francois Cretaux, LEGOS, Toulouse, France
- Margaret Dix, Universidad del Valle de Guatemala, Guatemala
- Martin Dokulil, Mondsee, Austria
- Hilary Dugan, Center for Limnology, University of Wisconsin-Madison, USA
- Gideon Gal, Yigal Allon Kinneret Limnological Laboratory, Israel Oceanographic and Limnological Research, Migdal, Israel
- Claudia Giardino, Istituto per il Rilavemento Elettromagnetico dell'Ambiente, National Research Council of Italy, Italy
- Johanna Korhonen, SYKE, Helsinki, Finland
- April James, Nipissing University, Canada
- Ilga Kokorite, University of Latvia and Latvian Environmental Geology and Meteorology Centre, Latvia
- Ben Kraemer, Leibniz institute for freshwater ecology and inland fisheries, Berlin, Germany
- Alo Laas, Estonian University of Life Sciences, Tartu, Estonia
- Eric Leibensperger, Center for Earth and Environmental Science, SUNY Plattsburgh, USA
- Junsheng Li, Institute of Remote Sensing and Digital Earth, Chinese Academy of Science, China
- Alessandro Ludovisi, Dipartimento di Biologia Cellulare e Ambientale, Universita degli Studi di Perugia, Italy
- Shin-ichiro Matsuzaki, National Institute for Environmental Studies, Japan
- Linda May, Centre for Ecology and Hydrology, Edinburgh, Scotland UK
- Ghislaine Monet, UMR CARRTEL, Thonon le Bains, France
- Tiina Nogesand Peeter Noges - Estonian University of Life Sciences, Tartu, Estonia
- Sebastiano Piccolroaz, Institute for Marine and Atmospheric Research, Department of Physics, Utrecht University
- Pierre-Denis Plisnier
- Don Pierson, Uppsala University, Sweden
- Antti Raike, SYKE, Helsinki, Finland
- Alon Rimmer, Yigal Allon Kinneret Limnological Laboratory, Israel Oceanographic and Limnological Research, Migdal, Israel
- Geoffrey Schladow, UC-Davis Tahoe Environmental Research Center, USA
- Eugene Silow, Irkutsk State University, Russia
- Lewis Sitoki, Department of Earth Environmental Science and Technology, Technical University of Kenya, Nairobi
- Evangelos Spyrakos, Biological and Environmental Science, University of Stirling, Scotland UK
- Wim Thiery, Department of Earth and Environmental Sciences, KU Leuven, Belgium
- Piet Verburg, NIWA, New Zealand
- Gesa Weyhenmeyer, Department of Ecology and Genetics, Uppsala University, Sweden
- Caroline Wynne, Environmental Protection Agency, Ireland
References
Carrea, L., Embury, O. and Merchant, C. J. (2015) Datasets related to in-land water for limnology and remote sensing applications: distance-to-land, distance-to-water, water-body identifiers and lake-centre co-ordinates. Geoscience Data Journal, 2(2). pp. 83-97. ISSN 2049-6060 doi:10.1002/gdj3.32
Embury, O., Merchant, C. J. and Corlett G.K. (2012) A reprocessing for climate of sea surface temperature from the along-track scanning radiometers: Initial validation, accounting for skin and diurnal variability effects. Remote Sensing of Environment, 116. pp. 62-78. ISSN 0034-4257 doi:10.1016/j.rse.2011.02.028
Saunders, P.M. (1967) The temperature at the ocean-air interface. Journal of the Atmospheric Science, 24. pp. 269-273. doi:0.1175/1520-0469(1967)024<0269:TTATOA>2.0.CO;2