The L3 pan-Arctic products (referred to here onward as UW-L3) are generated from AATSR unprojected, ATS_NR_2P available through ESA’s online archive MERCI [40]. Some L2 AATSR (ATS_NR_2P) products were found to suffer from an error consist of a sequence of identical temperature regions. These images were identified and excluded from the final products calculation. In addition, images that have less than 100 pixels were excluded and flagged as “insufficient” data files. The UW-L3 products of MODIS are generated using Collection 5 L2 data acquired by both the Terra (MOD11_L2.5) and Aqua (MYD11_L2.5) satellites (Terra only 2000–2002; Terra and Aqua, 2002–2010) [41]. Although, MODIS quality information indicates the probability of “unseen” thin clouds, we decided to include all MODIS observations that labeled clear sky by MODIS cloud mask. The spatial averaging over 25 km pixels (625 L2 observations per scan) during a month is likely to counterbalance the effect of MODIS cloud mask failures. However, we did evaluate this assumption using the MODIS products of year 2009.

We selected the Northern Hemisphere EASE-grid, Lambert’s Equal Area Azimuthal projection, based on a sphere datum with a radius of 6,371.228 km, as the standard projection for the UW-L3 products. This projection was selected to match ESA’s GlobSnow archive [42], snow and ice datasets available from the National Snow and Ice Data Center (NSIDC) [43] and DUE Permafrost products [18], therefore, facilitating future combined use of snow and ice products with UW-L3 LST products over the Arctic. The initial grid was assigned a 1-km spacing similar to the observation intervals of L2 AATSR and MODIS products. Bilinear interpolation was used to interpolate irregular L2 observations to the center of the EASE-grid pixels (Figure 1). Local time for each EASE-grid pixel is calculated using UTC acquisition time and longitude extracted from the Annotation Data Set (ADS) information in the case of AATSR data, and from file name in case of the L2 MODIS products. This method is limited to an accuracy of ±15 minutes, which is sufficient for the creation of weekly and monthly products.

The pan-Arctic UW-L3 products are generated through aggregation of interpolated L2 1-km data into a 25-km EASE grid. In the current version of the processing chain, spatial aggregation requires a minimum of 32 1-km observations (5% of total observations), which was found to be reasonable for estimating a clear-sky LST mean. Pixels with less than 32 1-km observations are discarded (assigned a “no data” value). Valid aggregated pixels are separated into either a day-time bin (from 6 am to 6 pm local time) or a night-time bin (from 6 pm to 6 am of the next day). The definition of day and night is not determined by the solar angle (i.e., number of hours of daylight and darkness). Day (night) LST mean, is calculated as an arithmetic average of all pixels that fall in the day (night) bin during the week or month of interest. The final LST products are created by calculating a mid-range (balanced) mean of the day and night bins for the time period of interest in order to ensure that the calculated LST values are not biased towards the time period of the day with the greatest number of observations. Day average, night average, daily average and the number of clear-sky L2 observations (counts) are recorded in separate images for the period of interest (Figure 1).

The calendar month is used for calculating monthly UW-L3 products. For the weekly products, a sliding window (average of seven days) consisting of the date of interest and the six previous days is used. This follows the same convention adopted by the GlobSnow project for the creation of 25-km weekly snow water equivalent (SWE) products. The mean annual surface temperature (MAST) is then calculated as the average of the twelve UW-L3 monthly mean products per year for computational efficiency.

Validating coarse-resolution (km to tens of km) pan-Arctic products is a challenging task [44]. Ground-based LST measurements are limited spatially. In addition, up-scaling of ground measurements (e.g., thermal radiometer measurements) is also challenging due to surface heterogeneity and the number of temperature instruments that can be deployed in remote high-latitude regions. An alternate approach for the evaluation of coarse resolution satellite products is through intercomparison with existing products of the same or similar physical quantities [45,46]. Uncertainty in UW-L3 pan-Arctic (monthly and annual) products was therefore estimated by intercomparing UW-L3 of AATSR and MODIS to each other and against existing surface and near-surface temperature products derived from the Advanced Microwave Scanning Radiometer for EOS (AMSR-E), the Special Sensor Microwave/Imager (SSM/I), and from North American Regional Reanalysis (NARR). Products (summarized in Table 3) were projected to the exact EASE-grid projection with 25-km spacing to match that of the UW-L3 products. Interpolation of NARR 32-km data to the 25-km EASE-grid resulted in some contaminated pixels along shorelines of oceans and that of large lakes. A binary mask was therefore applied to remove all pixels within 25-km of the shorelines.

In order to explore the effect of including only clear-sky LST observations in the generation of UW-L3 products, in contrast to AMSR-E Ta, SSM/I and NAAR products that include all-sky (clearand cloudy-sky) observations, we first identified cloudy-sky observations at daily intervals in UW-L3 using the original cloud flags of L2 MODIS. The identified cloudy days were then used to produce a mask and applied to LST products from passive microwave (AMSR-E and SSM/I) and reanalysis (NARR) products to create clear-sky only monthly averages of these products.

UW-L3 weekly, monthly and annual products were evaluated against the AMSR-E Ta, SSM/I, and NARR temperature datasets by calculating the mean difference (MD) and root-mean-square-difference (RMSD) statistics. MD is an uncertainty measure that takes into account the direction of the difference, whether positive or negative, while RMSD is a measure that is sensitive to outliers and considers the magnitude of the difference without considering the sign. Both uncertainty measures were calculated by averaging the difference obtained at all pixels between the different datasets on corresponding time periods. We deliberately use the term difference rather than error to indicate that each dataset contains its own level of error. We prefer to reserve the term “error” as used when validating satellite-derived LST with LST measured with a thermal radiometer deployed in the field. Such values are reported in Tables 1 and 2, and reported in corresponding Sections 2 and 3 for MODIS and AATSR, respectively. Differences of UW-L3 products with AMSR-E Ta, SSM/I and NARR were calculated over North America above 50 degrees latitude, while comparison between UW-L3 of MODIS and AATSR products was possible for all pixels included in the entire Arctic region.

Examples of UW-L3 products are presented in Figure 2. Corresponding UW-L3 products of AATSR and MODIS were found to follow similar LST patterns. Yet, LST products of MODIS were more spatially and temporally complete than those of AATSR (Figure 2) because the number of L2 MODIS observations is approximately 10-fold that of AATSR. MODIS sensors have the advantages of being onboard of two satellite platforms, Aqua and Terra, in addition to having a larger swath than AATSR. Although at the moment the density of AATSR is low, in the future, the SLSTR (Sea and Land Surface Temperature Radiometer), which is a successor of AATSR, is to be included on the payload of the twin Sentinel-3 satellites. The satellite pair will enable a short revisit time of less than one day, which will restore the ratio between MODIS and AATSR / SLSTR observations.

The maximum count of L2 MODIS observations was found in the April to October period. Unexpectedly, this period is known to have the highest cloud cover in the high and low Arctic [52]. The minimum count occurred during the period from October to March (Figure 3), which is characterized by the existence of snow on ground. Despite the high cloud fraction during the snow-free period (June–August), clouds can be easily detected during summer because of the large difference in reflectance between land and cloud. In contrast, snow-cover interferes with the cloud masks [53,54] and causes more L2 observations to be rejected, and therefore the total number of observations to drop.

AATSR observations were found to drop significantly during the snow-free period, suggesting that AATSR cloud mask tends to reject more observations in summer (Figure 3). The total number of L2 AATSR observations increased during the year 2006 compared to the average number of L2 observation found between years 2005–2009. The most likely cause of the “false” clear-sky observations was the development of a thin coating film on the visible channel calibration. This thin film caused a drift and poor performance of the sensor and higher probability of cloud contamination during that year [55].

A sudden drop in LST was identified around 60°N in Eurasia and North America during the cold months, particularly evident during the months of December and January (Figure 4). The drop in LST was found to correspond to a significant loss of available L2 MODIS observations [56]. This problem has been traced back to the input L2 cloud mask (MOD35 product) resulting likely from use of latitude dependent algorithm paths to retrieve the cloud mask [57]. The issue is currently being addressed in preparation for the release of Collection 6 data expected in summer 2013 [58].

Comparison between weekly UW-L3 products of MODIS and AATSR of year 2008 indicated an annual average MD of −1.8 K (Min: −4.4 K, Max: 0.4 K) and an annual average RMSD of 4 K (Min: 2.7 K to Max: 6.1 K). Comparison of monthly products for the period 2000 to 2010 (Figure 5(a)) indicated a similar bias range of an average MD of −1.1 K (Min: −4.3 K, Max: 1.8 K) and an average RMSD of 3.8 K (Min: 2.2 K, Max: 5.3 K). Day-time averages were found to be more different (MD: −0.1 K, RMSD: 5.1 K) than night-time averages (MD: −2 K, RMSD: 3.2 K). Comparing annual products of both sensors over the 2005–2009 period showed the lowest difference with an average MD of −1.1 K (Min: −1.8 K, Max: 0.1 K) and an average RMSD of 2.1 K (Min: 1.6 K, Max: 2.8 K). The reported bias between the two sensors (AATSR and MODIS) can be attributed to differences in retrieval algorithms and density of observations (Figure 5(b)).

Differences between monthly UW-L3 of MODIS and NARR LST (all-sky) show an average MD of −3.6 K (Min: −6.8 K, Max: −1.0 K) and an average RMSD of 4.8 K (Min: 2.5 K, Max: 7.7 K). The results of comparing mean annual LST products estimated from MODIS and NARR over North America indicated similar bias range with a MD of −3.6 K (Min: −4.2 K, Max: −3.1 K) and a RMSD of 4.2 K (Min: 3.7 K, Max: 4.6 K). The average bias between MODIS and NARR is entirely negative. In contrast, the observed bias between monthly UW-L3 of AATSR products and NARR had positive and negative magnitudes with a MD of −2.5 K (Min: −5.8 K, Max: 2.4 K) and a RMSD of 5.1 K (Min: 2.9, Max: 7.2 K). The positive difference during summer (AATSR warmer) could be traced to overestimating LST by L2 AATSR observations [38].

The bias between monthly UW-L3 of MODIS and AATSR products and NARR (all-sky), depicted in Figure 6, is systematic during the period 2000–2010 with a maximum difference around mid-winter and a minimum around July/August of every year. However, the maximum cloud coverage during the warm period (June–August) corresponds to the minimum difference. Remarkably, the systematic bias pattern is opposite to the interannual variation of cloud cover at high latitudes (see Figure 3). The observed negative difference during the cold months (snow on ground season) suggests that the admixing of top-of-the-cloud temperatures influences the winter LST estimates more than the summer. The reported bias is not absolute. The accuracy of LST (NARR) is influenced by how accurate the land surface scheme utilized in NARR captures energy partition in higher latitudes. In addition, the number of assimilated weather stations in higher latitudes is smaller than in mid latitudes. For example, it has been reported that NARR data tend to be warmer than LST derived from SSM/I data [49].

Different sources of bias could influence UWL3 products. We will discuss two major types of errors: the bias towards clear-sky average and the erroneous observations of undetected clouds. The bias towards clear-sky days will result in a warmer (colder) LST average during winter (summer) than the average LST for all-sky conditions (true LST mean). The cloud cover increases LST during winter by re-emitting long-wave radiation to the earth’s surface, while decreasing LST during summer by blocking direct solar radiation (increasing reflection back to space). In contrast, the bias introduced by erroneous observations of undetected clouds will always result in a negative offset of average LST because top-of-the-cloud temperatures are much colder than that of the land surface regardless of the season.

The results (Figure 7) clearly show that contamination with top-of-the-cloud temperatures (as indicated by applying a conservative quality control filter) is reduced in monthly LST averages, but it is not the main source of bias. Using a less restrictive quality control filter resulted in a negative bias between 0.5 and 1 K (relative to NARR). The bias towards clear-sky days contributed much more to the overall error, as can be seen by comparing the left and right panels of Figure 7. Adding to that, the summer bias between MODIS and NARR all-sky (Figure 7(b)) was found to be less than the summer bias between MODIS and NARR clear-sky (Figure 7(a)), likely because the positive summer bias (toward warm clear-sky days) countered the apparent negative bias during the warm months. Remarkably, the sinusoidal bias structure dominates the bias curves in both left and right figures indicating that the source of that oscillating bias is independent from errors resulting from the top-of-the-cloud admixing and from the bias towards clear-sky days. With regard to the spatial structure of error, the patterns of bias between MODIS and NARR (clear-sky vs. all-sky) during the month of July are not different; however, the difference in the severity of the bias is clear during the month of January (Figure 8).