1. Introduction
Temperature, defined as a physical quantity that characterizes the average energy of random molecular motion within a substance [
1,
2], is an essential variable in the energy exchange between the Earth’s surface and the atmosphere [
3,
4,
5,
6,
7,
8]. In meteorology, air temperature at various heights and the temperature of the ground surface can be measured, among others [
1], the latter being the one that corresponds to land surface temperature (LST) obtained from satellites.
Air temperature (
Ta) can also influence the behavior of the active permafrost layer, as described for James Ross Island (JRI)—one of the largest permafrost regions in the northeastern Antarctic Peninsula (PA)—by Hrbáček et al. [
9], who found a significant effect of
Ta on the thermal regime of the soil, especially in the absence of snow and during the advection of warm air masses that cause important gradients in
Ta during winter. In addition, major changes in temperature could lead to changes in the snowpack on the Antarctic ice sheet, which would have both immediate and long-term impacts on the global mean sea level [
10]. Therefore,
Ta constitutes an essential parameter in a wide range of environmental applications, such as hydrology and climate change studies [
11]. Surface temperature, for its part, is the primary climatic factor that governs the existence, spatial distribution and thermal regime of permafrost which is a major component of the terrestrial cryosphere [
12]. In summary, both Ta and LST constitute essential factors for permafrost, which is determined by the climate and the geothermal gradient [
13].
Ta and LST generally present a similar behavior that can be considered as a sinusoidal oscillation with a period of one year. It is also known that ground temperature can delay
Ta variations [
14], therefore it could be considered that this energy exchange can favor the correlation between
Ta and LST. However, it must be taken into account that
Ta and LST are two different variables, and the correlation between them could vary throughout the year because in the cold season the snow acts as an insulating layer [
15]. In addition, it has been described that the correlation also depends on land cover and sky conditions [
16,
17], as well as elevation changes [
18], latitude and solar radiation. On the other hand, land and air have different thermal capacities [
19,
20] and
Ta and LST have different physical meanings and are measured with different techniques: for example, in the case of the Moderate Resolution Imaging Spectroradiometer (MODIS), LST data indicate the land surface temperature in an area of 1 km
2, while, in this work,
Ta refers to the temperature measured 2 m above the ground surface. Also, unlike
Ta, LST provides direct information on long-wave radiation from the Earth’s surface [
19].
However, its different nature has not prevented LST from becoming an important alternative to overcome the lack of
Ta data, generally only obtained at certain points with very limited area coverage [
18]. Likewise, LST has been used to validate in situ data on surface air temperature (
Ts) [
21]. Furthermore, as indicated by Sobrino et al. [
22], it is known that there is a correspondence between LST and
Ta, in general terms—not only with MODIS sensor data—as has been shown for local studies, among others, by the works of Jin and Dickinson [
19,
23], Prihodko and Goward [
24] and Urban et al. [
25].
The possibility of having LST measurements from satellites has favored the development of studies for the estimation of
Ta. The high temporal resolution of MODIS has made it possible to widely develop studies on the temperature of the polar regions [
6]. The utility of applying statistical methods to MODIS sensor data, due to its high temporal resolution, is evident in the numerous works that have estimated
Ta with MODIS LST data in various study areas including Portugal [
11] and the United States [
26]. Also, statistical methods combined with the Temperature-Vegetation Index (TVX) have been used on different sites in Africa [
27]. The above examples highlight the variety of ecosystems that are represented in these studies.
In Antarctica, there have been several works that estimated
Ta from MODIS LST data with statistical methods. Also Meyer et al. (2016) [
28] estimated
Ta values using MODIS data and data from 32 weather stations in Antarctica, adding to LST other predictor variables to three machine learning algorithms: Random Forest (RF), Gradient boosting (GBM) and cubist. In addition, MODIS LST has been used to model the near-surface permafrost temperature [
29] in all ice-free areas in Antarctica and on Antarctic islands, using reanalysis data. Intercomparison studies between MODIS LST and
Ta in situ data have also been carried out; for example, Wang et al. [
30], based on the monthly average values and measurements of
Ta at 1 m and 2 m, in stations in Eastern Antarctica.
Analysis of numerous authors who have studied temperature variability [
31,
32,
33,
34,
35,
36,
37,
38,
39,
40,
41,
42,
43,
44] demonstrates the complexity of Antarctic climatology, on both a spatial and temporal scale, observed on and around the Antarctic Peninsula (AP), it is essential to continue monitoring it with both in situ
Ta and satellite LST data. But the in situ
Ta data, although continuous and precise, are spatially sparse, so the LST could help, after comparing with the
Ta, to extend the spatial coverage of these data, in spite of the known drastic reduction in LST data due to the frequent cloudiness in the regions of the Antarctic. On the other hand, the estimation of
Ta from LST will be more precise as the LST data are more precise, so the improvement of the LST data is essential, both with new methods and new sensors. In this scenario, the aim of this work is to estimate
Ta from daytime and nighttime LST data at Maritime Antarctica sites in the South Shetland Archipelago using empirical models, based on the addition of spatiotemporal variables to LST. The need to use satellites and sensors to estimate in situ data in polar areas is well known, given the impossibility of installing and maintaining a dense network of stations in these areas. To date, the estimation of
Ta from MODIS 6 (C6) collection data has not been performed on Livingston Island. In the case of Deception Island,
Ta has not been estimated with any MODIS collection so far. In addition, this work analyses the emissivity, taking into account that in C6 important changes were introduced with respect to collection 5 (C5), especially in the emissivity adjustment.
Following the Introduction (
Section 1), in Materials and Methods (
Section 2) we describe the study area, MODIS and in situ data, as well as the digital terrain model data and methodology used to compare
Ta with LST. The sections on Results, Discussion and Conclusions are provided in
Section 3,
Section 4 and
Section 5, respectively.
4. Discussion
Previous studies have shown that MODIS LST products can be used to estimate
Ta in Antarctica. With simple linear regression, R
2 ≥ 0.6 in more than 82 % of cases. This overall result is consistent with other studies in the area. For example, Wang et al. [
35] found, under clear sky conditions, a statistically significant correlation (R
2 = 0.41–0.98) between monthly averaged MODIS LST Collection 5 (C5) data and in situ data from eight stations over the Lambert Glacier Basin, East Antarctica, between 2000 and 2009, over a
Ta range of −70 °C to −15 °C. Our R
2 values, taking into account that
Ta and LST are different magnitudes, are in the range of those obtained in similar studies in different regions with MODIS, and also with other sensors, for example: Urban et al. [
25] compared daytime C5 MODIS LSTs and daily mean
Ta from ≈600 stations on the Pan-Arctic Scale (>60°N), and obtained an R
2 = 0.85 and R
2 = 0.64, for a
Ta range from −60 to 0 °C, and from 0 to 30 °C, respectively; these authors also compared LST products from (A)ATSR, MODIS and AVHRR, concluding that MODIS LSTs had the highest agreement with
Ta. Recondo et al. [
66], who obtained an R
2 = 0.87 comparing daytime Terra-MODIS data and daily mean
Ta from 331 stations in peninsula Spain; and Kawashima et al. [
72] obtained R
2 values ≈ 0.76 from Landsat data at several of their stations in the Kanto plain and surrounding mountainous area, in central Japan. Our minimum value is equal to the minimum R
2 obtained by Wang et al. [
30] between daytime MODIS data observed LST and air temperature 2 m above ground level, for Antarctica. Also, our error values are only slightly higher than those obtained by Duan et al. [
56], who found that, with daytime data, MODIS LST C6 was well correlated with in situ LST, with bias < 0.2 K and RMSE < 1.3 K. However, it must be taken into account that these authors compared MODIS LST with LST calculated with in situ data, a magnitude equivalent to MODIS LST; therefore, a better agreement between the two would be expected than that obtained with
Ta, given the differences between these variables [
19,
20]. Additionally, Duan et al. [
56] analyzed data from areas with different cover types. Thus, the proximity of our error values to those obtained by these authors is indicative of the daytime MODIS measurement quality in our study area and of the filters applied.
On the other hand, with nighttime data the fit between MODIS and in situ values is worse. R
2 values are in the range of 0.29 (JG, MOD11A1) to 0.71 (JCI2, MYD11A1) and RSE values are in the range of 1.38 (JG, MOD11A1) to 4.29 °C (GDC2, MOD11A1). This worse fit could be due to the effects of clouds on MODIS LST estimation, since cloud detection error rates are higher at night than those achieved during the day [
73,
74,
75,
76,
77], and undetected clouds introduce errors in MODIS nighttime LST estimate as stated by Zhang et al. [
78]. These failures in MODIS cloud detection algorithm cause the sensor to estimate an LST value that actually corresponds to the temperature of the clouds and not to the Earth’s surface temperature, which leads, according to Westermann et al. [
79], to erroneous measurements, with the consequent bias. In fact, in other polar regions, such as Svalbard, considerably higher error values have been obtained on cloudy days (7 K) than on clear days (3 K) when comparing MODIS LST with
Ts measurements [
75]. This masking effect of the in situ measurement real value had already been described for LST in Siberia [
80] and it has been verified that the same occurs with the albedo in our study area: the cloud mask failure can cause MODIS estimate to correspond to cloud albedo values and not snow albedo values [
81].
Our study area presents high cloudiness, being close to 60°S, where it is estimated that there is cloudiness between 85% and 90% of the year [
82]. Specifically, in the SSI archipelago, abundant cloudiness is related to the dynamic circulation of air masses and atmospheric fronts [
83]. It has also been reported for other areas that MODIS products were largely contaminated by frequent cloud cover [
84,
85]. In our case, it must be taken into account that to minimize the cloudiness effect, our LST data were also filtered using MOD10A1 and MYD10A1 data, given their higher spatial resolution. Furthermore, although these results are consistent with those described for these same stations—except GDC—for the multiple linear regressions of MODIS C5 data, with C5 good results were not obtained in the simple linear regression (R
2 ≤ 0.4) [
64]. Therefore, the results presented here corroborate an improvement in MODIS LST C6 product performance.
Previous studies have suggested that the better results obtained with C6 compared to C5 could be related to the improvements in MODIS emissivity data. Taking into account the importance of emissivity to estimate LST from satellites [
86], we compared the C5 and C6 mean emissivity values of bands 31 and 32, as well as the mean and the difference of both bands for all stations (
Figure 6,
Figure 7,
Figure 8 and
Figure 9), except for GDC, for which we did not have C5 data. We included the mean and the difference because they were both parts of the equation for calculating MODIS LST using the GSW. The emissivity value of each band for each station was calculated, following the criteria of previous studies [
56], as the average of the emissivity values for the entire study period using the dates with data after applying all of the filters described in
Section 2.3.
As is well known, emissivity varies with wavelength or soil type, although it has also been suggested that soil moisture or surface viewing geometry may influence it [
87]. Specifically, on snow-covered surfaces, the emissivity varies with depth, density and grain size [
51,
88]. The behavior described by Salisbury et al. [
89] was that the snow emissivity decreased both with increasing particle size and with increasing snow density, while melting contributed to an increase in emissivity. However, it is very difficult to analyze the effects of grain size change on snow emissivity, since such information is not available on a large scale [
90].
Figure 6 shows the mean emissivity values of bands 31 and 32, as well as the mean and the difference of both bands for each station for daytime Terra data with C5 and C6. The values of bands 31 and 32 are in the same range in both collections (band 31: between 0.989 and 0.992 and band 32: between 0.987 and 0.989). In the case of the means of both bands, the behaviors of C5 and C6 are also very similar: in INC2 and JCI2, both collections present the same value (0.988) and in the rest of the stations the C5 values are far from those of C6 only in |0.001|. Regarding the differences (band 31–band 32), the results of both collections are very close for all stations, especially for CR, SOF and JG, where C5 and C6 are separated by 0.0002, 0.0003 and −0.0004, respectively. If we examine R
2 values of Terra C6 daytime data for the different stations on Livingston Island, we find that the best agreement occurs in JCI2 (R
2 = 0.78), followed by CR (R
2 = 0.77), INC2 (R
2 = 0.73), SOF (R
2 = 0.71), JCI1 (R
2 = 0.68) and JG (R
2 = 0.64). On the other hand, with C5 values, the simple linear regression between MODIS LST and
Ta of each station shows a low R
2 for all the stations (R
2 ≤ 0.4), values that, at least taking these data into account, would not be justified by problems only related to emissivity in C5.
Figure 7 shows C5 and C6 emissivity values with Terra nighttime data. The values from both collections are still very close. For example, in the case of the band 32 values (to the right, top, in the figure) and the band 31 and 32 mean values (to the left, bottom, in the figure) the differences between the emissivity values of both collections for all stations are in the range of 0 to |0.001|, except for station GJ, where it is 0.002. Emissivity values with the Aqua data are shown in
Figure 8. The difference between the values of bands 31 and 32 (to the right, bottom, in the figure) is in the range of 0 (CR and SOF) to −0.002 (JCI1). In addition, when analyzing the mean of both bands (to the left, below, in the figure) we find that at all the stations C5 and C6 values are separated by |0.001|, except for JCI1, where the C6 mean values exceed that of C5 by 0.003. Finally,
Figure 9 shows the results for the Aqua nighttime data, where C5 and C6 emissivity behavior is also similar. For example, the mean values of both bands (to the left, bottom, in the figure) differ in this case in both collections in the range 0 to |0.001|, except in JG, where the value of C5 exceeds that of C6 by 0.004. In all cases, the JG value is not representative due to the lack of nighttime Aqua data at this station.
As can be seen, the discrepancies in the emissivity values are not conclusive to explain the better or worse correlation of MODIS LST with respect to the in situ data, either separately from the values of bands 31 and 32 nor from their means and differences. Although Wan [
52] indicated that the GSW algorithm is more sensitive to the change in the difference in emissivity values in bands 31 and 32 than to the change in their mean, this behavior is not evident in our data, because although the correspondence with the in situ data is worse with the C5 data, there are no big differences in the emissivity behavior for both collections. It has been described that the highest and most unreliable uncertainties in emissivity are found beyond the 10.50 µm to 12.50 µm atmospheric window that MODIS uses for GSW algorithm retrievals [
91]. This allows us to state that, for our data, the emissivity of the surface is not the main factor that generates the discrepancy between the MODIS LST of C5 and C6, which is consistent with what was described by Duan et al. [
56].
The stations where MODIS achieves the best agreement with the in situ data are, in this order, GDC2 (R
2 = 0.82, with Terra daytime data), CR (R
2 = 0.81, with Terra and Aqua daytime data) and JCI2 (R
2 = 0.80, using Terra daytime data). In these stations, R
2 values contrast with those obtained for the study area with data from C5, where the R
2 average is 0.53 (with daytime data), being even lower with nighttime data (R
2 average = 0.49) [
64]. Our results are also superior to those obtained in other studies estimating
Ta from MODIS LST for Antarctica, such as Meyer et al. [
28], who also used other variables besides to LST, with data from 32 stations distributed in continental Antarctica (R
2 = 0.78) and in two of the stations studied by Wang et al. [
30] in the Lambert Glacier Basin in East Antarctica using MODIS daytime LST data (LGB00 and LGB20, with R
2 = 0.62 and R
2 = 0.74, respectively).
With respect to cross-validation, as in the case of C5 data [
64], the worst results are also obtained with C6 with the nighttime data. The problems that MODIS presents in cloud detection have already been commented on in previous sections. The validation of the two unique models confirms that our results are within the range of values obtained with the MODIS LST product assessment for Antarctica and other study areas. Thus, for example, if we compare our results with the model obtained by Shi et al. [
92] using LST MODIS data at nine sites around the world—America (3), Europe (1), Africa (1), Oceania (2) and Asia (2)—we can see that while our R
2CV average values (0.69 for the model that excludes the GDC2 station and 0.68 including it) are lower than the average value obtained in these nine regions (R
2 = 0.93), these authors reported higher errors (MAE = 2.1 °C, RMSE = 2.7 °C) than those obtained in our study in the cross-validation with daytime data (MAE
CV = 1.60 °C and 2.00 °C, RMSE
CV = 2.19 °C and 2.61 °C, in the unique model without GDC2; MAE
CV = 1.61 °C and 1.99 °C, RMSE
CV = 2.18 °C and 2.59 °C, in the unique model with GDC2). MODIS LST products have also been evaluated, for example, in Northeast China by Yang et al. [
93], who obtained the following results with their best model for
Ta average: R
2 = 0.94, RMSE = 3.60 °C and MAE = 2.80 °C. As can be seen, although R
2CV values obtained here are lower than those obtained in other study areas, cross-validation yields smaller errors, which is indicative of the performance of the models we evaluated.
Specifically in polar areas, in the Arctic, Westermann et al. [
79] found deviations from weekly averages with MODIS LST data versus Svalbard in situ data to be less than 2 K, which is consistent with the MAE
CV average value obtained here for all data (1.99 ° C and 2.03 °C for the models without GDC2 and with GDC2, respectively), and especially excepting Aqua nighttime data (1.9 °C for both models). Similarly, our results yield a RMSE
CV average (2.59 °C and 2.63 °C, in the unique model, without GDC2, and including it, respectively), which is consistent with the 3 K error between MODIS LST and temperature of surface measured in situ during the period between 2002 and 2009, also in Svalbard by Westermann et al. [
94]. Note that, on the one hand, higher errors would be expected in our study area, given the differences between
Ta and surface temperature, and that, furthermore, we could expect even worse agreements given the range of temperatures in SSI, since the performance of MODIS measurements degrades as LST values increase and approach 0 °C [
94].
Specifically in the studies on Antarctica, Meyer et al. [
28] obtained as best results an R
2CV value of 0.71 and an RMSE
CV of 10.51 °C, with
Ta data from 32 stations distributed in continental Antarctica. In the case of R
2CV, although its results are slightly higher than those obtained here for all data (R
2CV average = 0.69 and 0.68, in the unique model without GDC2 and with that station, respectively), our daytime data show better agreement in both cases (R
2CV = 0.75 and 0.73, Terra and Aqua, respectively). On the other hand, our errors, calculated from RMSE
CV are significantly lower, since they are in a range between 2 °C and 3 °C, except with Aqua nighttime data, which, however, do not exceed 3.13 °C. Regarding the MAE
CV, our errors do not exceed 2 °C, except with Terra nighttime data in the model without GDC2 and with Terra and Aqua nighttime data in the model with GDC2. These error values are also below the 3.4 °C obtained by Wang et al. [
30] from the mean of the standard deviation of the differences between LST MODIS and
Ta with data from the Lambert Glacier Basin. Another factor that should be taken into account when evaluating the performance of the models obtained here is that MODIS LST daily products provide data with a spatial resolution of 1 km. At both the Livingston and Deception Islands there is the advantage that at most sites there are no pixel coverage differences, thus minimizing the problem of surface coverage heterogeneity that has been noted as a potential biasing factor for MODIS LST studies in permafrost areas [
79].