Assessment of MODIS Surface Temperature Products of Greenland Ice Sheet Using In-Situ Measurements

Abstract

Satellite-based data have promoted the research progress in polar regions under global climate change, meanwhile the uncertainties and limitations of satellite-derived surface temperatures are widely discussed over Greenland. This study validated the accuracy of ice surface temperature (IST) from the moderate-resolution imaging spectroradiometer (MODIS) over the Greenland ice sheet (GrIS). Daily MODIS IST was validated against the observational surface temperature from 24 automatic weather stations (AWSs) using the mean bias (MB), the root mean square (RMSE), and the correlation coefficient (R). The temporal and spatial variability over the GrIS spanning from March 2000 to December 2019 and the IST melt threshold (−1 °C) were analyzed. Generally, the MODIS IST was underestimated by an average of −2.68 °C compared to AWSs, with cold bias mainly occurring in winter. Spatially, the R and RMSE performed the better accuracy of MODIS IST on the northwest, northeast, and central part of the GrIS. Furthermore, the mean IST is mainly concentrated between −20 °C and −10 °C in summer while between −50 °C and −30 °C in winter. The largest positive IST anomalies (exceeds 3 °C) occurred in southwestern GrIS during 2010. IST shows the positive trends mainly in spring and summer and negative in autumn and winter.

1. Introduction

The Greenland ice sheet (GrIS) covers an area of 1.7 × 10⁶ km² and is the second largest ice sheet in the world after the Antarctic ice sheet [1]. The GrIS had been relatively stable during the 1970s, 1980s, and the early 1990s with the accumulation and loss staying relatively balanced [2,3]. In recent decades, there has been much attention paid to the GrIS due to its enhanced melting [4,5,6], particularly relevant to abnormal climatic conditions [7,8,9] and accelerated mass loss [10,11,12,13]. The enhanced ablation was measured using infrared and passive-microwave records [14,15,16,17,18]. The analysis of in-situ observations, satellite data, and regional climate models (RCMs) all revealed the accelerating GrIS mass loss [10,19]. The GrIS mass balance has decreased more than twice since the 21st century comparing with the last century [20,21]. The mass loss of GrIS over the past decade is about six times that of the 1980s and the GrIS has contributed 13.7 ± 1.1 mm to the global sea level rise since 1972 [2]. If the acceleration continues, several studies using numerical model strategies to predict the future evolution of GrIS suggested that GrIS could contribute to a rise in global sea level of 2.3 cm by 2050 [22], and 3.5 cm to 76.4 cm by 2300 [23,24]. Surface temperature is an important factor in controlling mass balance [25]. Recently, the increase in GrIS mass loss is mainly because of the increased surface meltwater runoff [26], and prolonged surface temperature elevate exacerbates melting [27]. Surface temperature affects the basal melt [28] as its rise is an important step in the melt process [29]. Thus, tracking IST can better account for surface melt, mass balance as well as ice sheet surface processes.

Currently, observational stations, regional climate models, and remote sensing are the three main sources to obtain IST for studying [30,31,32,33]. The AWSs on the ice sheet provide observations of the GrIS, for example, the Greenland Climate Network (GC-Net) [1,34] and the Greenland Ice Sheet Monitoring Program (PROMICE), which monitors surface and air temperature [35,36,37]. Meanwhile, existing coastal and low-density weather stations on the ice sheet make it hard to get accurate surface temperature observations of the entire GrIS. The difficulties of manipulating station devices in severe polar regions also make it hard to obtain continuous and complete observations record. RCMs can compensate for data recorded at certain periods and locations which could not be obtained from weather stations, especially in the polar regions [38]. RCMs are reliable in the entire GrIS context, however, in the regional and local context, climate variability can be missed occasionally [39,40]. Hence, the way to measure the surface temperature through satellite remote sensing is more outstanding than AWSs and RCMs with high temporal and spatial resolution over the GrIS. Nonetheless, the most practical way to get a wide range of surface temperature data is through the satellite remote sensing [41]. Accurate determination of satellite-derived IST will improve modeling of ice sheet melt and other ice sheet processes and changes.

The GrIS is sensitive to climate change and plays a vital role in the global climate [1]. Hall showed an increase IST of ~0.55 ± 0.44 °C/decade and the northeastern GrIS experienced the largest increase of ~0.95 ± 0.44 °C/decade [8]. Under these conditions, detailed and accurate monitoring of the surface temperature on the entire GrIS is of significance. Recently, a multilayer, daily IST-albedo-water vapor product of Greenland, spanning from March 2000 to December 2019, has been developed. This paper mainly validated the accuracy of the MODIS IST products based on 24 AWSs from PROMICE. A description of the data and methods used in this study were given in 2. In 3, the IST performance was first demonstrated, and then based on its favorable performance, the temporal and spatial variability of GrIS IST, IST anomalies, and trends were further reproduced. Finally, in 4, the MODIS IST melt threshold −1 °C using daily ablation data from 8 AWSs was also analyzed.

2. Materials and Methods

2.1. Observational Data

The 24 AWSs (Figure 1) data used in this study are from the PROMICE, which is operated by the Geological Survey of Denmark and Greenland (GEUS) in collaboration with the National Space Institute at the Technical University of Denmark [35]. The time series of the AWS has different observation periods and completeness (

Table 1. Geographical characteristics of the meteorological stations used in this study.

Station Name	Latitude (°N)	Longitude (°W)	Elevation (m)	Elevation-ELA (m)	Start Date
KPC_L *	79.91	24.08	370	−707	17 July 2008
KPC_U	79.83	25.17	870	−207	17 July 2008
EGP	75.62	35.97	2660	1583	1 May 2016
SCO_L	72.22	26.82	460	−1192	21 July 2008
SCO_U	72.39	27.23	970	−682	21 July 2008
MIT	65.69	37.83	440	−970	3 May 2009
TAS_L	65.64	38.90	250	−1160	28 August 2007
TAS_U	65.70	38.87	570	−840	15 August 2007
TAS_A	65.78	38.90	890	−520	23 August 2013
QAS_L	61.03	46.85	280	−1190	24 August 2007
QAS_M	61.10	46.83	630	−840	11 August 2016
QAS_U	61.18	46.82	900	−570	7 August 2008
QAS_A	61.24	46.73	1000	−470	20 August 2012
NUK_L	64.48	49.54	530	−885	20 August 2007
NUK_U	64.51	49.27	1120	−295	20 August 2007
NUK_N	64.95	49.89	920	−495	25 July 2010
KAN_L	67.10	49.95	670	−745	1 September 2008
KAN_M	67.07	48.84	1270	−145	2 September 2008
KAN_U	67.00	47.03	1840	425	4 April 2009
UPE_L	72.89	54.30	220	−843	17 August 2009
UPE_U	72.89	53.58	940	−123	17 August 2009
THU_L	76.40	68.27	570	−493	9 August 2010
THU_U	76.42	68.15	760	−303	9 August 2010
CEN	77.17	61.11	1880	817	23 May 2017

* L: Lower station, M: Middle station, U: Upper station.

). In this study, the daily IST data was derived from measured downward and upward longwave irradiance and surface emissivity is set to 0.97 [42]. Mostly located in the coastal area of the GrIS, the spatial layout of the AWSs is uneven. Roughly 25% of the stations are located below 500 m and only 16.7% are above 1000 m (Table 1). Moreover, only EGP, KAN_U, and CEN are located at the accumulation area according to the ELA (equilibrium line altitude) of the corresponding basin in 2017 [43].

2.2. MODIS IST Products

MODIS is a polar-orbiting, 36-channel, and across-track scanning spectro-radiometer whose images cover the whole region of the planet every one to two days [44]. MODIS has been flying aboard NASA Earth Observation System (EOS) Terra satellites since 2000 and Aqua since 2002 [45], so the GrIS swath-based daily gridded products began to be available at the beginning of 2000. Terra’s orbit around the Earth is timed so that it passes across Greenland in the afternoon (14:30–16:30 UTC) [41]. MODIS Collection 6.1 IST used in this study is from multilayer, daily IST-albedo-water vapor products developed using standard MODIS datasets from the Terra satellite [28].

Its algorithm was derived from the algorithm developed for the MODIS sea ice product MOD29 or MYD29 (MOD and MYD stand for Terra and Aqua products, respectively) [45,46,47]. The “MOD29” IST products during March 2000~December 2019, provide daily and monthly mean ISTs with polar-stereographic grids at 0.78 × 0.78 km resolution [28]. As the input product of the MOD29 algorithm, MOD35 determines the cloud obscuration of the MODIS IST products. The cloudy grid cells without IST value are excluded calculation process.

2.3. MODIS Products Preprocessing

The cloud coverage affects the accuracy of MODIS IST products. The internal cloud cover incorrectly identified instances of the ice surface as cloudless on a lot of days in JJA (June, July, and August) [8]. Therefore, the ISTs are actually the cloud top temperatures, far below the clear surface temperature (up to 30 °C) [8]. Therefore, we checked the MODSI IST of each grid over two standard deviations to reduce outliers caused by cloud masks or other reasons. The percentage of the removed pixels was shown in

Table 2. The percentage of the MODIS IST pixels that had been removed.

Year	Percentage (%)	Year	Percentage (%)
2000	0.56	2010	0.46
2001	0.55	2011	0.49
2002	0.52	2012	0.48
2003	0.52	2013	0.50
2004	0.52	2014	0.50
2005	0.49	2015	0.50
2006	0.48	2016	0.47
2007	0.50	2017	0.48
2008	0.50	2018	0.48
2009	0.48	2019	0.45

.

As the ice/snow surface cannot exist in the form of ice/snow above 0 °C [8], the fake pixels values (≥0 °C) found on the ice sheet are removed. The most likely reason for a few pixels with the wrong IST is incorrect cloud masking caused by MODIS cloud mask [48]. Additionally, some pixel values representing no data, cloud cover, and fill value, are also removed.

2.4. Statistical Indexes for MODIS IST Assessment

The daily MODIS ISTs used to make a comparison with the observations were extracted respectively from the nearest gridded MODIS IST at the ice sheet. Meanwhile, three classical statistical indexes were used to estimate the performance of the MODIS product, including the bias, root mean square (RMSE), and the correlation coefficient (R). The calculation formulas are as follows:

$$\mathrm{Bias}=\frac{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}({\mathrm{M}}_{\mathrm{i}}-{\mathrm{O}}_{\mathrm{i}})}}{\mathrm{N}}$$

(1)

$$\mathrm{RMSE}=\sqrt{\frac{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}({\mathrm{O}}_{\mathrm{i}}-{\mathrm{M}}_{\mathrm{i}}{)}^2}}{\mathrm{N}}}$$

(2)

$$r=\frac{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}({\mathrm{O}}_{\mathrm{i}}-\overline{\mathrm{O}})({\mathrm{M}}_{\mathrm{i}}-\overline{\mathrm{M}})}}{\sqrt{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}{({\mathrm{O}}_{\mathrm{i}}-\overline{\mathrm{O}})}^2}}\sqrt{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}{({\mathrm{M}}_{\mathrm{i}}-\overline{\mathrm{M}})}^2}}}$$

(3)

where N is the available number of the sample, O is the observed IST, and M is the corresponding MODIS IST.

2.5. Anomaly and Trend Analysis

As the MODIS IST dataset of the GrIS used in this study is available since March 2000, the annual anomaly of 2000 was not calculated. For each grid, the annual IST anomalies were calculated using the monthly data, relative to the multi-year average annual IST from 2001 to 2019. The summer IST includes the year 2000. When seasonal and annual IST trends of each grid were calculated, only the grids with 20-year data are used to calculate.

3. Results

3.1. MODIS IST Performance

To initially obtain the performance of MODIS IST, 12 AWSs are selected according to their location and relative data volume. Figure 2 shows that the MODIS IST agrees well with the observations, with the R > 0.84 (TAS_ L and QAS_L). A significant characteristic is MODIS IST has a cold deviation of 2–3 °C. MODIS IST shows the best agreement with KPC_U AWS observation, with a low RMSE (RMSE = 3.71 °C) and a high correlation (R = 0.97), whereas, generally poorer agreement with TAS_L and QAS_L.

Figure 3 also suggests that the bias indicates the IST tends to underestimate surface temperature (2–4 °C). The R and RMSE indicate that MODIS IST is highly consistent with ground measurements. Spatially, the performance of MODIS IST in the north is better than that in the south, with a higher average R (0.96) and lower average RMSE (4.19 °C). Compared with other regions, the R and RMSE show a better accuracy of MODIS IST on the northwest, northeast, and central GrIS. In these areas, the R is close to 1 and the range of RMSE is concentrated at 3.7–4.9 °C, indicating a high correlation. The MODIS IST shows the best agreement with observations in the northeast, with the highest average R (0.96) and the lowest average RMSE (3.93 °C), however, generally poorer agreement with AWSs observations in the southeast.

Figure 4 compares MODIS IST time series with AWSs to analyze the typical temporal variation and (in) consistency of time series derived from MODIS IST. KPC_U, KAN_U, and EGP were selected for their high-continuity data. The MODIS IST performance is good, and the time series are consistent well with observations, only with a slight underestimation in summer. Notably, the cold bias of MODIS IST mainly occurs in winter. The average cold deviations in summer and winter are 2.04 °C and 4.35 °C, respectively.

Overall, the performance of MODIS products is good, and the range of the cold bias is concentrated at 2–4 °C. A cold deviation of 0.98 °C was found when MODIS IST was compared with the observation data measured by Summit Station of Greenland [28]. The results of thermochron data on the Summit also showed a cold deviation of 3.14 °C of MODIS IST [45]. Given the limited coverage of AWSs on GrIS, we suggest that ISTs obtained from the MODIS IST products are characterized by high accuracy and could be reliably used as an alternative or supplement to Greenland temperature monitoring, especially for IST in summer. Generally, fine space-based climate information obtained by analyzing MODIS IST can effectively evaluate the temporal and spatial variability of the GrIS and be applied in many fields.

3.2. Temporal and Spatial Variability of GrIS IST

Figure 5 are clear-sky IST maps illustrating the MODIS IST on the GrIS at seasonal and annual scales. The MODIS IST revealed distinct spatial gradients varying from outer (low elevation) to inner (high elevation), showing a colder inside with higher elevation. Apparently, the IST in the northern GrIS is much lower than that in the southern region while there is little difference between the eastern and western GrIS. The mean autumn IST is a little colder than the spring and annual mainly reflected in the inner of the GrIS. The mean summer IST is mainly concentrated between −20 °C and −10 °C with the highest IST (−1.33 °C) in the southwest margin of the GrIS and the lowest IST (−17.98 °C) in the inner. Conversely, the mean winter IST is mainly concentrated between −50 °C and −30 °C with the highest and lowest IST values of −9.85 °C and −50.13 °C, respectively. The minimum IST values in winter are found at high elevations just as in summer but tend to migrate farther north, indicating that the coldest ISTs are not always located at the highest elevations, consistent with the previous study [49].

3.3. GrIS IST Anomalies

Figure 6 shows the temporal and spatial annual IST anomaly from 2001 to 2019 respectively. The IST anomalies between −1 °C and 1 °C are dominated in most of the study period except for some abnormally warm and cold years. The largest positive IST anomalies (exceeds 3 °C) occurred in the southwest in 2010 when almost the entire GrIS experienced positive IST anomalies consistent with [8]. In that year, Greenland experienced melting up to another 60 days compared with the 1960–2010 average (1980–2010 were simulated data), with the largest differences occurring at the southwestern and western margins of the GrIS [6]. The previous study shows that the unusual melt event in 2010 was trigged by large positive air temperature anomalies during May, accelerating snowpack metamorphism, and premature bare ice exposure [50]. The GrIS IST anomaly in May 2010 may be relevant to the eruption of Eyjafjallajőkull volcano on March 2010 which once had a profound influence on the environment and climatic conditions of neighboring regions [51]. With the positive albedo feedback and the fact that the wet snow could absorb up to three times more incident solar energy than dry snow, the further melt was fostered [52].

In addition to 2010, relatively large positive IST anomalies (1–3 °C) were mainly found in 2003, 2012, 2016, and 2019. The high positive IST anomalies were mainly found in central and southern GrIS in 2003, southeastern GrIS in 2012, and almost the entire GrIS in 2016 while northern and western GrIS in 2019. By comparison, a relatively large region of negative IST anomalies that between –1 °C and –3 °C were found in 2001, 2011, 2013 and 2015. Especially, the largest negative IST anomalies (below −3 °C) was found in 2015 in southwestern GrIS.

Figure 7 shows the temporal and spatial characteristics of the mean summer IST anomaly from 2000 to 2019, respectively. As reported by Hall et al. [8], the summer of 2012 is the warmest summer, with the southeast experiencing the largest positive IST anomalies (exceeding 3 °C), related to the strongest North American heat wave since 1895 [53]. In 2012, almost the entire GrIS (98.6%) experienced an extreme melting event., even at Summit [7]. This extreme melt event was associated with an abnormal ridge of warm air that was stagnant over Greenland and the radiative effects of low-level liquid layer clouds [7,54]. Previous studies had shown that on 11–13 July, the GrIS experienced the most extensive melt extent based on satellite-derived data [7,9]. Figure 8 are the MODIS IST on 11–13 July of the GrIS showing that the IST is concentrated between 0 °C and −3 °C, especially the western GrIS experienced the largest area of highest IST.

In addition, a relatively large area of positive IST anomalies (0–2 °C) in summer was mainly found in 2007, 2010, 2016 and 2019. The recent warm summer found in our study period is the summer of 2019, agreeing with the previous study [55,56]. The increasing IST is the basic factor causing the increasing GrIS melting [57]. The warm, moist air intrusions could potentially drive melt events [54,58]. At the end of July 2019, the GrIS melting extent achieved more than 62%, even at Summit [55] and the ablation zones continued through August and September, entering accumulation conditions in early October [59]. Contrary to the above-average warmth conditions during the summers mentioned above, relatively large areas of negative IST anomalies were found in 2000, 2001, 2002, 2006, 2009, 2013, 2017 and 2018.

3.4. IST Trend Analysis of GrIS

Figure 9 shows the positive IST trends mainly in spring and summer and negative trends in fall and winter. The annual IST trends are much more moderate than the seasonal IST trends and the most IST trends are concentrated ~+−0.5 °C/decade, showing a lower spatial variability. On an annual scale, the northern GrIS mainly experienced positive IST trends (0~0.5 °C/decade) while the southern GrIS mainly experienced negative IST trends (−1~0 °C/decade). Autumn cooling was common and Hall et al. [8] also found a decrease in IST (~−1.49 ± 1.20 °C/decade) in autumn during the period 2000–2012. Similarly, this work also showed the northwestern GrIS experienced positive IST trends except autumn during the study period, especially in spring with the positive IST trends reaching up to 1.5 °C/decade. Furthermore, the minimum negative IST trends (below −1.5 °C/decade) were found in the small area of southern and eastern GrIS in winter and a small area of southern GrIS in autumn.

4. Discussion

GrIS melt and runoff have increased rapidly since the early 1990s [26]. The low density and discontinuity of AWS measurement are the inherent limitation in evaluating the melt on the entire ice sheet. Surface melting conditions that cannot be measured due to the AWSs constraints can only be obtained by remote sensing [60]. Given the significance of remotely sensed melt products for monitoring GrIS melt, an assessment of the MODIS IST-derived melt data is essential.

The MODIS IST-derived melt data is determined by the melt threshold (−1 °C) and a non-cloud-obscured IST grid that is ≥−1 °C is defined as “melt”. Therefore, the melt threshold is the key factor to be validated. The IST products have an accuracy of ±1 °C and melt may occur when temperatures are slightly below freezing with a strong solar radiation [28]. Given the completeness and amount of the AWSs daily ablation data, the daily ablation data from 5 AWSs at low elevations and 3 AWSs at high elevations were selected to validate the melt threshold. The daily ablation data was set to 25 mm based on the accuracy of the pressure transducer [37]. Only the data that were both available for MODIS and AWS were used in this study. The days that ware detected melt or no melt both from MODIS and AWS were recorded as “BB” and these days can be considered well-matched days. If the days were detected melt only from MODIS or AWS, then they were recorded as “MM” or “SM” and these days can be considered failed-matched days.

Figure 10 shows the day-to-day melt agreement conditions between MODIS and AWSs. Obviously, AWSs at low elevations reproduced more melt occasions than at high elevations. In the strong ablation year of 2012, except for the KPC_L, MODIS IST reproduced fewer melt conditions than AWSs. In general, MODIS IST performed an average agreement of 55% with the highest agreement in KPC_L (61%), located at the northeastern GrIS. NUK_L at southwestern GrIS had the lowest consistency of 41% with much fewer melt days than that detected by MODIS IST. Other AWSs show an agreement between 50% and 60%. Overall, the −1 °C melt threshold suggested a 41–61% consistency of melt conditions (BB) between MODIS and AWSs, and generally, fewer MODIS melt days occurred than AWSs. Given the fewer MODIS melt days based on −1 °C and the cold bias found in this work, lower melt thresholds were used to possibly achieve higher consistencies (Figure 11). The results showed that consistencies were 46–66% and 51–70% based on melt thresholds respectively −1.5 °C and −2 °C, respectively. With the melt threshold decreasing, the consistency was generally becoming higher especially for the stations QAS_L, NUK_L, KAN_L, and TAS_U while KPC_L and QAS_U had lower consistencies.

5. Conclusions

In this study, based on PROMICE AWSs observations, three statistical indexes, bias, RMSE, and R are used to comprehensively evaluate the performance of the MODIS IST over the GrIS. Generally, MODIS IST has a widespread cold bias over the GrIS, ranging from −2 to −4 °C and this cold bias mainly occurred in winter. The consistency between MODIS IST and in-situ IST is very high on the GrIS, with an average of R = 0.92 and RMSE = 4.36 °C. Spatially, the R and RMSE agree on a better accuracy of MODIS IST on the northwestern, northeastern, and central GrIS.

The MODIS IST can well demonstrate the temporal and spatial variability of GrIS. The mean summer IST is mainly concentrated between −20 °C and −10 °C with the highest and lowest IST values of −1.33 °C and −17.98 °C, respectively. Besides, the largest positive IST anomalies (exceeding 3 °C) occurred in southwestern GrIS in 2010. Summer 2012 was the warmest summer with the largest positive IST anomalies (exceeding 3 °C) in the southeastern GrIS. The positive IST trends are mainly in spring and summer and the negative IST trends are in autumn and winter. Importantly, most of the annual IST trends are concentrated between ~±0.5 °C/decade.

This work also verifies the rationality of the MODIS IST melt threshold of −1 °C using the daily ablation data from AWSs. The results show that the −1 °C melt threshold suggested a 41–61% consistency of melt conditions between MODIS and AWSs. Furthermore, with the melt threshold decreasing (−1.5 °C, −2 °C), the consistency is generally becoming higher.