The Effect of Arctic Dust on the Retrieval of Satellite Derived Sea and Ice Surface Temperatures

Large quantities of dust are transported annually to the Arctic, primarily from Asian deserts. The influx of dust into the polar environment changes the radiative properties of clouds while the deposition of dust onto ice and snow decreases the surface albedo. Atmospheric and surface dust may be identified with space borne radiometers by comparing infrared energy in the 11 μm and 12 μm regime. Between 2007 and 2017 satellite infrared data revealed persistent low-level dust clouds in the vicinity of Amundsen Gulf in the Western Canadian Arctic during the melting season. Evidence suggests that the subsequent deposition of atmospheric dust in the region affected the surface emissivity in the thermal infrared regime. As a result, satellite derived sea and ice surface temperature algorithms were rendered inaccurate in these areas. Moreover, the ubiquitous nature of dust in the region may play a role in the rapidly vanishing cryosphere.

in the satellite record, leading to the conclusion that the observation was a physical manifestation of atmospheric and surface properties.

Methods
Satellite Detection of Dust. Naturally occurring materials such as ice, snow, water and sand are selective emitters in which the emissivity is a function of wavelength 29,30 . The difference in emissivity between wavelengths can help to identify a type of cloud or surface viewed by a satellite radiometer. Space based sensors that operate in the thermal IR regime utilize bands centered on 11 μm and 12 μm, both of which correspond to atmospheric windows and the region of peak radiation emission of the Earth. Figure 2 illustrates the change in emissivity between 11 μm and 12 μm for vegetation, snow/ice, desert/sand and ocean/water surfaces. For snow/ice there is a sharp drop in emissivity while the opposite is true for desert surfaces. There is a slight decrease for water surfaces and slight increase for vegetation. Generally, a satellite radiometer will record different brightness temperatures for the two channels as a result of the different emissivity values for these wavelengths. The Brightness Temperature Difference between the 11 μm and 12 μm channels (BTD [11][12] ) is highly positive for ice, somewhat positive for water, somewhat negative for vegetation and highly negative for sand 31 . The BTD [11][12] for ice, water and desert surfaces can be extended to an ice cloud, water cloud and dust cloud respectively. Hence, the differentiation between surface and cloud types is possible through the analysis of satellite BTD [11][12] data.
Dust clouds may be identified from space using thermal IR channels on the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA's Terra satellite. The sensor has 36 channels, including Channel 31 centered at 11 μm (10.8 μm to 11.3 μm) and Channel 32 centered at 12 μm (11.8 μm to 12.3 μm). Spatial resolution for the thermal IR channels is 1 km. Using MODIS BTD [11][12] values, the extent and intensity of a dust event can be ascertained. Dust events are identified when BTD 11-12 < −0.5 K in cloudy regions 31 . In the development of the algorithm, it was determined that the surface contributed −0.5 K to the BTD 11-12 , necessitating a threshold at that value. The intensity of the dust event may be determined by taking the difference of MODIS Channel 29 centered at 8.5 μm (8.4 μm to 8.7 μm) and Channel 31, or BTD [8][9][10][11] . In a region that meets the parameters of a dust event, BTD 8-11 > 0 represents a relatively strong dust region while BTD 8-11 < 0 signifies a relatively weak dust region 31 . A comparison of concurrent (±1 hour) AVHRR and MODIS imagery in this study demonstrated that AVHRR Channel 4 (10.3 μm to11.3 μm) minus Channel 5 (11.5 μm to 12.5 μm), with a spatial resolution of 1.1 km, is comparable to MODIS in identifying BTD [11][12] dust signatures.
In the arctic environment over areas of ice, snow, water and tundra where BTD 11-12 > 0, there is no requirement for a threshold of −0.5 K. In this study, a cloud with a negative BTD 11-12 is classified as a dust cloud, while any ice, snow or water surface exhibiting a negative BTD [11][12] under clear skies implies that there is layer of deposited material on top. Space based radiometers measure the skin temperature of the sea surface, which is less than 0.1 mm, so a very thin layer can return a signature that is indicative of a desert instead of a body of water or ice sheet. While it is possible that a substance other than mineral dust is causing the negative BTD 11-12 signature, such as anthropogenic aerosols, there is nothing in the literature to support this theory. In situ sampling will solve the issue of chemical composition and provenance of the material. Regardless, it is a known certainty that a negative BTD [11][12] cannot be produced by a pure water, ice or snow surface. Sea and Ice Surface Temperature Retrieval. Clear skies are required for satellite-derived sea and ice surface temperatures. Satellite sea surface temperature (SST) algorithms for temperate oceans utilize BTD [11][12] to determine clear sky atmospheric absorption between the sensor and sea surface. The presence of water vapor causes a positive BTD 11-12 so a more positive value leads to a greater correction to the main estimator at 11 μm (BT 11 ) to account for this absorption. AVHRR SST algorithms generally take the form, 11 11 12 11 12 where a, b, c and d are coefficients normally based on a regression analysis of concurrent satellite and in situ data, and θ is the sensor zenith angle. The final term in the algorithm accounts for amplified atmospheric absorption resulting from increased path length to the satellite sensor. For the computation of SSTs a moist atmosphere has a BTD 11-12 ≥ 0.7 K while a dry atmosphere corresponds to BTD 11-12 < 0.7 K 32 . Since a positive BTD 11-12 is expected, negative values will return an inaccurately cool SST. An ice surface temperature (IST) algorithm takes the same form as equation (1), but in this case the BTD 11-12 term accounts for a number of factors including modeled directional snow emissivities 33 . An IST BTD 11-12 term should be positive, with a negative value returning incorrect cooler temperatures. In the case of dust, which has a lower emissivity than water and ice for BT 11 (Fig. 2), the error of the retrieved temperature using equation (1) is further amplified.
Surface Temperature Retrieval using CASSTA. The Arctic is a dry climate, with an annual mean distribution of specific humidity near the surface of 1 g kg −1 , compared to 18 g kg −1 in equatorial regions 34 . In accordance with the paradigm for temperate water SST algorithms, atmospheric BTD 11-12 over Arctic waters should be low; however, high values are observed, particularly during the colder months 35 . Inflated atmospheric BTD [11][12] in the Arctic is a function of clear sky ice crystals that absorb more 12 μm (BT 12 ) energy than water vapor 35 . Arctic BTD [11][12] values are highest in the winter, with values exceeding 2 K in regions of ice fog in the vicinity of leads, and reducing to approximately 0.5 K during the summer 35 . As a result, temperate water algorithms overestimate Arctic SSTs by 2 to 3 K 28 . CASSTA, which was developed with AVHRR and concurrent in situ data, combines three different algorithms to determine the temperature of seawater, marginal ice zones and ice, each of which is designated by the BT 11 value. A single channel (BT 11 ) is used to determine Arctic SST to avoid high BTD [11][12] brought about by enhanced BT 12 absorption by atmospheric ice crystals, which is not statistically related to absorption of the main estimator. Marginal ice zones use a weighted average between a standard IST and the Arctic SST, while ice regions use the IST algorithm 28 . In the case of dust deposition on the surface, the Arctic SST will return an inaccurately cool temperature as a result of the lower emissivity of dust compared to water for BT 11 (Fig. 2).
Satellite Data for the Study. The Arctic is a challenging environment to retrieve satellite derived SSTs.
With the exception of polynyas, open water is limited to the melting season, which is characterized by the persistence of arctic stratus 36 with estimates of July cloud cover exceeding 90% 37 . In this study METOP-A and National Oceanic and Atmosphere Administration (NOAA) AVHRR satellite imagery was initially used to determine BTD [11][12] in the Amundsen Gulf region. AVHRR was chosen for the initial search since CASSTA could be applied to these images to evaluate the magnitude of surface temperature discrepancies. Anomalies were identified as a Figure 2. Angularly averaged emissivity (ε) of snow/ice, ocean/water, desert/sand and vegetation for thermal IR wavelengths 29 and BTD 11-12 for surface and cloud types 31 . The resulting difference in brightness temperature between 11 μm and 12 μm channels, or BTD [11][12] , allows the differentiation between surface and cloud types. The emissivity of bare land is a function of rock or soil type, though BTD [11][12] is commonly zero to highly negative 29,31 . In this study, land areas consisting of tundra, generally showed highly positive BTD [11][12]  negative BTD [11][12] in areas of water and ice that appeared cloud free. The initial search for anomalies focussed on June to August for the years 2007 to 2017. One to three images were available per day between 17Z and 21Z of which about 20% were sufficiently cloud-free for further examination. Negative BTD [11][12] anomalies were detected at least once for June, July and August for every year from 2007 to 2017. Most imagery showed negative BTD [11][12] in optically thin, low lying clouds in the vicinity of Amundsen Gulf.

AVHRR Imagery. CASSTA was developed in the North Water Polynya (NOW) situated between Elsmere
Island and Greenland (78 N 76 W). Initially, the purpose of this study was to compare CASSTA to temperate ocean SST algorithms in more southern regions of the Arctic. The application of these algorithms to the Amundsen Gulf region (70 N 120 W) focussed on a relatively cloud free period 1 to 12 July 2016. During this timeframe, there were instances when temperate ocean SST algorithms returned cooler surface temperatures than CASSTA, which is not possible under normal atmospheric conditions. Further investigation showed that these anomalies were the result of negative BTD [11][12] . Examination of 210 AVHRR satellite images for June to August 2016 revealed the following observations.
• A persistence of low-level, optically thin clouds with BTD 11-12 ranging from 0 to −1 K.
• A regular occurrence of optically thick clouds with BTD 11-12 ranging from 0 to +2 K.
• Under apparent clear skies, ice and seawater in the Amundsen Gulf region and Great Bear Lake periodically showed negative BTD 11-12 that is not characteristic of these surface types.
The initial CASSTA study of the area was expanded to include other years. Over 2,300 AVHRR for the Amundsen Gulf region were examined for June to August 2007 to 2017. Similar anomalies were detected every year to varying degrees, the detection of which was often hindered by clouds. Figure 3 shows the extent of negative BTD [11][12] for selected AVHRR scenes in 2010, 2013 and 2016. A closer inspection of Amundsen Sound for the 09 July 2016 image shows a mixture of positive and negative BTD [11][12] . Figure 4 compares the BTD 11-12 mapping and CASSTA retrieval for this day. Areas of open water with a positive BTD 11-12 are 6 K to 10 K warmer than water with negative BTD 11-12. Since CASSTA uses BT 11 solely to calculate SST, this result implies that negative BTD [11][12] regions have reduced emissivity in the 11 μm regime. The negative BTD [11][12] in concert with lower temperatures, is consistent with the properties of dust. MODIS Imagery. The dust detection algorithm was applied to MODIS imagery to corroborate the AVHRR analysis. In all cases of concurrent AVHRR/MODIS imagery there was agreement between the two sensors. The thermal channels for MODIS are 0.5 μm wide compared to 1 μm for AVHRR, which generally resulted in more negative BTD 11-12 dust signatures for that sensor. Dust clouds were generally observed as optically thin, low lying features as illustrated in Fig. 5. The addition of the 8 μm channel on MODIS allowed the intensity of the dust event to be evaluated using BTD [8][9][10][11] . The majority of dust clouds tested were classified as relatively weak, although there were examples of strong dust events that persisted for extended periods.
Similar to AVHRR, the MODIS imagery showed ice and water surfaces with negative BTD 11-12 under apparent clear skies. It is possible that optically thin dust clouds close to the surface are creating a signature that mimics one of dust deposition. For these clear sky cases, the negative BTD 11-12 follows the coastline, while the adjacent land surfaces generally show highly positive BTD [11][12] . On many occasions during the study period Great Bear Lake, approximately 200 km south of Amundsen Gulf, also displayed negative BTD 11-12 under apparent clear skies. In these instances, the land between the two bodies of water showed positive BTD [11][12] values with no visual cues of a dust cloud. Figure 6 compares a MODIS image of Great Bear Lake and Amundsen Gulf to the BTD 11-12 mapping of the same scene. The evidence suggests that dust deposition has occurred on Great Bear Lake and Amundsen Gulf. Some of the land close to Amundsen Gulf is less positive, which may be the result of dust deposition or the natural state of the surface for August.
The formation, evolution and dissolution of anomalous sea and ice surface BTD 11-12 is difficult to observe with satellite data as a result of persistent cloud cover. While there are cases of the phenomenon dispersing within 24 hours, there are also examples in the satellite record that demonstrate the potential for anomalies to last for extended periods. Figure 7 shows four consecutive days in July 2016 of negative BTD 11-12 on the surface of Amundsen Gulf. Clouds obstructed clear observation of the onset and dissolution of the phenomenon, which could be seen intermittently through breaks in the cloud for twelve days. In this series of images, the BTD 11-12 of ice is more negative than that of water. If the area was covered by an optically invisible dust cloud, the opposite would be true since the positive contribution of BTD [11][12] from an ice surface is greater than that of water. An explanation for this observation is that dust can collect on ice in greater density than a water surface that moves under the influence of wind and current.
Winter Imagery. Dust clouds identified by negative BTD 11-12 on MODIS images are readily apparent during the winter months in the Arctic since the underlying icescape is highly positive. Figure 8a [11][12] over ice and water are approximately 400,000 km 2 . The BTD [11][12] signature in these areas is not representative of water or ice surfaces and is in stark contrast to the expected positive BTD [11][12] . (Images created with Harris Geospatial ENVI 5.3 software, http://www.harrisgeospatial.com).  [11][12] and positive BTD [11][12] . SST values for negative BTD [11][12] are significantly lower, indicating that there is a reduction in emissivity in these regions. (Images created with Harris Geospatial ENVI 5.3 software, http://www.harrisgeospatial.com). Figure 5. MODIS true color images on the left are shown for 08 and 09 August 2013 with corresponding BTD 11-12 mappings on the right. A dust cloud is identifiable in dark blue for both days. Portions of these clouds meet the criteria for a strong dust event (BTD 11-12 < 0, BTD 8-11 > 0). Higher clouds, identified by positive BTD [11][12] , block out portions of the lower dust cloud for both images. The surface is visible through much of the dust cloud. (Images created with Harris Geospatial ENVI 5.3 software, http://www.harrisgeospatial.com).
that gives the warmest temperature, while leads in the ice pack also show elevated thermal signatures. Figure 8b shows the corresponding BTD 11-12 mapping with negative values superimposed on the thermal image. Similar to the melting season, the dust clouds identified by negative BTD [11][12] are low-level features that are blocked by higher ice clouds that exhibit highly positive BTD [11][12] . Clouds at any level may potentially contain dust aerosols that have been coated with ice and subsequently give the highly positive BTD 11-12 signature of an ice cloud. In many cases the regions identified as dust clouds are optically thin, appearing as a haze that allows the ice features to be viewed below (Fig. 3a). The signature of dust clouds in the winter are generally less negative than those observed in the warmer months as a result of the contribution of the highly positive surface BTD 11-12 . In the Figure 6. A MODIS true color image and the corresponding BTD 11-12 mapping is shown for 04 Aug 2013. The negative BTD 11-12 of the water surfaces indicate that the signature cannot be the result of water, suggesting that a material such as dust is coating the surface. Textural difference in the land between Great Bear Lake and Amundsen Gulf in the true color image correspond to relatively low BTD [11][12] . This may be the result of dust deposition or the natural state of the surface for August. Highly positive BTD 11-12 indicate high ice clouds. (Images created with Harris Geospatial ENVI 5.3 software, http://www.harrisgeospatial.com). with a corresponding negative BTD [11][12] overlay to the right is shown for each day. A dust signature persists on the water (to the south) and ice (to the north) for at least four days. Clouds prevented observation of the onset and dissolution of the phenomenon. The more negative BTD [11][12] signatures appear on the ice, which supports the hypothesis that it is a surface phenomenon. (Images created with Harris Geospatial ENVI 5.3 software, http://www.harrisgeospatial.com). case of open water, a highly positive ice fog signature 35 may overwhelm the negative BTD 11-12 dust signature (Fig. 3b). In apparent cloud free areas, some ice exhibits negative BTD [11][12] , suggesting dust deposition on the surface (Fig. 3b). Although this study did not focus on the colder months, dust clouds were commonly observed in winter satellite imagery of the region.

Discussion
In this study significant dust clouds, identified by negative BTD 11-12 using satellite radiometers, were detected in the in the Amundsen Gulf region during the melting season. Dust signatures were also observed on ice and water surfaces under apparent clear skies, suggesting dust deposition. Surrounding land surfaces in the region generally did not show negative BTD [11][12] , although values may have been less positive as a result of dust deposition. The presence of dust prevents accurate SST and IST retrieval. Regardless of whether the phenomenon is a surface or a near-surface feature, the problem for satellite surface temperature retrieval using AVHRR remains the same.
Previous studies of Arctic SSTs in the NOW 28,38 , were not hampered by dust clouds or dust deposition on the surface. The observed persistence of dust in the vicinity of Amundsen Gulf is notable since the apparent deposition of particulates changes the emissive properties of the surface. In this study, the application of CASSTA to negative BTD 11-12 regions led to an underestimation of 5 K to 10 K as a result of the lower emissivity of dust compared to water. The application of temperate water SSTs applied to the same area results in an even greater error as a result of the BTD 11-12 term in the algorithm that is designed to account for absorption by atmospheric water vapor. In the case of a negative BTD 11-12 this corrective term, which is supposed to add to the main estimator (BT 11 ), will erroneously reduce the value by an additional 1 K to 5 K depending on the BTD 11-12 value and sensor zenith angle. These are significant errors for SST algorithms, which are projected to determine sea surface skin temperatures with an accuracy of better than 0.5 °C 39 . A mask could be used to identify and avoid areas of negative BTD [11][12] ; however, in some cases dust may simply reduce highly positive BTD [11][12] generally observed in the Arctic to values that are still above zero but erroneous. As such, significant errors could still be propagated in the data.
An IST algorithm was not utilized in this study, since summer ice in Arctic waters typically exceed the BT 11 threshold (271 K) in which the SST portion of the CASSTA is employed. This aspect of the algorithm leads to better ice differentiation during the summer months 28 . During colder months an IST algorithm will potentially return even cooler anomalies than water as a result of having a higher emissivity differential compared to dust (Fig. 2). Additionally, the BTD 11-12 component of the algorithm will further increase the error. The observed dust in winter scenes may also promote the formation of ice fog in the vicinity of leads and polynyas as a result of ice nucleation. Ice fog is difficult to detect with satellite imagery and prevents the determination of surface temperature 28 . Application of an IST algorithm to these regions will overestimate the surface temperature as a result of inflated BTD 11-12 (>2 K) that is related to atmospheric ice crystals 35 . It should be noted that the ice fog BTD [11][12] signature is a function of enhanced IR absorption in the 12 μm regime.
While atmospheric and surface dust is problematic for SST and IST satellite retrieval, the ubiquitous nature of the phenomenon highlights issues that extend beyond the discipline. It is estimated that the introduction of mineral dust into the Arctic climate plays a role in Arctic Amplification 9 and the subsequent loss of ice. On the other side of the globe, surface air temperatures in Antarctica have not risen to the same extent as the Arctic and ice extent has trended upward since 1979 40 . Approximately 16 times more dust is transported to Arctic latitudes compared to equivalent Antarctic latitudes 14 as a result of fewer dust sources in the southern hemisphere coupled with unfavourable wind trajectories 41 . This suggests that the reduction in dust aerosols transported to Antarctica may be a factor in the disparity between climatic trends of the two Polar Regions. In the Arctic, the greatest decline in sea ice over the past two decades is the Western Arctic, which is losing ice at a faster rate than predicted by models 42 . The high frequency of dust events observed in the vicinity of Amundsen Gulf may be a contributing factor to the rapid disappearance of the cryosphere in the region.