ETM+ data and SGS image of MOA dataset description

This study uses 1073 scenes of Landsat-7 ETM+ images acquired during austral summers over 1999–2003 that cover the portion of the Antarctic continent north of 82.5° S, and an SGS image of the MOA dataset acquired during the 2003/ 04 austral summer covering the area south of 82.5° S, where Landsat has no coverage because of its near-polar orbit. The ETM+ raw data were taken for the LIMA project (Reference BindschadlerBindschadler and others, 2008) and were downloaded from the US Geological Survey website (http://lima.usgs.gov/access.php) and the SGS image of the MOA dataset (Reference Haran, Bohlander, Scambos and FahnestockHaran and others, 2005; Reference Scambos, Haran, Fahnestock, Painter and BohlanderScambos and others, 2007) was provided by the US National Snow and Ice Data Center (NSIDC; http://nsidc.org)

To correct for snow and ice saturation in the visible spectrum of ETM+ data due to the high reflectance of these surfaces, a linear regression method obtained from the relationship of different bands was adopted for DN saturation adjustment of ETM+ data, and clear textures were observed after the adjustment. The Landsat ETM+ data were then converted into planetary reflectance. The multispectral bands (bands 1–5 and 7) with 30m resolution were then sharpened into 15 m spatial resolution using band 8. Finally, the Landsat-7 ETM+ mosaic image of Antarctica north of 82.5°S was assembled by combining all multispectral ETM+ scenes at 15 m spatial resolution (Reference HuiHui and others, 2013).

The SGS image of the MOA dataset was achieved from the MOA acquired during the 2003/04 austral spring and summer, and re-gridded using image super-resolution methods to 125 m spacing (150–200 m true resolution) (Reference Scambos, Haran, Fahnestock, Painter and BohlanderScambos and others, 2007).

The ETM+ data with 1 5m spatial resolution and the SGS image of the MOA dataset with 125 m gridding space were used to map the BIAs. During BIA mapping, the method of thresholds based on band ratio and reflectance was employed in ETM+ data and the method of threshold values was applied in the SGS image of the MOA dataset. Details are described in the next subsection.

Thresholds based on band ratio and reflectance for ETM+ data

Various methods have been used to distinguish BIAs from snow, firn or exposed rock, including supervised and unsupervised classification (Reference Winther, Jespersen and ListonWinther and others, 2001; E and others, 2011; Reference Yu, Liu, Wang, Jezek and HeoYu and others, 2012), image segmentation based on image texture (Reference Liu, Wang and JezekLiu and others, 2006; Reference Yu, Liu, Wang, Jezek and HeoYu and others, 2012), and threshold discrimination based on a band ratio or grain size obtained from a band ratio (Reference Orheim and LucchittaOrheim and Lucchitta, 1990; Reference Bronge and BrongeBronge and Bronge, 1999; Reference Brown and ScambosBrown and Scambos, 2004; Reference Scambos, Haran, Fahnestock, Painter and BohlanderScambos and others, 2007; E and others, 2011). Good results have been achieved using these methods. The band ratio method is based on the differences in spectral reflectance of blue ice and snow, which is a clear physical basis for discriminating between blue ice and snow, moraines or exposed rock. This method has been widely used in studies of snow grain size (Reference DozierDozier, 1989; Reference Fily, Bourdelles, Dedieu and SergentFily and others, 1997; Reference Nolin and DozierNolin and Dozier, 2000; Reference Li, Stamnes, Chen and XiongLi and others, 2001; Reference ScambosScambos and others, 2012). The ratio can be defined as

The ratio between two bands can remove much of the effect of illumination and minimize the influence of surface slopes. Ratios can also be used as an index for grain size, though one specific ratio cannot give an exact grain-size value. Studies show that certain ratios, i.e. those that use Landsat band 5 (1.55-1.75 m m), are more sensitive to terrain conditions than other ratios, i.e. those that use band 7 (2.08-2.35 m m). Band 4 (0.76-0.90 m m) is less sensitive to atmospheric effects and pollution than bands 2 (0.52-0.60 m m) and 3 (2.08-2.35 m m) (Reference DozierDozier, 1989; Reference Fily, Bourdelles, Dedieu and SergentFily and others, 1997; Reference KorenKoren, 2009). It was found that values of R₄₇ match the theoretical curves of snow grain size better than those of R₄₅, though both may have had saturation problems in the ratio for large grain sizes in that study (Reference Fily, Bourdelles, Dedieu and SergentFily and others, 1997). This study selects R ₄₇ as the best ratio to obtain the thresholds for mapping BIAs using Landsat-7 ETM+ data. Figure 1 shows the clear differences in spectral reflectance among blue ice, snow and rocks. These differences are caused by the scattering properties of materials with different grain sizes (Reference DozierDozier, 1989; Reference Nolin and DozierNolin and Dozier, 1993). In Figure 1 the spectral reflectance was measured in December 2009 near Zhongshan station (69° 22ʹ S, 76°22ʹE), Antarctica, using an Analytical Spectral Devices (ASD) FieldSpec Pro FR spectroradiometer over the 350–2500nm spectral range. The weather was sunny and breezy when spectra measurements were taken. The six bars in Figure 1 represent six bands (bands 1–5 and 7) of the ETM+ sensor. Figure 1 shows that snow had the highest reflectance in the visible and near-infrared (VNIR) wavelengths of the spectrum, with two peaks in reflectance in the shortwave infrared (SWIR) wavelength. Blue ice had higher reflectance than snow in the blue spectrum, hence its bluish appearance, but strongly absorbed solar radiation in the red and near-infrared portion of the spectrum and had very low reflectance in SWIR. The spectral signatures of exposed rock were very different from those of snow and blue ice. Exposed rock had strong absorption in the visible spectrum, but the reflectance changed little from the near-infrared wavelength to SWIR, and the reflectance values in SWIR were higher than those of both blue ice and snow. Comparison of the spectral curves for blue ice, snow and exposed rocks indicates that the red, infrared and SWIR bands have the most distinct spectra for discriminating blue ice from snow and exposed rocks. The slope of the reflectance curve of blue ice in band 4 was larger than that of snow in band 4, and the values of blue ice and snow in band 7 were small and relatively constant. Consequently, it was inferred that values of R ₄₇ for blue ice and snow should be large and that those of snow should be less than those of blue ice because of the different slopes of the reflectance curves. Values of R ₄₇ for exposed rocks should be negative or very small. Therefore, R ₄₇ can clearly distinguish BIAs from snow and exposed rocks.

Fig. 1.

Spectral reflectance of blue ice, clean snow and exposed rocks near Zhongshan station, Antarctica, in December 2009. The spectra of blue ice, snow and rocks were measured at 69 ° 27.242• S, 76 ° 21.134• E on 22 December 2009, at 69 ° 19.314• S, 76 ° 27.517• E on 12 December 2009 and at 69 ° 22.286 •S, 76 ° 22.061• E on 24 December 2009. It was sunny and breezy when the spectra measurements were taken. The six bars from left to right are the blue, green, red and near-infrared bands respectively (bands 1–4) and two SWIR bands (bands 5 and 7).

To cover the whole Antarctic continent requires a great amount of Landsat-TM/ETM+ data. In this study, 14 sample ETM+ scenes in different places were randomly chosen to define the threshold values for blue-ice identification. Previous studies indicated that BIAs, snow, rock, shadowed snow or BIAs, and shadowed exposed rock could be identified and differentiated in the Landsat TM/ETM+ images. BIAs were classified as wind-induced or melt-induced, depending on the dominant climate process, which is closely related to the mechanism of formation of the BIAs (Reference WintherWinther, 1994; Reference Liston, Bruland, Winther, Elvehøy and SandListon and others, 1999; Reference Winther, Jespersen and ListonWinther and others, 2001). They were also classified as smooth, level or rough, according to image texture information from ETM+ data. Smooth BIAs were related to wind-induced BIAs with smooth image texture, level BIAs were associated with pond and lake ice (seasonal ice) with homogeneous texture (a level, specular surface) and rough BIAs were connected with troughs, ridges and surface crevasses with a relatively rough surface (Reference Yu, Liu, Wang, Jezek and HeoYu and others, 2012). Pixels of smooth BIAs and level BIAs were almost pure, but those of rough BIAs were likely to be mixed pixels containing both blue ice and snow. In this study, training areas for different surface features were identified by visual interpretation. BIAs were only classified as smooth or rough according to pixel type (pure or mixed pixels). The former category was closely related to small grain size, while the latter was better related to larger grain size.

Six different types of surface features were used as training samples: smooth BIAs; rough BIAs; snow; exposed rock; shadowed snow or BIAs; and shadowed exposed rock. First, we selected training samples for these six types through visual interpretation of 14 randomly selected false-color composite images, which were combined by bands 7, 4 and 1. Because the 14 sample images were randomly selected, not all the six different types of surface features were included in each sample image. Next, some statistical variables – maximum, minimum, mean, standard deviation of the apparent reflectance and R ₄₇, and the number of pixels – were calculated for all the training samples in each image. Finally, comparative analyses were carried out to determine the threshold values for BIA identification.

The mean and standard deviation of apparent reflectance of six bands for each surface feature type as derived from our training samples in 14 sample images are listed in Table 1, and the mean R ₄₇ value of each surface feature type in each sample image as derived from our training samples is shown in Table 2.

Table 1.

The mean (p) and standard deviation (SD) of apparent reflectance of six bands for each surface feature type as derived from our training samples

Table 2.

The mean R ₄₇ value of each surface feature type in each sample image as derived from our training samples

Table 1 shows that of the six types of surface, snow had the highest apparent reflectance in the VNIR wavelengths (bands 1–4) while shadowed exposed rock had the lowest; both had only slight fluctuations and small standard deviations. The apparent reflectance of smooth and rough BIAs was the next highest after snow, but the roughest BIAs had greater ranges and high standard deviations, likely caused by the mix of snow and BIAs in the pixels. The apparent reflectance of shadowed snow and BIAs was lower, but changes in the apparent reflectance from band 1 to band 4 were successively reduced, as were those of smooth BIAs, rough BIAs, snow and shadowed exposed rock. Rock showed the highest reflectance of all six types for the SWIR wavelength (bands 5 and 7). The results are consistent with those obtained from the LIMA dataset for Lambert Glacier basin (Reference Yu, Liu, Wang, Jezek and HeoYu and others, 2012).

Table 2 shows that R ₄₇ values of smooth and rough BIAs were >0.900 (minimum value 0.901), those of snow were <0.900 (maximum value 0.894), those of rock were <0.500 and those of shadowed exposed rock and shadowed snow or BIAs were unsteady with wide fluctuations and a maximum value of 0.914; the large range is likely associated with terrain, grain size and scattered light. In Table 1, the R₄₇ values of smooth and rough BIAs are >0.90, the highest for any area.

Tables 1 and 2 suggest that pixels with R₄₇ values greater than 0.90 cannot be identified as blue ice because pixels containing shadowed snow or BIA pixels could be misidentified as blue ice. Previous studies showed that BIAs have a lower albedo (0.50-0.70) than snow (0.80-0.90) (Reference Bintanja and BroekeBintanja and Van den Broeke, 1994, Reference Bintanja and Van den Broeke1995; Reference Bintanja, Jonsson and KnapBintanja and others, 1997; Reference BintanjaBintanja, 1999; Reference Reijmer, Bintanja and GreuellReijmer and others, 2001; Reference Warren, Brandt, Grenfell and McKayWarren and others, 2002), so the apparent reflectance could be used as another criterion to identify BIAs. In Table 1 the reflectance of bands 3 and 4 distinguishes among the six different surface features well. Band 4 exhibits the greatest difference between itself and the other bands.

The band 4 average reflectance of smooth and rough BIAs (0.482 ±0.138 and 0.635 ±0.255, respectively) was lower than that of snow (0.865 ± 0.027) and much higher than that of shadowed snow or BIAs (0.148 ±0.112). Maximum and minimum values of reflectance of the six surface features in the 14 sample images were obtained and plotted (Fig. 2). Figure 2 clearly shows that the maximum and minimum reflectance of both smooth and rough BIAs in band 4 were lower than those of snow and much greater than those of the other three surface features. The maximum reflectance of BIAs was lower than that of snow. Pixels with band 4 reflectance between 0.30 and 0.70 could be identified as BIAs. Therefore, the criteria to identify BIAs were defined as having both R ₄₇ > 0.900 and band 4 reflectance between 0.3 and 0.70.

Fig. 2.

Maximum and minimum reflectance values of the six surface features in the 14 sample images. In each bar the top is the maximum, and the bottom the minimum, value of apparent reflectance.

The difference in spectral characteristics between smooth and rough BIAs is clearly very small (Fig. 2). It was therefore very difficult to distinguish smooth from rough BIAs using just their spectral signatures. In this study, we focused on mapping BIAs and identifying their spatial correlation with ice velocity and elevation; we gave less attention to identifying the types of BIAs.

Threshold-based snow grain size for MOA dataset

BIAs covering the area south of 82.5° S were mapped using the SGS image of the MOA dataset. BIAs were considered when snow grain size was >400 (Reference Scambos, Haran, Fahnestock, Painter and BohlanderScambos and others, 2007).

All data were processed using a procedure developed using ENVI software (Exelis Inc., McLean, VA) and Interactive Data Language (IDL; Exelis Inc., McLean, VA, USA).

Noise reduction

BIAs were identified using R₄₇ values and band 4 reflectance, and the resulting images were converted into binary images. To reduce this ‘salt-and-pepper’ effect and produce more realistic maps, a median filter with 5 x 5 pixel windows was applied. This filter is less sensitive than linear techniques to extreme changes in pixel values; it can remove salt-and-pepper noise and preserve useful details in images without significantly reducing the sharpness of an image. This method proved effective at removing most of the noise, though some noise still remains in the final images.

Accuracy assessment

It is too difficult to carry out field surveys in Antarctica to assess the accuracy of our remotely sensed BIA mapping. Instead, we use results from recently published papers as validation data. Approximately 60 000 km² and possibly as much as 241 000 km² of BIAs were observed in NOAA AVHRR data acquired during 1980-94 (Reference Winther, Jespersen and ListonWinther and others, 2001). That study’s maximum area of 241000 km² is essentially consistent with our results of 234 549 km². The Grove Mountains contained 601.9 km² of BIAs according to ETM+ data acquired on 18 January 2001 (E and others, 2011), 610.1 km² according to MODIS data acquired on 17 January 2001 (E and others, 2011) and 745.3 km² according to MODIS data acquired during the 2003/04 austral summer (Reference Scambos, Haran, Fahnestock, Painter and BohlanderScambos and others, 2007). Our objective study found 624.2 m² of BIAs in the Grove Mountains, which agrees with the results of E and others (2011) but differs by >100km² from the MODIS summertime data of 2003/04. Reference Scambos, Haran, Fahnestock, Painter and BohlanderScambos and others (2007) reported five BIAs as blue-ice regions using the MOA dataset. Only one BIA was found in our result (69.5660 S, 43.566° E). The other four sites were not identified, for reasons discussed in the following subsection.

The differences between previous results and our results may be due to temporal changes in BIAs, the use of different identification algorithms or the different spatial and spectral resolutions of the sensors.

The size of BIAs may have changed during the past few decades and over the years between AVHRR, MODIS and ETM+ image acquisitions due to changes in wind speed, air temperature (solar radiation) and ice flow velocity, factors that could affect BIAs. The areas and spatial distributions of BIAs determined from satellite images at the same place but at different acquisition times could vary widely. Smaller areas of blue ice would become exposed when strong winds blow thin snow over BIAs, while the melting of that thin snow in periods of higher air temperatures would expose larger BIAs. The seasonal and annual changes of BIAs may be significant, so BIA changes need to be confirmed by more detailed studies.

The different spatial and spectral resolutions of the AVHRR, MODIS and ETM+ data likely also affected the differences in the various BIA calculations. Pixels containing both blue ice and snow in the 1100m resolution AVHRR data and in the 250 m resolution MODIS data (or in the 125 m gridding-space MOA data) can either be omitted or counted as BIAs, leading to either an underestimate (pixels being omitted) or an overestimate (pixels being counted) of BIAs. However, these pixels were clearly recognized in the 15 m resolution ETM+ data. Previous studies suggest that uncertainties such as mixed pixels, crevasses and shadows could account for the difference between identified BIAs (~60000km²) and the potential maximum area (241000 km²) obtained from the NOAA AVHRR data (Reference Winther, Jespersen and ListonWinther and others, 2001). Figure 1 indicates that the spectral signatures of the six-band ETM+ sensor discriminated surface features more sensitively than those of the two-band AVHRR and MODIS data, which discriminated BIAs from exposed rock and snow. The higher spatial resolution and better spectral signatures of the ETM+ data improved BIA recognition. In the future, 16-bit images with nine spectral bands from the Landsat-8 Operational Land Imager (OLI) will play an important role in mapping BIAs.

The areas and spatial distribution of BIAs obtained by different methods may have differed because of different underlying assumptions and theories. In summary, our estimates for the areas and spatial distribution of BIAs are likely to be an improvement over previous results from AVHRR and MODIS data.

Spatial distribution

BIAs identified from 1999-2003 ETM+ data covered ~1.67% of the area of Antarctica and were scattered over the continent but generally located in coastal or mountainous regions. Figure 3 shows that BIAs were concentrated in four regions: Victoria Land, the Transantarctic Mountains, Dronning Maud Land and the Lambert Glacier basin. They were sparsely distributed in other regions. These results are in agreement with previous studies (Reference Winther, Jespersen and ListonWinther and others, 2001; Reference Brown and ScambosBrown and Scambos, 2004; Reference Scambos, Haran, Fahnestock, Painter and BohlanderScambos and others, 2007; E and others, 2011; Reference Yu, Liu, Wang, Jezek and HeoYu and others, 2012). These four regions either contain nunataks or mountains with high elevations and steep slopes or are at relatively low elevation and have notable glacier dynamics. BIAs seldom occur in West Antarctica and the Antarctic Peninsula. This is generally due to higher snow accumulation and a relatively brief melt season, and the predominantly springtime acquisition time of the images used in LIMA and MOA. High snow accumulation regions require an intense melt season before the winter-spring snow accumulation is converted to surface ice. Moreover, winds are more variable, and more frequently tied to synoptic weather systems rather than highly sublimating katabatic processes. Although a late-summer acquisition and BIA analysis is possible, problems would arise in distinguishing wet snow from blue ice using the methods we present here, so in our results BIAs are less distributed in West Antarctica and the Antarctic Peninsula.

BIAs were associated with surface mass balance and ice velocity features (Reference Ligtenberg, Lenaerts, Van den Broeke and ScambosLigtenberg and others, 2013). Overlay analysis between BIAs and Antarctic ice velocity data with 450 m spatial resolution (Reference Rignot, Mouginot and ScheuchlRignot and others, 2011a, Reference Rignot, Mouginot and Scheuchlb) and digital elevation models (DEMs) with 1000 m spatial resolution (Reference Bamber, Gomez-Dans and GriggsBamber and others, 2009a, Reference Bamber, Gomez-Dans and Griggsb) was conducted (Fig. 4 and 5). Because there were great differences in map scale between the BIA data, ice velocity data and the DEM, approximately 9342 km² and 15 122 km² of BIAs in the overlay analysis of ice velocity data and the DEM were ignored. Statistical analysis was performed only on overlaid data between BIAs and ice velocity and DEM.

Fig. 4.

BIA distribution of different classes of ice velocity.

Fig. 5.

BIA distribution of different classes of surface slope.

The results in Table 3 demonstrate that the ice velocity in -74.26% of BIAs was <100ma^–1, and in about half these cases <1 ma–¹, while 9.70% of BIAs had an ice velocity of 100-200 m a–¹, 8.96% an ice velocity of 200-500 ma^–1, and 5.62% an ice velocity of 500-1000ma–¹. Ice velocity exceeded 1000 m a–¹ for only ~1.46% of BIAs. Regions with an ice velocity of 200-500ma^–1 were situated on the Ross, Shackleton, Amery, Riiser-Larsen and Ronne ice shelves, which are closely associated with glacier dynamics. The results in Table 3 show that 54.24% of BIAs had surface slopes less than 18, and 23.35% had surface slopes of 1-28. Surface slopes of only 5.51% of BIAs, at most, were >58, these areas being primarily located in the Transantarctic Mountains and Victoria Land.

Table 3.

Areas and percentage of BIAs for different classes of ice velocity and surface slope

Surface slopes of BIAs with ice velocities greater than 500 m a–¹ ranged from 0 to 39.058, but 96.65% of these BIAs had slopes smaller than 5° and were distributed along ice-shelf coastlines. This suggests that BIAs with higher ice velocities may occur on flatter glaciers with smaller surface slopes. BIAs with surface slopes greater than 5° had ice velocities between 0 and 2832 m a–¹, but the ice velocities of 95.27% of these BIAs were <500ma^–1. However, BIAs with slopes greater than 5° and ice velocities greater than 500 m were dispersed across Antarctica. BIAs with gentle slopes but greater ice velocity were mainly distributed on coastal ice shelves, and inland BIAs with steeper slopes but slower ice velocity were mainly distributed along or near mountains. This suggests that there was no significant relationship between ice velocity and surface slope; the relationship may be evident when rough and smooth BIAs are separated.

The spatial resolution of the Antarctic ice velocity map and DEM (900 m and 1000 m, respectively) was much coarser than that of the BIA map (15 m): at least 60 x 60 pixels in the ETM+ data constitute only one pixel in the Antarctic ice velocity map and DEM. As a result, the ice velocity and surface slope of the BIAs could not be described at high resolution. Errors may have resulted from this large difference in spatial resolution among the datasets. The dates of data acquisition also varied greatly: ETM+ data were acquired from 1999 to 2003, ice velocity data were created from SAR data acquired from 2007 to 2009, and the DEM was generated using Satellite Radar Altimeter (SRA) data from European Remote-sensing Satellite- ^₁ (ERS-1) acquired from 1994 to 1995, and Geosciences Laser Altimeter System (GLAS) data from the Ice, Cloud and land Elevation Satellite (ICESat) acquired from 2003 to 2008. Temporal changes could also lead to inconsistent spatial alignments of BIA locations, ice velocities and the DEM, which could also lead to errors.