The ability of Visible and Near-Infrared (VIS-NIR) satellite sensors, notably the Landsat series and the Advanced Very High Resolution Radiometer (AVHRR), to reveal previously unknown features of the Antarctic continent and its coastline was broadly recognized in the 1980s by glaciologists and cartographers. Later, a host of studies showed how careful processing of image radiometry could provide unprecedented information about the ice sheet surface revealing details of ice flow, sub-ice bedrock structure, and wind-related features in the interior of the ice sheet by detailed portrayal of the subtle surface morphology at the approximately 1 km spatial resolution scale of AVHRR (Orheim and Lucchitta 1988), (Bindschadler and Vornberger,1990), (Scambos and Bindschadler 1991) and (Seko et al. 1993). The U. S. Geological Survey created continent-wide mosaics using AVHRR (USGS 1991) and (USGS 1996). USGS and other groups then created regional ice feature image maps using Landsat (Ferrigno et al. 1994) and (Swithinbank et al.1988).

Recognizing that a red-infrared band combination can provide grain size information, Winther et al. (2001) used the VIS-NIR data from the USGS/Ferrigno AVHRR mosaics to generate an approximate estimation of total blue ice area for the Antarctic continent. This followed earlier experiments at mapping snow grain size and blue ice extent from space (Bourdelles and Fily 1993) and (Orheim and Lucchitta 1990).

The normalized difference band radiance ratio correlates with surface grain size because snow reflectivity decreases in the infrared as grain size increases (Warren 1982), (Fily et al. 1997) and (Painter and Dozier 2004). In detail, this reflectivity change is due to absorptive interactions between infrared light and ice crystal electronic structure. Decreasing reflectivity in the infrared region of the electromagnetic spectrum is the principal observable contributing to the capability to remotely determine snow grain size, and it is a function of the greater mean absorbing path length within a larger grain, that is, increased absorption with longer path length through the crystal structure. However, optical path within ice is also a function of grain shape: complex, feathery grains will have a small optical grain size even though individual grains may have a much larger maximum dimension. Moreover, smaller grain sizes have a larger single-scattering component off the crystal surface. For most albedo or energy balance investigations, optical grain size is the desired parameter, because the fundamental physics of the study concerns ice interacting with light.

Geolocation and Processing

Satellite image data swaths of Band 1 and Band 2 from MODIS Level-1B MOD02QKM (Terra) and MYD02QKM (Aqua) files, together with illumination and viewing angle data from Level-1A MOD03 and MYD03 files, were geolocated and resampled onto the projection grid using NSIDC's MODIS Swath-to-Grid Toolbox (MS2GT) software. The software interpolates the 1 km resolution latitude and longitude data from the Level-1A files to 250 m resolution and then resamples the Level-1B data to the grid using a forward elliptical weighted average (EWA) algorithm (Greene et al. 1986).

Destriping of MODIS Image Data

The MS2GT algorithm was modified to remove striping artifacts incurred by the 40-detector whiskbroom scanner and the two mirror sides of the MODIS Band 1 and Band 2 sensor (Haran et al, 2002). These artifacts are a known problem with all Terra and Aqua MODIS Level-1B data at 250 m (MOD02QKM and MOY02QKM data). Inter-detector variations are as large as 1 percent or 50 Digital Numbers (DN) in the 12-bit MODIS data, contributing to distinct horizontal striping in contrast-enhanced images. This primary striping pattern appears to be caused by poor inter-detector calibration among the 40 detectors that constitute a single scan of the 250 m data. A secondary variation in brightness appears between successive 40-line scans that is due to mirror side effects in the double-sided MODIS scan mirror. A third artifact appears as an every-fourth-pixel brightness shift in detectors 28 and 29 (a stitching artifact in appearance). These three artifacts limit the usefulness of MODIS over ice sheet interior images, because they induce brightness variations as large or larger than the shading due to subtle topography.

To correct the problems caused by artifacts, investigators conducted a Lambertian solar zenith angle normalization on the swath data for both bands. Telemetry noise and line drops in the MODIS scenes, having the appearance of chads in the projected images, were reset to zero, that is, treated as masked cloud areas. Though the mosaic composite is nearly cloud-free, some areas of thin cloud and small fog or cloud patches exist.

Two composite images were created from the geolocated Band 1 and Band 2 images: a high-pass filtered Band 1 image composite, which emphasizes the surface morphology, and a normalized-difference band-ratio image, which provides semi-quantative information about mean surface grain size over snow and ice surfaces.

Cloud Masking

The geolocated swath images were manually masked to remove clouds, cloud shadows, fog, blowing snow, and heavy surface frost. To do this, the images were compared to the RAMP AMM-1 SAR mosaic image and to initial versions of the MOA composite to identify cloud, fog, and blowing snow areas. Refer to the RAMP Basics Web page. Investigators conducted cloud and surface artifact masking using the 750 m images. They then applied the same mask to the 125 m scenes. Some small cloud and fog features at less than 750 m scale were not masked, and in a few persistently cloudy areas, some thin cloud and cloud-shadow contaminated images had to be used to cover the enitre continent with multiple scenes.

The mosaic composite is nearly perfectly cloud-cleared. Some areas of thin clouds, cirrus cloud shadows, and fog or low-lying small clouds are present in the northeastern Ronne Ice Shelf, which was persistently cloudy throughout the 2003-2004 austral summer. In numerous areas, there are small patchy clouds at spatial scales of less than 1 km as a result of using a 750 m resolution image for the cloud clearing. More regionally, some areas are partly impacted by blowing snow, particularly in the East Antarctic.

Compositing the Image Swaths via Data Cumulation

Investigators created two composite images from the geolocated Band 1 and Band 2 images: a high-pass filtered Band 1 image composite, emphasizing the surface morphology; and a normalized-difference band-ratio image that provides quantitative information about mean surface grain size over snow and ice surfaces. The processing and assembly steps of these two composites is discussed separately.

High-Pass Filtered Surface Feature Composite

After cloud masking, the geolocated, destriped Band 1 images were high-pass filtered to reduce non-Lambertian illumination and to reset the mean grayscale range to a common value for compositing. Investigators set the filter size for the images to 511 x 511 pixels, or the ground equivalent of 64 km x 64 km. The mean brightness of these filtered images was set to the same value (the integer 16000) to match gray levels for compositing.

Investigators then applied a weighting scheme to the masked and filtered image swaths, creating a weight image for each gridded image that contains a scalar value for each non-masked pixel in the swaths. Weights for the pixels range from 0 to 50000. The weight was computed as the product of a fractional scan weight, a fractional mask weight, and a scale factor (50000). The scan weight was determined by the proximity to the nadir track, favoring near-nadir areas, and the mask weight by the proximity to an image edge or mask edge to feather the edges of the component images.

wscan, the scan weight image, is computed as follows:

wmask, the mask weight image, is computed as follows:

weight, the weight image, is then:

The high-pass filtered Band 1 and weight images were then combined using stacking techniques, also called image super-resolution or data cumulation. See Scambos et al. 1999 to improve spatial and radiometric detail beyond the 250 meter and 12-bit characteristics of single Band 1 and Band 2 MODIS images. This improvement in part is due to reduction of random image noise by cancellation and improvement of relative radiometric resolution by combining repeated measurements of the surface in the form of multiple digital images. The additional spatial resolution in the stacked image composite is a result of knowing the pixel center locations to a high precision (50 m), a precision that is smaller than the width of the pixel sample area (250 m).

Images were added to the mosaic with weighting applied according to the following scheme:

Given a set of n high-pass filtered Band 1 images (B_i) and a corresponding set of n weight images (W_i), compute the composited band image (B_c), composited weight image (W_c), and count image (N_c).

B_c, W_c, and N_c are initially all zero, and N_i is set to 1 for each pixel. Then, for each i from 1 to n, where B_i is not zero, and W_i is not zero:

A set of n intermediate composites can themselves be composited into a single composite by setting the values of B_i, W_i, and N_i to the corresponding values of B_c, W_c, and N_c for each composite (in this case N_i is not set to 1), and then applying the given algorithm.

The image stacking allows multiple images to contribute to the representation of a single grid cell in the MOA composite. Image count ranges from 38 to 1; 98.53 percent of the imaged area is made up of six or more contributing images. Mean image count is 14.7. Areas of low image count are the northeastern and far northwestern Ronne Ice Shelf, which had persistent cloud cover throughout the summer of 2003-2004, the ice sheet region between the Executive Committee Range and Crary Mountains in West Antarctica, and the area between the Dome Fuji and Plateau Station camps in the East Antarctic Plateau. This area is in the region of local noon for the UTC time range; therefore, it has fewer near-nadir passes. Using the equations and models developed in Scambos et al. 1999, investigators infer that regions with five or more scenes contributing have resolutions of approximately 200 m or better ranging to a best resolution value of approximately 150 m for images composed of 10 or more scenes. Mean weight of the image pixels for the MOA grid cells is 30322, ranging from 4800 to 49980. A large region of lower weights (5000 - 9000) is centered on the South Pole, because all images are significantly off-nadir in this area.

In general, image quality is best in regions with high image counts and high composited weight.

Optical Mean Snow Grain Size

Two simple grain-size composite images were generated by applying a normalized difference algorithm to Band 1 and Band 2 of the composite scenes.

This ratio takes advantage of the decreasing reflectivity of snow in the infrared range, creating an image that is sensitive to grain-size variations (Warren 1982) and (Fily et al. 1997). To maintain a quantitative ratio, these images were not processed beyond geolocation, calibration, and destriping. Calibration is based on the information provided in the MOD02QKM and MYD02QKM files, as applied by the MS2GT software. The image is provided as a 16-bit unsigned integer image of optical mean snow grain size in µm.

A composite image of optical surface grain size was generated by applying a model-derived lookup table to images of normalized difference MODIS Band 1 (red light; 620 nm - 670 nm) and MODIS Band 2 (near-infrared; 841 nm - 876 nm) radiances and solar zenith angle:

G_i,j represents a normalized radiance ratio image, b1_i,j and b2_i,j are the calibrated band radiance values at grid location (i,j) for the respective MODIS bands, Z_i,j is the solar zenith angle image for each MODIS pixel in the scene, f(G_i,j, Z_i,j) is the lookup table, and S_i,j is the optical grain radius image. The spring-only grain size image was constructed from 122 MODIS scenes acquired between 1 November and 17 December 2003. Calibration scale factors are provided in the MOD02QKM and MYDO2QKM data for the Terra and Aqua MODIS sensors, respectively.

Lookup table values for the grain size conversion were derived from runs of the Santa Barbara DISORT Atmospheric Radiative Transfer (SBDART) software (Ricchiazzi et al. 1998). These lookup tables were then applied to the MODIS scenes to produce a corresponding set of grain size images. The grain size images were then composited together using the same weighting and compositing scheme as that used for the surface morphology image, except that in addition to masked areas, values of 5 or 1105 µm were treated as missing data. Values that led to unmodeled grain sizes, that is less than 10 µm or greater than 1100 µm, had the associated grain size value set to marker values of 5 µm and 1105 µm, respectively. Count and weight images for the grain size composites were slightly different from the corresponding surface morphology count and weight images because of these additional missing values. In cases where all images gave out-of-range grain size results, the grid cell value was set to either 5 µm or 1105 µm.

Runs of the SBDART software provided predictions of MODIS Top Of Atmosphere (TOA) Band 1 and Band 2 radiance values for a series of snow grain sizes (10 µm to 1100 µm in 10 µm increments) at a series of solar illumination solar zenith angles from 0 degrees to 89 degrees in 0.1 degree increments. Investigators provided SBDART with snow optical properties, that is, BRDF spectral reflectance information from a recent snow reflectance model (Painter and Dozier 2004), and created files (one for every 10 µm of grain size) that used the data to represent the reflecting substrate in the atmospheric radiative transfer model. Investigators held the sensor viewing angle at nadir for all SBDART evaluations. The weighting scheme favoring nadir-viewing conditions was appropriate to reduce off-nadir BRDF effects in the final optical snow grain size composite.

To validate the optical grain size scheme, researchers with the Antarctic Remote Ice Sensing Experiment used a field spectrometer to collect six surface-snow spectral measurements during three locally clear-sky days in October 2003 (Massom et al. 2006). MOA investigators averaged 60 separate spectra of snow (spectral range 0.4 to 2.5 µm), normalized to 40 interspersed spectra of an Etalon calibration target, and inverted the data to the optically equivalent grain-size radius using conversions developed by Nolin and Dozier (2000) with 50 µm - 100 µm grain radius uncertainties. The same sites were imaged by six MODIS scenes, five from the Terra platform and one from Aqua. In situ grain size radius ranged from 66 µm, for snow immediately following a storm accumulation, to 170 µm, for wet snow in warmer conditions. Comparing grain sizes predicted by the method used for these images with in situ measurements shows that, in general, the satellite algorithm underestimates grain size by 15 µm to 151 µm.

Validation data for the optical grain size measurement are shown in Table 5. Same-day sites are within 100 m of each other. SGS refers to Snow Grain Size.

Error Sources

Wolfe et al. (2002) estimated the accuracy of the Level-1A geolocation data to be 50 m, considerably better than the Level-1B ground-equivalent nadir pixel size of 250 m.

The accuracy and precision of this geolocation was tested using known surface sites, such as South Pole Station, Vostok Station, Siple Dome camp and traverse trail, Megadunes Camp runway, and Dome Concordia camp, and areas of well-mapped coastline such as Ross Island and the northern Antarctic Peninsula. Investigators did not find discrepancies greater than 125 m in the projected location of a fixed object among the 260 scenes or relative to well-mapped coastline positions. Furthermore, identical features in areas where separate images overlap on the grid align to within one grid cell. As such, in areas where the position of mapped coast differs from the MOA image, the investigators believe that the MOA is more accurate relative to, for example, the CIA coastline database or the Antarctic Digital Database.

Discrepancies did not exceed 125 m in the projected location of a fixed object among the 260 scenes or relative to well-mapped coastline positions. Further, overlapped areas of separate images showed identical feature locations on the grid to within one grid cell.

The mean image weight and image count 750 m data products are provided as a means for users to assess image quality in various parts of the MOA. In general, the MOA quality is higher in areas of both high count and high weight.

Three sources of error are described below: geolocation error, surface obscurations, and snow grain size error.

The component MODIS scenes and the final MOA mosaic are geolocated to within an estimated 50 m, considerably less than the satellite image pixel size or the final grid spacing. This estimate was checked by comparisons with coastlines, field camps, mountain peaks, and other mosaics and maps. Coastline and grounding line files are estimated to be accurate, tracking the best estimate of these features, to within 250 m, or two grid cells. The coastline grounding-line interpretation may not be accurate in all areas.

The mosaic composite is almost perfectly cloud-cleared. Some areas of thin clouds, cirrus cloud shadows, and fog or low-lying small clouds are present in the northeastern Ronne Ice Shelf region, which was persistently cloudy throughout the 2003-2004 austral summer. Other known sites where clouds remain are the grounding line near Bailey Ice Stream (79.67° S, 33.1° W), the ice tongue of Jutulstraumen Glacier (70° S, 0° E), and the Mobiloil Inlet and adjacent Solberg Inlet (68.4° S, 66.5° W). In numerous areas, small patchy clouds and shadows, less than approximately 1 km in size, resulted from using the 750 m resolution images for cloud evaluation and masking. In addition, the investigators identified several areas with features that appear to be blowing snow on the scale of hundreds of square kilometers, particularly in East Antarctica over the upper slopes of the ice sheet. Blowing snow appears as low-contrast mottlings of the surface, often arranged in approximately linear bands oriented near the mean katabatic wind direction, as mapped by Parish and Bromwich (1991). Also, widespread, low-contrast artifacts from hoar frost patches can be seen in regions where fog or emerging vapor from the snowpack have formed frost crystals on the surface. In general, the artifacts appear as sharp-edged patches, often with a sawtooth or flame-like outline, that cross-cut undulations or other topography on the ice surface. Images from late spring or early summer were selected where possible to reduce the number of hoar patches, and then averaged over a sufficiently long period of time (frost patches change on a scale of days) to reduce the intensity of these features in the composite image.

Optical snow grain size error is approximately +/- 50 micrometers, estimated by compariing in situ spectra of varying snow grain sizes with near-simultaneous MODIS images that were processed in the same manner as the MOA grain size composite scenes. However, snow grain size varies greatly over the period of image acquisition for the MOA. As such, large ranges of snow grain sizes were averaged together in some areas, for example, in melting areas or warm but sub-freezing areas that experienced numerous snowfalls followed by snow diagenesis