Satellites are well suited to the measurement of snow cover because the high albedo of snow presents a high contrast with most other natural surfaces except clouds. Spectral reflectivity of snow depends on grain size and shape, impurity content, liquid water content, depth, surface roughness, and solar elevation angle (Hall and Martinec 1985). Reflectance of fresh snow is very high in the visible wavelengths, but decreases in the near-infrared wavelengths especially as grain size increases. Because of natural aging and other factors such as soot or volcanic ash deposition, reflectance of snow decreases over time. Fresh snow can have a reflectance up to about 80 percent, but its reflectance may decrease to below 40 percent after snow crystals metamorphose (Hall et al. 1998).

Snow and Cloud Discrimination
Snow and cloud discrimination techniques are based on differences between cloud and snow/ice reflectance and emittance properties. Clouds typically have high reflectance in visible and near-infrared wavelengths, while reflectance of snow decreases in near-infrared wavelengths.

Special Considerations for Dense Forests
The mapping of snow cover becomes limited in areas where snow cover is obscured by dense forest canopies. A forested landscape is never completely snow-covered because tree branches, trunks, and canopies may not be covered with snow. Often, in boreal forests, snow that falls on the coniferous tree canopy will not stay on the canopy for the entire winter because of sublimation. Thus, even in a continuously snow-covered area, much of the forested landscape will not be snow-covered. Furthermore, snow that falls onto the ground through the canopy may not be visible from above.

A canopy reflectance model (GeoSAIL) for discontinuous canopies is the basis for determining the fraction of sunlit crown, sunlit background, crown reflectance, canopy transmittance, shadowed crown, and shadowed background within a forest stand. Reflectances for the sunlit snow are calculated using the Wiscombe and Warren (1980) model, while reflectances from other components were measured from direct observations. A significant difference exists between the high reflectance of snow and the low reflectance of soil, leaves, and bark. Depending on the vegetation type, snow may also cause a decrease in the mid-infrared reflectance of the forest stand. In addition, reflectance in the visible spectrum often increases with respect to the near-infrared reflectance. This lowers the Normalized Difference Vegetation Index (NDVI), a complement to the NDSI.

The NDSI and NDVI are used together to discriminate between snow-free and snow-covered forests (Klein, Hall, and Riggs 1998). Forested pixels have higher NDVI values compared with non-forested pixels. Thus, by using the NDSI and NDVI in combination, it is possible to lower the NDSI threshold in forested areas without compromising the algorithm performance in other land covers (Hall et al. 1998).

The MODIS science team is responsible for algorithm development. The MODIS Data Processing System (MODAPS) is responsible for product generation and transfer of products to NSIDC.

The snow-mapping algorithm, based on the NDSI, identifies the presence of snow by reflectance or radiance properties in each 500 m pixel. The NDSI is a ratio of the difference in reflectance of snow in the visible and near-infrared wavelengths. The NDSI partially compensates for a number of illumination conditions including atmospheric effects. The algorithm uses MODIS Bands 4 (0.55 µm) and 6 (1.6 µm) from MYD02HKM to calculate the NDSI (Hall et al. 1998).

NDSI = (Band 4 - Band 6) / (Band 4 + Band 6)

The fractional snow cover map is based on the regression technique of (Salomonson and Appel 2004). The fractional area (in percent) of each pixel covered by snow is calculated for both land and inland water bodies not covered by cloud and over the range of NDSI values from 1-100. Fractional snow may be mapped over the whole NDSI range indicative of snow (Salomonson and Appel 2006).

Snow Fraction = -0.01 + 1.45 * NDSI

Processing Steps

Analysis for snow in a MODIS swath is constrained to pixels that:

• have nominal Level-1B radiance data
• are in daylight
• are over land or inland water
• are unobstructed by clouds
• have an estimated surface temperature less than 283 K

Constraints are applied in the order listed. Only pixels having a daylight, clear-sky view of the land surface are then analyzed for snow. The NDSI can generally separate snow from most obscuring cumulus clouds, but it cannot always discriminate optically-thin cirrus clouds from snow. Instead, clouds are masked using data from the MODIS Cloud Mask data product (MYD35_L2). If the cloud mask algorithm was not applied to a MYD10_L2 pixel, the snow algorithm proceeds while assuming that the pixel is unobstructed by cloud (Hall and Riggs 2006).

V005 data uses the summary cloud result field and the unobstructed field-of-view flag from MYD35_L2 to generate a single cloud mask for the snow algorithm. In some cases, MYD35_L2 identifies snow as clouds, which prevents the snow algorithm from mapping true snow extent.

A 1 km resolution land/water mask within the MODIS geolocation product (MYD03) is used to mask oceans and inland water. Ocean pixels are not analyzed for snow. Inland water pixels are analyzed for the condition of snow-covered inland water (primarily lakes).

MYD02HKM Level-1B data are screened for missing and unusable data. Unusable data result if sensor radiance data fail to meet acceptable criteria during processing. If unusable data are encountered, then a no decision result is written for the affected pixels. Similarly, pixels are labeled missing data if missing data are encountered.

Snow detection is achieved through the use of two groups of criteria tests for snow reflectance characteristics in the visible and near-infrared regions. One group is for detection of snow under a wide range of conditions. A pixel is identified as snow if all the following conditions are met:

• (Band 4 - Band 6) / (Band 4 + Band 6) is greater than or equal to 0.4
• Band 2 reflectance is greater than 0.11
• Band 4 reflectance is greater than 0.10

Another group of criteria tests is used to better detect snow in dense vegetation, such as forests. In this case, a pixel is identified as snow if the following conditions are met:

• The pixel has NDSI and NDVI values in a defined polygon in a scatter plot of the two indices
• Band 1 reflectance is greater than 0.10
• Band 2 reflectance is greater than 0.11

This last set of criteria is applied without regard to the ecosystem, so it should not be interpreted strictly as snow-covered forest. Snow-covered ice on inland water is determined by applying the first group of criteria tests used for snow detection to pixels mapped as inland water by the land-water mask.

Intermediate checks for theoretical bounding of reflectance data and the NDSI ratio are made in the algorithm. In theory, reflectance values should lie within the 0-100 percent range, and the NDSI ratio should lie within the -1.0 to +1.0 range. Summary statistics are kept for pixels that exceed these theoretical limits; however, the test for snow is done regardless of violations of these limits.

The snow algorithm also identifies missing data and reports them in the output product. Certain expected anomalous conditions may exist with the input data, such as a few missing lines or unusable data from the MODIS sensor. In these cases, the snow algorithm makes no snow decision for an affected pixel. The pixel value is set to missing data or no decison and the algorithm moves to the next pixel. The Valid EV Obs Band x and Saturated EV Obs Band x local attributes are propogated from the MYD02HKM input product to assist in QA and analysis (Riggs, Hall, and Salomonson 2006).

Error Sources

The snow mapping algorithm incorporates tests for known anomalous conditions. If input data are missing, the snow mapping algorithm indicates that in the output product. Where dead detectors are found, the result is a no decision result. No averaging or other processing is done for dead detectors. If other known anomalous conditions are encountered, the snow mapping algorithm makes no decision for that pixel.

As with any upper level product, the characteristics of and anomalies in input data may carry through to the output data product. The following products are input to the snow mapping algorithm:

Note: NSIDC does not archive or distribute the following products, and does not maintain the links to these products. Thus, if a link does not work, please contact the MODAPS or the LP DAAC.

• MYD02HKM - MODIS/Aqua Calibrated Radiances 5-Min L1B Swath 500m, Version 5
• MYD021KM - MODIS/Aqua Calibrated Radiances 5-Min L1B Swath 1km, Version 5
• MYD03 - MODIS/Aqua Geolocation Fields 5-Min L1A Swath 1km, Version 5
• MYD35_L2 - MODIS/Aqua Cloud Mask and Spectral Test Results 5-Min L2 Swath 250m and 1km, Version 5

Errors may exist in the reflectance calculations used to determine whether snow is present due to the anisotropy of snow and ice. Snow is not a Lambertian Reflector because snow reflects more in a forward direction. As snow ages, its anisotropy increases. The increase in forward scattering with snow age is greater in the near infrared wavelengths, relative to the visible wavelengths. Such errors will likely be greater at larger angles, 30 percent or more off nadir, as the amount of reflected solar irradiance varies with view angle. Additionally, errors in precise reflectance value due to anisotropy related to topographic variability is inherent in the data (Hall et al. 2001a).

The snow-mapping algorithm is sensitive to small changes in NDSI or Normalized Difference Vegetation Index (NDVI) particularly over dark, dense vegetation. Since this can result in erroneous snow detection, particularly over the dark, dense forests of the tropics, a thermal threshold of 283 K is used. If the surface has an estimated temperature greater than 283 K, then it will not be mapped as snow. This threshold is several degrees above the melting temperature of snow since the effective brightness temperature of a pixel may be high due to mixed contributions in the field of view, for example, warm tree crowns in a snow-covered region or sun warmed rock in regions of patchy snow. Thus, bright warm surfaces that have similar spectral characterists as snow are removed because of the screen.

Errors with the snow mapping algorithm are lowest in tundra and prairie regions. The maximum expected errors are 15 percent for forests, 10 percent for mixed agriculture and forest, and 5 percent for other land covers. Estimating snow cover is difficult in forests because trees partially or completely conceal underlying snow. The expected maximum monthly and annual errors in Northern Hemisphere snow-mapping methods from the algorithm have been estimated. The maximum monthly errors are expected to range from 5 percent to 9 percent for North America, and from 5 percent to 10 percent for Eurasia. The maximum aggregated Northern Hemisphere snow mapping error is estimated to be 7.5 percent. The error is highest, around 9 percent to 10 percent, when snow covers the Boreal Forest roughly between November and April (Hall and Riggs 2006).

Quality Assessment

Quality indicators for MODIS snow data can be found in three places:

• AutomaticQualityFlag and the ScienceQualityFlag metadata objects and their corresponding explanations: AutomaticQualityFlagExplanation and ScienceQualityFlagExplanation located in the CoreMetadata.0 global attributes
• Custom local attributes associated with each SDS, for example snow cover
• Snow Cover Pixel QA data field.

These quality indicators are generated during production or in post-production scientific and quality checks of the data product. For more information on local and global attributes, go to one of the following links:

• MOD10L2 and MYD10L2 Local Snow Cover Attributes, Version 5
• MOD10L2 and MYD10L2 Global Attributes, Version 5

The AutomaticQualityFlag is automatically set according to conditions for meeting data criteria in the snow mapping algorithm. In most cases, the flag is set to either Passed or Suspect, and in rare instances, it may be set to Failed. Suspect means that a significant percentage of the data were anomalous and that further analysis should be done to determine the source of anomalies. The AutomaticQualityFlagExplanation contains a brief message explaining the reason for the setting of the AutomaticQualityFlag. The ScienceQualityFlag and the ScienceQualityFlagExplanation maybe updated after production, either after an automated QA program is run or after the data product is inspected by a qualified snow scientist. Content and explanation of this flag are dynamic so it should always be examined if present in the external metadata file.

The snow algorithm identifies missing data and reports them in the output product. Certain expected anomalous conditions may exist with the input data such as a few missing lines or unusable data from the MODIS sensor. In these cases, the snow algorithm makes no snow decision for an affected pixel. Summary statistics are calculated for these conditions and reported as Valid EV Obs Band x percent and Saturated EV Obs Band 1 percent local attributes (Riggs, Hall, and Salomonson 2006). In addition to these data values, the product contains quality information at the pixel level.

The Snow Cover Pixel QA data field provides additional information on algorithm results for each pixel within a MODIS scene and is used as a measure of usefulness for snow-cover data. The QA data are stored as coded integer values and tell if algorithm results were nominal, abnormal, or if other defined conditions were encountered for a pixel (Riggs, Hall, and Salomonson 2006). For example, intermediate checks for theoretical bounding of reflectance data and the NDSI ratio are made in the algorithm. In theory, reflectance values should lie within the 0-100 percent range, and the NDSI ratio should lie within the -1.0 to +1.0 range. Summary statistics are kept for pixels that exceed these theoretical limits; however, the test for snow is done regardless of violations of these limits.

The NASA Goddard Space Flight Center: MODIS Land Quality Assessment Web site provides updated quality information for each product.