Processing Steps
Snow Cover
The MODIS snow cover algorithm detects snow by computing the Normalized Difference Snow Index (NDSI) (Hall and Riggs, 2011) from MODIS Level 1B calibrated radiances. Data screens are then applied to alleviate errors of commission and to flag uncertain snow detections. The final output consists of NDSI snow cover plus the location of clouds, water bodies, and other algorithm results of interest to data users. The following sections briefly describe the approach used to detect snow. For a detailed description, see the Algorithm Theoretical Basis Document (ATBD).
Input Products
Table 3 lists the MODIS products that are used as inputs to the snow detection algorithm:
Product ID
|
Long Name
|
Data Used
|
Table 3. Inputs to the MODIS snow algorithm
MOD02HKM
|
MODIS/Terra Calibrated Radiances 5-Min L1B Swath 500m
|
Band 1 (0.645 μm); Band 2 (0.865 μm); Band 4 (0.555 μm); Band 6 (1.640 μm)
|
MOD021KM
|
MODIS/Terra Calibrated Radiances 5-Min L1B Swath 1km
|
Bands: 31 (11.03 μm )
|
MOD03
|
MODIS/Terra Geolocation Fields 5-Min L1A Swath 1km
|
Land/Water Mask (see note); Solar Zenith Angle; Latitude; Longitude; Geoid Height
|
MOD35_L2
|
MODIS/Terra Cloud Mask and Spectral Test Results 5-Min L2 Swath 250m and 1km
|
Unobstructed Field of View Flag; Day/Night Flag
|
The new land/water mask greatly improves the accuracy of lake and river locations compared with Version 5. Users will likely notice that many larger rivers are more continuous and that the number of mapped lakes has increased, especially in regions with small lakes such as northern Minnesota to the Northwest Territories.
The algorithm reads radiance data from MOD02HKM, geolocation data and the land/water mask from MOD03, and the cloud mask and day/night flag from MOD35_L2. The radiance data is checked for quality and converted to top of the atmosphere (TOA) reflectance. The NDSI is then computed for all land and inland water pixels in daylight using Band 4 (0.55 µm) and Band 6 (1.6 µm) reflectances as follows:
NDSI = (Band 4 - Band 6) / (Band 4 + Band 6)
Snow typically has a very high reflectance in visible bands and very low reflectance in the shortwave infrared, a characteristic which distinguishes snow cover from non snow-covered land and most cloud types. As such, pixels with NDSI > 0.0 are deemed to have some snow present. Pixels with NDSI ≤ 0.0 are classified as snow free land.
In the previous version of this data set (Version 5), fractional snow cover was computed from the NDSI using a regression technique. This approach has been abandoned for Version 6 because the NDSI is directly related to the presence of snow in a pixel and thus more accurately describes snow detection compared with FSC. The MODIS Science Team believes this change will offer users more flexibility to apply MODIS snow cover data sets to their research. Importantly, the change does not disrupt data continuity because the snow detection algorithm in Version 6 is essentially the same as Version 5 without the FSC calculation. Users who wish to estimate FSC can apply the FSC regression equation from Version 5 to Version 6 NDSI snow cover data.
In addition, the binary (snow/no snow) snow-covered area (SCA) map in Version 5 has been abandoned for Version 6. This SDS was computed by: a) setting a snow threshold of 0.4 ≤ NDSI ≤ 1; and b) applying an additional test to pixels with 0.1 ≤ NDSI ≤ 0.4 which used the Normalized Difference Vegetation Index (NDVI) to increase snow detection sensitivity in forested landscapes. However, this algorithm effectively prevented snow detections for NDSI < 0.4 on any landscape. Again, the MODIS Science Team believes this change offers the research community more flexibility. Users who wish to construct a binary SCA map can choose their own threshold for snow using the Version 6 NDSI Snow Cover, the raw NDSI data, or a combination of both.
The NDSI has proven effective at detecting snow cover on the landscape given clear skies and good viewing geometry and solar illumination. However, other illumination conditions can diminish the technique's effectiveness and induce errors of commission or omission. During the course of the MODIS mission, the Science Team and user community have identified several frequently occuring sources or error, for example, confusion between snow-covered land and certain cloud types or surface features with snow-like reflectances.
Examining the NDSI relationship more closely provides a means to circumvent many of these potential errors. For example, some bright surface features with snow-like NDSIs have MODIS Band 6 reflectances that exceed expected values for snow, while others have visible/near-infrared reflectance differences that are too low. As such, pixels determined to have some snow present are subjected to a series of screens that have been specifically developed to alleviate snow commission and omission associated with the most common error sources. In addition, snow-free pixels are screened for very low illumination conditions to prevent possible snow omission errors. The following sections describe these data screens.
Low Visible Reflectance Screen
This screen is applied to prevent errors from occuring when the reflectance is too low for the algorithm to perform well, such as in very low illumination or on surface features with very low reflectance. This screen is also applied to pixels that have no snow cover present (snow-free pixels) to prevent possible snow omission. If the MODIS Band 2 reflectance is ≤ 0.10 or the Band 4 reflectance is ≤ 0.11, the pixel fails the screen and is set to no decision in the NDSI snow cover SDS. The results of this screen are tracked in bit 1 of the NDSI_Snow_Cover_Algorithm_Flags_QA
SDS.
Low NDSI screen
Pixels detected as having snow cover with 0.0 < NDSI < 0.10 are reversed to no snow and flagged by setting bit 2 in the NDSI_Snow_Cover_Algorithm_Flags_QA
SDS. This flag can be used to find pixels where snow cover detections were reversed to not snow.
Estimated surface temperature and surface height screen
This screen serves a dual purpose by linking estimated surface temperature with surface height. It is used to alleviate errors of commission at low elevations that appear spectrally similar to snow but are too warm. It is also used to flag snow detections at high elevations that are warmer than expected. Using the estimated MODIS Band 31 brightness temperature (Tb), if snow is detected in a pixel with height < 1300 m and Tb ≥ 281 K, the pixel is reversed to not snow and bit 3 is set in the NDSI_Snow_Cover_Algorithm_Flags_QA
SDS. If snow is detected in a pixel with height ≥ 1300 m and Tb ≥ 281 K, the pixel is flagged as unusually warm by setting bit 3 in the NDSI_Snow_Cover_Algorithm_Flags_QA
SDS.
High SWIR reflectance screen
This screen also serves a dual purpose by: a) preventing non-snow features that appear similar to snow from being detected as snow; b) allowing snow to be detected where snow-cover short-wave infrared reflectance (SWIR) is anomalously high. Snow typically has a SWIR reflectance of less than about 0.20; however, this value can be higher under certain conditions like a low sun angle. The SWIR reflectance screen thus utilizes two thresholds. Snow pixels with SWIR reflectance > 0.45 are reversed to not snow and bit 4 of NDSI_Snow_Cover_Algorithm_Flags_QA
SDS is set. Snow pixels with 0.25 < SWIR reflectance ≤ 0.45 are flagged as having an unusually high SWIR for snow by setting bit 4 in the NDSI_Snow_Cover_Algorithm_Flags_QA
SDS.
Solar zenith screen
When solar zenith angles exceed 70°, the low illumination challenges snow cover detection. As such, pixels with solar zenith angles > 70° are flagged by setting bit 7 in the NDSI_Snow_Cover_Algorithm_Flags_QA
SDS. This solar zenith mask is set across the entire swath. Note: night is defined as a solar zenith angle ≥ 85°. Night pixels are assigned a value 211.
Lake Ice
Ice/snow covered lake ice are detected by applying the snow algorithm specifically to inland water bodies. These data are provided so that the MODIS user community can evaluate the efficacy of this technique. Inland water bodies are flagged by setting bit 0 in the NDSI_Snow_Cover_Algorithm_Flags_QA
SDS. Users can extract or mask inland water in the NDSI snow cover SDS using this flag. The algorithm relies on the basic assumption that a water body is deep and clear and therefore absorbs all of the solar radiation incident upon it. Water bodies with algal blooms, high turbidity, or other relatively high reflectance conditions may be erroneously detected as snow/ice covered.
Cloud Masking
Clouds are masked using the Unobstructed Field of View (UFOV) cloud mask flag from MOD35_L2. Values in the 1 km mask value are applied to the four corresponding 500 m pixels. If the cloud mask flags “certain cloud,” the pixels are masked as cloud. Values of “confident clear," “probably clear,” or “uncertain clear” are interpreted as clear in the snow cover algorithm.
Abnormal Condition Rules
If radiance data are missing in any of the MODIS bands used by the algorithm, the pixel is set to "missing data" and is not processed for snow cover. Unusable radiance data are set to "no decision."
Version History
See the MODIS | Data Versions page for the history of MODIS snow and sea ice product versions.
Error Sources
Anomalies in the input data can propagate to the output. Table 3 lists the MODIS products that are used as input to the snow cover algorithm. Although developing a global snow cover detection algorithm presents a variety of challenges, the NDSI technique has proven to be a robust indicator. Numerous investigators have utilized MODIS snow cover data sets and reported accuracy in the range of 88% to 93%. Consult the MODIS Snow Products Collection 6 User Guide for more details about potential sources of error in the MODIS snow cover data sets.