Close

Service Interruption


Quality Control Summary for Northern Hemisphere EASE-Grid 2.0 Weekly Snow Cover and Sea Ice Extent, Version 4 Data

by Mary J. Brodzik

For more information about Northern Hemisphere EASE-Grid 2.0 Weekly Snow Cover and Sea Ice Extent, Version 4 data, please refer to the complete data set documentation.



Table of Contents:

  1. Introduction
  2. Solutions to Quality Problems
  3. Side Effects
  4. Summary
  5. Acknowledgements

1. Introduction

New Data Set Version, New EASE-Grid Version

Please note that the term EASE-Grid is used throughout this document to refer to the target grid, regardless of the processing version of the data set. This data set was gridded to the original EASE-Grid through Version 3.1, and was changed to EASE-Grid 2.0 beginning beginning with Version 4.

Summary of Quality Problems

The NSIDC development team encountered four categories of problems during the production of the Northern Hemisphere EASE-Grid Weekly Snow Cover and Sea Ice Extent Version 4 product:

  1. A "hole" of missing data at the pole.
  2. Shallow coastal water on inland lakes during summer was misclassified as ice.
  3. Ocean pixels interior to the ice pack were not classified as ice.
  4. Land pixels above the snow line were not classified as snow.

The effects of all four problems can be seen in Figure 1. (Click on the image for a larger version).

Zoomed Image of Raw, Regridded Combined
Data for June 17-23, 1996

Figure 1. Zoomed image of raw, regridded combined data for June 17-23, 1996

2. Solutions to Quality Problems

Missing Data at the North Pole

Neither the Nimbus-7 nor the DMSP satellites pass directly over the pole. To eliminate the resulting "hole" at the North Pole, all regridded ocean pixels above 83 degrees North latitude were reset to sea ice (assigned the byte value for QC sea ice).

Inland Lake Sea Ice Misclassification

For Version 3 processing, one of the main reasons that we switched the input sea ice data to the Sea Ice Concentrations from Nimbus-7 SMMR and DMSP SSM/I-SSMIS Passive Microwave Data time series was the improved algorithm for detecting and eliminating coastal contamination and weather, both of which resulted in misclassifications as ice in the earlier version of this data set. Beginning with Version 3, the sea ice data did not suffer from these problems. See Figure 2 for a comparison of Version 2 data and Version 3 data. (Click on the images for larger versions).

Sea Ice Example, Version 2

Figure 2a. Version 2 data, December 21-27, 1998

Sea Ice Example, Version 3

Figure 2b. Version 3 data, December 21-27, 1998 (Note sea ice improvements in weather filter and along coastlines)

Difference between Version 2 and Version 3

Figure 2c. Difference image

However, the sea ice data used for Version 3 and Version 4 still returned "ice" pixels on inland lakes in the summertime (see the coastlines of the Great Lakes in Figure 1). The following monthly sea ice extent climatology was used to reset likely misclassified ice pixels back to open ocean. The sea ice extent climatology represents the monthly maximum extent of sea ice derived from satellite passive microwave sensors (SMMR and SSM/I) from 1978 through 2003. Weekly files that span two months are filtered with the climatology for the month that includes four or more days of the week.

Monthly Sea Ice Climatology Masks

Figure 3. Monthly Sea Ice Climatology Masks (Click on the image for a larger version).

Open Ocean Pixels within the Ice Pack

The input sea ice data land mask classified pixels as coastline, land, or ocean; it differed slightly from the EASE-Grid land-ocean-coastline-ice (LOCI) mask, resulting in consistent open ocean pixels along coastlines within the ice pack. In order to minimize these differences, the sea ice land mask was systematically compared to the EASE-Grid land mask prior to regridding.

Each EASE-Grid ocean pixel location was regridded to the source land mask. The EASE-Grid pixel was identified as a mismatch if the source mask was tagged as either coastline or land, because nearest neighbor regridding would never classify the mismatched pixels as ice. The mismatched pixels were then further categorized.

  1. one-neighbor category: Mismatched pixels in this category had at least one of the immediately surrounding eight neighbor source mask pixels (plus or minus one row or column from the mismatch) classified as ocean.
  2. two-neighbor category: Mismatched pixels in this category had at least one of the immediately surrounding 24 neighbor source mask pixels (plus or minus two rows or columns from the mismatch) classified as ocean.
  3. "others" category: Mismatched pixels in this category were the remainders (i.e, those pixels for which more than plus or minus two rows or columns were searched to find an ocean pixel in the source mask).

The resulting mismatched pixels are displayed by category in Figures 4a and 4b. (Click on the image for a larger version).

EASE-Grid vs. Sea Ice Data Land Mask Differences
Figure 4b. EASE-Grid vs. Passive Microwave Data Set Land Mask Differences
Figure 4a. EASE-Grid vs. Passive Microwave Data Set Land Mask Differences, V3.1 Figure 4b. EASE-Grid 2.0 vs. Passive Microwave Data Set Land Mask Differences, V4

Figure 4a shows EASE-Grid vs. Passive Microwave Data Set Land Mask Differences for Version 3.1 data. Mismatched EASE-Grid ocean mask pixels that are not ocean in the passive microwave sea ice data land mask are shown in yellow (for mismatched pixels with an ocean neighbor one pixel away), pink (for mismatched pixels with an ocean neighbor two pixels away), or red (for mismatched pixels with no ocean neighbor less than three pixels away).

Figures 4b shows EASE-Grid 2.0 vs. Passive Microwave Data Set Land Mask Differences for Version 4 data. Mismatched EASE-Grid ocean mask pixels that are not ocean in the passive microwave sea ice data land mask are shown in yellow (for mismatched pixels with an ocean neighbor one pixel away), pink (for mismatched pixels with an ocean neighbor two pixels away), or red (for mismatched pixels with no ocean neighbor less than three pixels away).


After visual inspection, we decided that pixels in the first two categories were acceptable coastline errors resulting from three different land masks in two projection/grids derived from various coastline databases and algorithms. The most significant areas in the third category included small lakes (in northwestern Russia and central Canada) that simply were not included in the source masks. Pixels in this category have been assigned a byte value that corresponds to "unclassifiable water."

The pixel classification in the first two categories was used during the actual ice data regridding. Up to seven SMMR or SSM/I daily sea ice concentration grids were combined for each corresponding week of NOAA snow data. A pixel was classified as sea ice if it was at least 15 percent sea ice for at least half of the component days. These interim data were then regridded to the NL EASE-Grid via nearest neighbor interpolation. When the target EASE-Grid pixel mapped to an ice pixel in the source grid, it was set to ice (to retain as much of the original information as possible). However, when the EASE-Grid pixel was in the "one-neighbor category," it was classified as sea ice (assigned the byte value for QC sea ice) if any of the eight adjacent source grid interim pixels were sea ice. Similarly, when the EASE-Grid pixel was in the "two neighbor category," it was classified as sea ice (assigned the byte value for QC sea ice) if any of the 24 adjacent source grid interim pixels were sea ice.

Pixels set to sea ice via simple nearest neighbor regridding are distinguished from pixels set during the QC processing by the value for sea ice vs. QC sea ice.

Open Land Pixels Above Snow Line

Source snow chart pixels ranged in size from 125 x 125 km to 205 x 205 km. EASE-Grid pixels were 25 x 25 km. This regridding resolution difference, combined with the NOAA ocean mask, resulted in consistent open land pixels above the "snow line" (Fig. 1). In order to minimize these differences, the NOAA ocean mask was systematically compared to the EASE-Grid mask prior to regridding.

Each EASE-Grid land pixel location was regridded to the NOAA mask and identified as a mismatch if the NOAA mask was classified as ocean. If the source snow data only occurred in pixels identified as NOAA mask land, nearest neighbor regridding would never classify the mismatched pixels as snow. To retain as much of the actual data as possible, these mismatched pixels were further classified.

  1. lower latitude snow neighbor category: pixels with at least one of the immediately surrounding eight neighbor NOAA mask pixels (plus or minus one row or column from the mismatch) at a lower latitude and classified as land
  2. any snow neighbor category: the remaining mismatched pixels, (i.e. those pixels that failed the first category test by having only higher latitude NOAA land neighbors or no land neighbors)

The resulting mismatched pixels are displayed by category in Figures 5a and 5b. (Click on the image for a larger version).

EASE-Grid vs. NOAA Ocean Mask Differences
EASE-Grid 2.0 vs. NOAA Ocean Mask Differences, Version 4 Data
Figure 5a. EASE-Grid vs. NOAA Ocean Mask Differences, V3.1 Figure 5b. EASE-Grid 2.0 vs. NOAA Ocean Mask Differences, V4

Figure 5a shows EASE-Grid vs. NOAA Ocean Mask Differences for Version 3.1 data. Mismatched EASE-Grid land mask pixels that are NOAA ocean are shown in yellow (for mismatched pixels with a lower latitude land NOAA neighbor one pixel away), or pink (for the remaining mismatched pixels).

Figure 5b shows EASE-Grid 2.0 vs. NOAA Ocean Mask Differences for Version 4 data. Mismatched EASE-Grid land mask pixels that are NOAA ocean are shown in yellow (for mismatched pixels with a lower latitude land NOAA neighbor one pixel away), or pink (for the remaining mismatched pixels).


The NOAA mismatched pixel classification was then used during the actual snow data regridding. The NOAA snow chart data were regridded to the NL EASE-Grid via nearest neighbor interpolation. When the EASE-Grid pixel mapped to a NOAA pixel that was classified as snow, it was simply set to snow (to retain as much of the original information as possible).

When the EASE-Grid pixel was not classified snow via nearest neighbor regridding, but was in the "lower latitude snow neighbor" category, it was classified as snow (assigned the byte value for QC snow) if any of its adjacent lower-latitude pixels were snow. This test for lower-latitude snow neighbors (rather than "any" snow neighbor) was introduced to avoid erroneously "growing" the southernmost edge of the snow line. EASE-Grid pixels not classified snow via nearest neighbor regridding, but in the "any neighbor snow category," were classified snow (assigned the byte value for QC snow) if any of the eight adjacent source grid pixels were snow.

Some fixed pixel classifications were also introduced. This data set is designed to represent seasonal fluctuations in snow cover. The Greenland ice sheet and other areas classified as permanent snow or ice in the BU-MODIS mask are always masked as snow covered (assigned the byte values for QC snow). These fixed pixels are indicated in the land mask file EASE2_N25km_loci_land50_coast0km.720x720.bin that is included in the tools directory of the data set distribution.

Finally, there were some land areas (at sub-pixel resolution with respect to the NOAA grid) that failed the snow mismatches test, and resulted in year-round "snow-free" areas, even when the surroundings were classified snow or QC snow. Through visual inspection, we identified the area of Coat's Island (in Hudson's Bay, Canada) to be reset to QC snow if the nearby higher-latitude land was snow or QC snow.

3. Side Effects

The object of the QC algorithms was to create the most realistic snow and ice maps by reasonably compensating for the differences in resolution and source grid projections. However, we have noted an unfortunate side effect in Iceland during the summer months. The Vatnajökull Glacier in southeast Iceland is large enough to be classified snow in the NOAA snow maps. Our QC procedures will therefore set the adjoining higher latitude pixels in Iceland to QC snow. This is obviously incorrect, since the rest of Iceland does experience a snow-free summer. By way of this example, users should be made aware of the overall limitation of the snow classification in this data set, and note that this data set is not intended for such small, relatively sub-resolution area studies, but rather for continental-scale studies of snow cover over time.

4. Summary

The combination of all quality control algorithms is shown in Figure 6. Figure 6a is post-quality control version of Figure 1, with the various QC pixels indicated separately. Figure 6b is of the same data, with the various QC pixels displayed as the new classification (QC snow pixels are displayed as snow, etc). After visual inspection of the entire time series, we are confident that the majority of gross errors due to regridding and algorithm misclassifications have been corrected.

Zoomed Image of Post-QC Combined Data for June 17-23, 1996

Figure 6a. QC pixels indicated separately

Zoomed Image of Post-QC Combined Data for June 17-23, 1996

Figure 6b. QC pixels indicated as corresponding category

Figure 6 shows zoomed images of post-QC combined data for June 17-23, 1996. 6a: Snow QC pixels are indicated in red, sea ice QC pixels in light pink, and ocean QC pixels in bright pink. 6b: All QC pixels are indicated as corresponding type (i.e. ocean QC pixels are displayed as open ocean, snow QC pixels are displayed as snow, and ice QC pixels are displayed as ice). (Click on the images for larger versions).

5. Acknowledgements

The EASE-Grid land-ocean-coastline-ice (LOCI) mask used for Version 3 was derived from the K. Knowles (2004). The version of the LOCI mask used for this data set was derived with no coastlines, so pixels were classified land, ocean (water), or permanent ice. Version 4 was derived from M. J. Brodzik (2012). EASE-Grid 2.0 Land Cover Data Resampled from Boston University Version of Global 1 km Land Cover from MODIS 2001, Version 4. Boulder, Colorado USA: National Snow and Ice Data Center.


Go to the complete data set documentation.