Thermal and geometric thresholds in the mapping of snow with MODIS

Jonathan S. Barton | Dorothy K. Hall | George A. Riggs

One of the problems facing the MODIS snow-mapping algorithm is the mapping of snow in regions where it is known not to exist. One of the more common locations for this problem is in dark forests, particularly in the tropics. The nature of the snow-mapping algorithm is such that it is particularly sensitive to small changes in the NDSI or NDVI over dark, heavy vegetation, and can more easily be tricked by various agents. Speculatively, a list of these agents might include aerosols in the atmosphere above the canopy; water under the canopy, a common occurrence in many rainforests; shadows of clouds falling on the canopy; and clouds that have not been identified by the MODIS cloud mask.

Figure 1. Distribution of 11.03m thermal brightness temperature for 350,000 pixels falsely identified as snow in nine scenes of central Africa.

To correctly map these areas of tropical forest as snow-free, the 11.03m thermal infrared band was used to estimate the surface temperature of the ground. This band was selected because it represents an atmospheric window, in which little of the emitted thermal radiation is absorbed by the atmosphere. The distribution of brightness temperatures was determined for a population of more than 360,000 misidentified pixels taken from nine five-minute granules of central Africa in February 2001. The mean of this sample fell at 289.3K, and 99% of the population lay greater than 279.5K. (Figure 1.) The distribution of brightness temperatures for a population of more than 3,500,000 pixels, assumed to be all correctly identified, from two scenes of northern North America was determined for comparison. These data, in a bimodal distribution, had a mean of 267.9K and 99% of the pixels had brightness temperatures less than 275.9. (Figure 2.) Using these results, a tentative threshold has been set for 277K, which lies between the 99% cutoffs of the two datasets. (Figure 3.) It is possible that some of the pixels in North America were not correctly mapped as snow, but in principle, because the number of snow pixels is so large, if the relatively few incorrectly mapped pixels were removed, it should not substantially alter the distribution. The bimodality of the North American data can be directly attributed to the two different scenes used for the dataset. If more datasets were used, the bimodality might disappear. It is unclear what role weather might play in determining the brightness temperature, and a warm day after a large snowstorm may prove troublesome. When this threshold is applied to the Congo data, it eliminates from 93% to 98% of the incorrectly mapped snow.

Figure 2. Distribution of 11.03m thermal brightness temperature for 3,500,000 pixels identified as snow in two scenes of northern North America.

To further reduce the amount of false snow, a second test was devised. It was observed visually that there is a weak relationship between scan angle and the presence of false snow, with more false snow mapped at higher scan angles. Numerical tests show that from 35% to 60% of the snow in a given scene can be found at scan angles greater than 45°. It is also observed in the data that almost all of the false snow is mapped on the sunward side of the sensor. This phenomenon is probably due to forward scattering off of various atmospheric or ground-level agents. For these reasons, the scan angle threshold was set at 45° on the sunward side, and at the maximum, 55°, looking away from the sun.

Figure 3. Comparison of the two distributions shown in Figures 1 and 2. Note that those figures gave the actual number of pixels on the y-axis, while this figure is normalized by total number of points, for visibility.

Although the scan angle threshold does not have a great deal of predictive power on its own, a combination of these two tests proves to be more powerful than either of its components. As can be observed in Figure 4, showing misidentified snow from 19 February 2001 (10:15Z), although there are many pixels with brightness temperatures less than 277K (the thermal mask has only a 93.5% efficiency in this scene), by eliminating the points above 45° as well, the efficiency is raised to 99.2%. Figure 5 shows another example from 16 February 2001 (9:45Z) that shows improvement provided by using both masks, though the thermal mask alone deals with 99.3% of the misidentified pixels, and adding the scan angle mask improves this to 99.7%.

Figure 4. Relationship between thermal brightness temperature and scan angle for 19,470 pixels misidentified as snow in the 19 February 2001 (10:15Z) scene of central Africa. Note that very cold brightness temperatures only occur at very high scan angles.
Figure 5. Relationship between thermal brightness temperature and scan angle for 41,819 pixels misidentified as snow in the 16 February 2001 (9:45Z) scene of central Africa. Note that most of the pixels are warmer than 277K on this plot.

Future studies will focus on how this thermal/scan angle mask will affect the mapping of snow in temperate regions, such as North America, particularly in the U.S. Pacific Northwest, and the U.S. South, where there is currently also mapping of snow in regions known to be snow free. Another area of interest will be the central Great Plains where MODIS snow lines agree well with other published snow maps, so observation will be necessary to ensure that this mask will not erode the measured snow line dramatically.

Tests using the 11.03m thermal emissivity band and the scan angle of the MODIS sensor in central Africa suggest that these parameters can be used to eliminate approximately 99% of false snow mapped in these tropical regions. It remains to be seen if these tests can be successfully applied globally, and in particular in areas where snow actually exists.