Cloud Masking Improves MODIS Products Accuracy
One of the biggest hindrances to satellite remote sensing
of hydrological, ecological, and even atmospheric conditions
is our very cloudy Earth. To a cloud microphysicist, routine
satellite observations of clouds might seem terrific. But
what if you are a climatologist measuring snow pack in the
Arctic? Or an ecologist mapping forest cover in Madagascar?
Or an oceanographer, measuring ocean temperatures during El
Niño events? In those cases, clouds are simply a nuisance.
From a satellites perspective, they can look like snow,
they can block the view, and they can make ocean temperatures
seem colder than they are.
To help scientists decide if MODIS measurements of a
particular location at a specific point in time were based
on a cloud-free view of the Earths surface, Paul Menzel
and his colleagues at the University of Wisconsin-Madison
have developed the MODIS Cloud Mask. The product is called
a mask because it gives scientists the opportunity to block
out, or mask, pixels within a data set that might be contaminated
by clouds. (See inset for a discussion of pixels).
|MODIS collects data by pixels.
In that sense, then, it is similar to a digital camera,
where an image is made up of row upon row of thousands,
sometimes millions, of tiny little boxes filled with patterns
of color. These little boxes are called pixels. MODIS
collects data (makes images) of the Earth using pixels
of different sizes. The smallest pixels represent an area
on the Earth 250m by 250 mthe largest, 1km by 1km.
All MODIS scientists, whether they ultimately
want to know about vegetation, temperature, phytoplankton,
or snow cover, start with the same basic measurements: radiances,
or the amount of light or heat energy registered by MODIS
photon detectors for a particular Earth scene pixel.
Scientists convert radiances to other land, ocean, or atmospheric
characteristics using what they already know about how different
objects interact with different kinds of energy. For example,
scientists know that vegetation absorbs a lot of red light,
and reflects a lot of infrared light. So if MODIS passed over
a dense rainforest, the detectors would likely report low
radiance values for red light, and high radiance values for
infrared light, and biologists could be confident that the
area was vegetated.
But what if the area over the rainforest was cloudy? Then
the radiance measured by MODIS wouldnt be the true radiance
of the forest. If a biologist based a calculation of the amount
of forest cover present on the cloud-contaminated radiance
value, the estimate could have serious error. Similar cloud-contamination
problems arise for other MODIS products, such as snow and
sea ice detection, and sea and land surface temperature.
The MODIS Cloud Mask aims to minimize the potential errors
resulting from cloud contamination by labeling every pixel
of data as either confident clear, probably clear, uncertain,
or confidently cloudy. In general, when MODIS receives energy
coming from clouds, that energy is both brighter and colder
than it would be if MODIS had a clear view of the Earth. So
initial cloud detection begins with these electromagnetic
However, there are many situations when the colder-brighter
rule doesnt hold up: clouds occurring over snow and
ice, thin cirrus clouds, and nighttime stratus clouds that
are very close to the surface, just to name a few. So Menzels
team takes advantage of the fact that MODIS is designed to
measure a large number of spectral bands, or ranges of wavelengths
of radiation. Those wavelengths interact with water or ice
droplets in clouds in different ways, giving the scientists
wide variety of tests they can use to look for cloud contamination.
Menzels team uses 17 of MODIS 36 spectral bands
for their tests, which they translate into a computer program
that screens the MODIS data stream, testing every pixel of
data for cloud contamination. For each of the spectral tests,
the program assigns a probability between 1 and 0, where 1
indicates a near 100% probability that the pixel is clear,
and 0 indicates almost zero probability that it is clear.
Depending on how the pixel performed on the various tests,
the program then combines the results of all the tests to
classify the pixel as either confident clear, probably clear,
uncertain, or cloudy. To be most useful of course, most pixels
would fall into the first and last categories, and the teams
initial results indicate that on an average day 90% of pixels
are labeled as confidently clear, or confidently cloudy.
The program isnt perfect; it still makes occasional
mistakes over snow and ice and sometimes the desert, whose
bare ground can look as bright as a cloud. But as MODIS continues
to collect data, the Cloud Mask tests will become more and
more refined, yielding increasingly accurate and precise cloud
detection, ultimately improving the quality of all MODIS products
that depend on it.