Cloud Detection: Optical Depth Thresholds and FOV Considerations
Steven A. Ackerman, Richard A. Frey, Edwin Eloranta, and Robert Holz
Cooperative Institute for Meteorological Satellite Studies, Space Science and Engineering Center, University of Wisconsin-Madison,
Cloud Mask and Spectral Tests
Cloud Detection Issues
What is a cloud?
The answer to that question is determined by the application. What is considered a cloud in
some applications may be defined as clear in other applications. For example, detection of thin
cirrus clouds is important for applications of infrared remote sensing of sea surface
temperature, but of little concern for microwave remote sounding of atmospheric temperature.
So, in one regard, cloud detection capabilities is determined by the application.
Detection of clouds is also a function of instrument capability and algorithm design. Cloud
detection is a function of contrast between the target (e.g. cloud) and the background. Contrast
can be:
•Spatial:
•Temporal:
•Spectral:
surface.
Large fov are generally more uniform lowering contrast
Clouds can be detected in a sequence of images if the clouds are moving
Spectral contrast is determined by the radiative properties of the cloud and
GCMs make extensive comparison with satellite derived cloud amount. Total cloud
amount from different satellite algorithms can vary significantly even among accepted
standards, as shown below in a comparison of annual zonal mean cloud fraction from
snow
CLAVR, ISCCP and UW-HIRS. Global distributions demonstrate
expected patterns but
can differ in magnitude by more then 10%. This paper investigates how spectral testing,
field-of-view size and cloud optical depth impact cloud detection algorithms.
1.38um Cloud Model Reflectance
A note on thresholds
Lowering the Cirrus Cloud detection
threshold will capture much thinner cirrus
(optical depth of 0.08 in this example)
B26 Cloud test threshold
in V3.0.0 (current)
New B26 Cloud test
threshold under evaluation
Optical depth detectable
with new threshold
Cloud Optical Depth
Top: Thresholds accuracy is a function of instrument noise and spectral
contrast with background. In this example, the optical depth of the cloud
detection at 1.38 microns is estimated through simulations of an water scene.
Observations indicate that optical depth detection limits are more on the order
of 0.1 to 0.3.
Total
Cloud
Amount
The above two figures show the 0.86 reflectance (x-axis) versus the percentage of pixels less then that value (e.g. cloud
fraction if this reflectance was a threshold for ocean scenes equatorward of 60 degrees and away from sun glint for
various solar zenith angles and viewing angle s between 0- 10 degrees (left) and 30 –40 degrees (right). The geometrically
defined sun-glint areas (reflectance angle 0-36 degrees). For different viewing geometries, the cloud detection threshold
varies. At low reflectances (less then 10 %) a small change in the threshold can result in a large change in cloud amount.
Comparison of satellite cloud amounts should account for the varying viewing and solar geometry.
These histograms show observations of radiance data as a
function of final clear sky confidences according to the MODIS
cloud mask. The top two figures also define thresholds for the
tests depicted. For example, the top left plot shows how the
distribution of visible ratios changes with clear sky confidence.
The vertical lines define the threshold interval for this cloud test
(1.0 at left to 0.0 at right). One may conclude that the thresholds
have been chosen properly as very few, if any, clear sky
confidences >0.95 fall within the interval. In the figure above,
however, one sees that part of the distribution of observations
denoted as clear (blue) or probably clear (green) falls inside the
threshold interval. One could conclude that these thresholds
should be made smaller (moved left on the graph).
Top: Zonal mean frequencies of cloudy conditions for October 16,2003, daytime ocean scenes as a
function of three cloud detection tests and the combination of all 16 tests from MODIS.
Bottom: Zonal mean frequencies of cloudy conditions for October 16,2003, daytime land scenes as
a function of three cloud detection tests and the combination of all 16 tests from MODIS.
Field of View
Right: Accumulated zonal mean cloud occurrence as various tests are added to the detection
algorithm for daytime land (top) and daytime ocean (bottom) scenes.
Bottom: Cloud fraction increase due to addition cloud detected by the MODIS
1.38 micron channel.
The percent of total observations of clear (blue),
high cloud (green) and total cloud (red) as a
function of MODIS fov size. Observations are over
ocean between 60 S and 60N during the daytime
and away from sun glint.
To estimate cloud optical detection limits cloud mask results from the
MODIS and GLI were compared to ground based observations from the
High-Spectral Resolution Lidar (HSRL), which measures visible optical
depth. Comparisons were also made using the ER-2 borne cloud
physics lidar and collocated observations of the MODIS Airborne
Simulator (MAS)
Left: GLI (and MODIS) observations were compared to the HSRL site over the University of
Wisconsin-Madison. The HSRL directly measures cloud optical depth at visible wavelengths.Initial
results indicate that when the MODIS or GLI flag a cloudy region as Uncertain Clear, the optical depth
is less then approximately 0.3.
snow
Above: The total cloud fraction as a function of cloud optical depth,
and the cloud fractions when considering the plus and minus sigma
values of optical depth from June 2004. Each optical depth time
profile has an associated error bar due to the molecular return and
the density profile.
Top: The number of occurrences that MAS scene was identified as clear and
the cloud physics lidar detected a cloud optical depths (visible wavelengths).
This figure suggests that the cloud limit is less then approximately 0.4
Bottom: The number of occurrences of lidar determined optical depth when the
MAS cloud mask algorithm detected a cloud.
Right: A temporal resolution comparison of cloud fraction for the
month of June 2004, for a 300-meter spatial average and a 1, 3, and
5-minute temporal average of the HSRL. When the time average is
reduced to a 1-minute average, the lidar signal tends to “see” the
same amount of optically thin cloud and a slight increase in the
percentage of thicker clouds (optical depths greater than 1.0).
Increasing to a 5-minute average, the amount of detectable thin
clouds remains the same, the cloud fraction for thicker clouds is
reduced. This is due to averaging the edges of clouds with
surrounding clear sky – effectively reducing the measured optical
depth.
Far right: The impact of vertical average on the derived cloud
fraction from the HSRL.
MODIS Cloud
MPL/MMCR Cloud
MODIS No Cloud
MPL/MMCR Cloud
24 cases
9 cases
MODIS Cloud
MPL/MMCR No Cloud
MODIS No Cloud
MPL/MMCR No Cloud
12 cases
8 cases
Lidar/radar/cloud mask validation results are shown in the above table. MODIS cloud
mask output (5x5-pixel regions) and combined SGP CART site ground-based MPL
(Micro-pulse Lidar) and MMCR (Millimeter-wavelength Cloud Radar) cloud data were
collocated for the month of November, 2000. MODIS 5x5s were considered to be cloudy if
at least half of the individual 1-km pixels had clear sky confidences < 0.95. The
distribution of the four possible cloud/no cloud results between the cloud mask and the
MPL/MMCR is indicated by the numbers in the table. There are inherent difficulties in
comparing data with vastly different spatial and temporal resolutions and sensitivities.
Discussion
Cloud fraction has been derived from various satellite platforms for over 25
years. New satellites have more capability, in terms of improved performance,
more spectral channels and higher spatial resolution. Fractional cloud
coverage is a function of the optical depth threshold limit of the instrument, and
will also vary with the spatial, spectral and temporal resolution of the
instrument. In some scenes, a single (or few) channel method yields similar
results to mutlispectral tests. Cloud detection thresholds also vary as a
function of viewing geometry, scene illumination and thermal structure of the
scene. Comparison of various cloud detection methods/results from different
instruments or platforms need to account for these differences.
The
dependence of cloud detection on these parameters and the need to monitor
with changing instruments and satellites, will likely make it difficult to compare
cloud amounts from different approaches and achieve the 1% accuracy needed
for long-term monitoring of cloud amount.