January 13, 2009


Cloud detection

Over Ocean

Over Ocean, cloud detection is much easier than over land surfaces because the surface is a relatively uniform and dark background. This background generates a nice contrast with the clouds. As a consequence, a simple reflectance based on threshold on the longer wavelength channel measurements eliminates most of the clouds. On the other hand, it is desirable for aerosol objectives to keep aerosol-contaminated pixels. Aerosol, and in particular large dust events, may yield a significant reflectance, comparable to that generate by thin or broken clouds. It is then possible to distinguish clouds and aerosol with a simple threshold on the reflectance.
Analysis of satellite reflectance images show that the aerosol fields are much more spatially homogeneous than the cloud fields. This observation suggests the use of simple test based on the spatial variability of the reflectance. We make use of a simple threshold on the 3x3 pixels standard deviation. The cloud detection over ocean is based on the combination of a simple rough test on the top of the atmosphere reflectance, and another test on the spatial variability of the reflectance. We found that the later is the primary test (i.e. few clouds are detected only by the former).

Over land surfaces

Retrievals of tropospheric aerosol properties and vegetation cover over continental surfaces are only possible over cloud-free regions. The first step in satellite data processing thus requires a reliable cloud detection scheme based on the specific radiometric signature of clouds. Such a scheme is implemented in POLDER processing and takes advantage of POLDER multi-spectral and multi-directional specific capabilities. The scheme consists in a simple cloud filtering based on a threshold method.

Four tests are applied to each pixel at POLDER Level 2 (full resolution grid). These tests are non-exclusive and one proved positive is sufficient to declare the pixel cloudy. The first one compares the apparent pressure in oxygen band to the surface pressure. The second one is a threshold on the reflectance in the blue channel. The third consists in detection of the presence of polarized rainbow, indicative of the presence of water droplets. Once all these tests applied, some of the pixels declared cloudy by the blue channel test are considered again in order to prevent from abusive elimination of actually clear pixels. It is highly probable to be the case above snow covers. This kind of difficulty is also encountered with the apparent pressure test above dense forests. Finally, pixels adjacent to a detected cloud are considered to be contaminated and are rejected for further processing.

Performances of this cloud detection scheme have been evaluated against ground-based sky observations. For more details about this validation or about the detection method itself, the reader is referred to Bré,on and Colzy, 1999. Hereafter follows a brief presentation of each one of these cloud detection tests implemented in Polder Level 2 data processing for Land Surfaces thematic.

Apparent pressure in O2 band


This test involves 2 POLDER channels of different wide (a narrow one of 10 nm and a broader one of 40 nm) centered on the oxygen "A" absorption band at 765 nm. The ratio of the reflectances measured in each of theses 2 channels is a good indicator of atmospheric oxygen absorption (Bréon and Bouffiès, 1996). From this ratio, an apparent pressure of the reflector can be derived that is a physical quantity directly comparable to surface pressure. This comparison allows concluding about the presence of a cloud layer since these scattering layers (in particular high clouds) have a large effect on the apparent pressure (Bréon and Colzy, 1999). The pixel is declared cloudy if the apparent pressure is markedly lower than surface pressure.
Psurf-Papp>P, where Papp is result of average over all available viewing angles such as to reduce random uncertainties.

An NDVI dependant threshold

A clear surface shows a spectral signature that affects the apparent pressure (lower apparent pressure over the vegetation). As a consequence, the threshold depends on the vegetation coverage, quantified by the NDVI. Bréon and Bouffiès, 1996 highlighted that the difference Psurf-Papp is NDVI dependant. An empirical representation of Psuf-Papp=f(NDVI) for clear sky cases enabled to advance a linear relation:
P = a*NDVI + b
In POLDER Class 1, a=60 and b=140. For Class 2, a finest matching is applied with b=120.

Use of meteorological surface pressure

In POLDER first algorithms, the surface pressure considered the simplified hypothesis of a uniform standard atmosphere, accounting only for the surface altitude. Advanced algorithm makes use of a more realistic surface pressure derived from meteorological data (ECMWF).

Blue channel reflectance


While ground reflectance increases with wavelength in visible and near-infrared, cloud reflectance is roughly uniform within these windows. Thus the contrast between ground and cloud reflectances is the largest in the blue channel (443 nm) that is the shorter wavelength POLDER channel. However, this test cannot be applied to raw R443 measurements as long as the molecular diffusion dominates the signal and screens the ground signature. Also, the observed pixel is declared cloudy by this test if 443 nm reflectance corrected from molecular diffusion is above a given threshold, 443:
R443-R443mol > 443

A reference base of ground R443min

Whereas in first algorithm, the blue channel test was rather simple with only a coarse distinction between dense vegetation covers (443=0.10) and deserts (443=0.15), advanced algorithm introduces a finer tuning of the threshold taking into account the spatial variability of surface reflectance. A base of surface reflectances R443min has been constructed using the hemispheric reflectances calculated at Polder Level 3 over the 8 months of POLDER-1 data. The pixel is declared cloudy when the observed blue reflectance, corrected for molecular scattering, is larger than a spatially varying threshold:
R443-R443mol > R443min+’443 where ’443 is fixed to 0.05, supposing that reflectance increase due to vegetation growth during annual cycle do not exceed this rate.

Polarization rainbow

In presence of liquid water clouds, the reflected radiance shows a maximum in the 142° scattering direction and this maximum is highly polarized. This property is exploited through multi-directional and polarized POLDER data. The test on the presence of a polarized rainbow is realized in the 865 nm channel, the less contaminated by atmosphere (molecules and aerosols) single scattering. If R865pol within 142° is significantly larger than R865pol away from this direction, the pixel is declared cloudy.

Snow covers

Since snow has a large radiance in the 443 nm wavelength and shows a spectral signature very similar to that of the clouds (in the spectral range covered by POLDER), the blue channel test fails, declaring snow pixels as cloudy. Although is appears not possible to unambiguously distinguish thin clouds from partial snow, it is useful to distinguish at least thick clouds from uniform snow covered region.

In configuration of:

    negative Papp test;
    negative polarized rainbow test;
    positive R443 test;
    large reflectance at 670 nm;
    pixel identified snow or ice by meteorological data;

    The pixel is put back to clear.

    Dense forests

    Tropical forests show a specific spectral signature. The surface reflectance, averaged over the narrow "763 nm" channel is significantly larger than that on the wide "765 nm" channel, when the apparent surface pressure algorithm assumes a spectrally neutral surface. This signature yields false cloud detection. To correct this problem, the pixels that are declared cloudy by the Apparent Pressure test are reconsidered:

    In configuration of:

      positive Papp test;
      negative polarized rainbow test;
      negative R443 test;
      large NDVI;

      The pixel is put back to clear.

      © CNES Copyrights 2002