Sub-pixel Cloud Research Articles

Application of neural networks to AVHRR cloud segmentation

The application of neural networks to cloud screening of AVHRR data over the ocean is investigated. Two approaches are considered, interactive cloud screening and automated cloud screening. In interactive cloud screening a neural network is trained on a set of data points which are interactively selected from the image to be screened. Because the data variability is limited within a single image, a very simple neural network topology is sufficient to generate an effective cloud screen. Consequently, network training is very quick and only a few training samples are required. In automated cloud screening, where a general network is designed to handle all images, the data variability can be significant and the resulting neural network topology is more complex. The latitudinal, seasonal and spatial dependence of cloud screening large AVHRR data sets is studied using an extensive data set spanning 7 years. A neural network and associated feature set are designed to cloud screen this data set. The sensitivity of the thermal infrared bands to high atmospheric water vapor concentration was found to limit the accuracy of cloud screening methods which rely solely on data from these channels. These limitations are removed when the visible channel data is used in combination with the thermal infrared data. A post processing algorithm is developed to improve the cloud screening results of the network in the presence of high atmospheric water vapor concentration. Post processing also is effective in identifying pixels contaminated by subpixel clouds and/or amplifier hysteresis effects at cloud-ocean boundaries. The neural network, when combined with the post processing algorithm, produces accurate cloud screens for the large, regionally distributed AVHRR data set.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Automatic cloud detection applied to NOAA-11 /AVHRR imagery

The imagery from the AVHRR on board NOAA polar orbiting satellites allows a description of cloud cover, oceanic, and continental surfaces that is used by Météo-France for nowcasting activities and as input for numerical weather prediction models (NWP). A real-time processing scheme has been designed at the Centre de Météorologie Spatiale (CMS) in Lannion to extract cloud cover and surface parameters from NOAA-11 AVHRR imagery received at CMS. The key step of this scheme is cloud detection. It is based upon threshold tests applied to different combinations of channels. Its main originality is its complete automation by the computation of the 11 μm infrared threshold from a monthly sea surface temperature (SST) climatology over the oceans and from air temperature (near the surface) forecast by NWP over land. A special test has been implemented to detect cloud edges and subpixel clouds over continental surfaces during daytime. It is applied daily in deferred time only to compute normalized difference vegetation index (NDVI). This scheme has been used operationally since February 1990, and its quality has been checked. It has enabled the routine production of various products. A nighttime cloud classification is sent to all French forecasters; NDVI values are computed daily and used to map the vegetation cover; and SST and thermal fronts are derived operationally from nighttime imagery.

Derivative reflectance spectroscopy to estimate suspended sediment concentration

Remotely sensed data can be used to estimate successfully the concentration of sediment in water. Such estimation has relied on relationships between suspended sediment concentration (SSC) and radiation in one or two broad wavebands where it is assumed that the effects of environmental variability (irradiance, subpixel cloud, etc.) are either small, or can be considered as spectrally additive constants in all wavebands. Where these assumptions do not hold, an alternative and theoretically more robust relationship is proposed, t that between SSC and derivative radiation (change in radiation per unit wavelength). Measurements of spectral reflectance (R λ), derivative spectral reflectance (DR λ) and SSC were measured in the laboratory, where the effects of environmental variability were small and at sea where the effects of environmental variability were large. There was a strong correlation between SSC and DR λ both in the laboratory (max R 2 = − 0.98) and at sea (max R 2 = − 0.83) and DR λ was used to estimate SSC in the laboratory to an error of less than 8 % of the mean SSC. The correlation between R λ and SSC was weaker at sea (max R 2 = 0.46) than in the laboratory (max R 2 = 0.96). This was due to the presence of large and spectrally variable environmental effects. We recommend the use of derivative spectra for the estimation of SSC when continuous spectra are available.

Remote sensing of water vapor in the near IR from EOS/MODIS

The LOWTRAN-7 code was used to simulate remote sensing of water vapor over 20 different surface covers. The simulation was used to optimize the water vapor channel selection and to test the accuracy of the remote sensing method. The channel selection minimizes the uncertainty in the derived water vapor due to variations in the spectral dependence of the surface reflectance. The selection also minimizes the sensitivity of the selected channels to possible drift in the channel position. The use of additional MODIS channels reduces the errors due to the effect of haze, subpixel clouds and uncertainties in the temperature profile. Remote sensing of the variation of water vapor from day to day will be more accurate, because the surface reflectances vary slowly with time. The method was applied to Airborne Visible Infrared Imaging Spectrometer (AVIRIS) data. >

Cloud-screening for Africa using a geographically and seasonally variable infrared threshold

A spatially variable monthly, infrared cloud-threshold data base has been used to screen cloud-contaminated observations from radiances measured by the NOAA-9 AVHRR over Africa. Cloud-screening through a monthly average infrared threshold based on measured surface air temperature, which is geographically dependent, shows an improvement over using a seasonally and geographically independent thermal cloud threshold of 287 K. It is found that differences in cloud-screening for these two thresholds occur for cases of lower altitude clouds or subpixel clouds where the radiative temperature is higher than the 287 K infrared threshold, yet colder than the variable threshold developed by Stowe et al. (1988) for the Nimbus-7 global cloud climatology. The variable IR threshold is shown to be effective over persistently cloud-covered regions, such as the coastal region of the Gulf of Guinea, but may introduce some erroneous cloud identifications over mountains.

The effect of subpixel clouds on remote sensing

A method for estimating the cloud effect on remote sensing is described, and it is applied to cloudiness in several climatological conditions. The algorithm is based on the Haurwitz (1948) measurements of the cloud layer transmission of solar radiation for an overcast sky and on an empirical interpolation of data for broken cloudiness by Pochop et al. (1968). Radiances for a sunny area observed directly from space and through a cloud, and for a shady area observed from space and through a cloud are computed. Methods for detecting the cloud effect from satellite images are discussed. The relation between cloud reflectance and cloud size is studied. It is observed that the subpixel clouds affect the detected radiance and vegetation index, and the effect depends on the cloud types and the dependence of the cloud transmissivity on cloud fraction. Procedures for decreasing or eliminating cloud effect are examined.

The effect of subpixel clouds on remote sensing

The effect of subpixel clouds on remote sensing of the surface reflectivity and of the vegetation index was simulated for a thin layer of rectangular clouds. The simulation is for a 4×4 km 2 field of view. It was found that a 20% cloudiness can increase the apparent surface reflectance by 0.02 for a model of Stratocumulus and by 0.08 for a fixed cloud reflectance of R C=0.5. For the measured cloudiness probability distributions it was found that the cloud effect can be eliminated if the best remote sensing case out of about 4 independent cases is chosen. This corresponds to 8–16 observation days.