Satellite Cloud Classification and Rain-Rate Estimation Using Multispectral Radiances and Measures of Spatial Texture

View More View Less
  • 1 National Institute of Water and Atmospheric Research, Wellington, New Zealand
© Get Permissions Rent on DeepDyve
Restricted access

Abstract

Twelve months of Southern Hemisphere (maritime) midlatitudes Advanced Very High Resolution Radiometer local area coverage data at full radiometric and spatial resolution have been collocated with rain-rate data from three Doppler weather radars.

Using an interactive computing environment, large independent samples of cloudy-altocumulus, cumulonimbus, cirrostratus, cumulus, nimbostratus, stratocumulus, stratus-and cloud-free scenes have been identified (labeled) in the collocated data. Accurate labeling was ensured by providing a supervising-analyst access to appropriate diagnostics, including difference and ratio channels, 3.7-µm reflected and emissive components, spectral histograms, Coakley-Bretherton spatial coherence plots, mean, standard deviation, and gray-level difference (GLD) statistics. This analysis yielded 4323 cloud and no-cloud samples at a spatial resolution of 8 × 8 instantaneous fields of view (IFOV), from 257 NOAA-11 and NOAA-12 orbits.

Bayesian cloud discriminant functions calculated from the labeled samples and utilizing feature vectors including radiometric and GLD spatial characteristics successfully classified scenes into one of the seven cloud and no-cloud classes with significant skill (Kuipers’ performance index 0.63). Utilizing the posterior probability of the classified samples enabled some clouds that were classified erroneously to be identified (and discarded), improving the skill of the discriminant functions by an additional 10% or so. Removing the GLD statistics from the feature vector reduced the skill of the cloud discrimination by about 20% (relative to the nondiscarding discriminant function), while increasing the misclassification of midlevel clouds. However, some cloud classes can only be discriminated from their multispectral signatures. Day and night discriminant functions show similar skill.

Within raining cloud classes, rain rate has been related to the spatial and radiometric characteristics of the cloud. The skill of the rain-rate estimates is dependent on the cloud type. For nimbostratus and altocumulus classes 20%–25% of the rain-rate variation can be explained by predictors that measure the temperature, spatial texture, and degree of isotropy in the sampled clouds. Raining and nonraining samples of altocumulus, cumulus, cirrostratus, and nimbostratus can be delineated with at least 60% accuracy.

This approach, whereby cloud classes are identified then rain rates estimated as a function of cloud type, would seem to resolve some of the usual problems associated with rain-rate analyses from midlatitudes infrared and visible satellite data. It also extends rain-rate diagnosis to nonconvective (frontal) cloud systems.

Abstract

Twelve months of Southern Hemisphere (maritime) midlatitudes Advanced Very High Resolution Radiometer local area coverage data at full radiometric and spatial resolution have been collocated with rain-rate data from three Doppler weather radars.

Using an interactive computing environment, large independent samples of cloudy-altocumulus, cumulonimbus, cirrostratus, cumulus, nimbostratus, stratocumulus, stratus-and cloud-free scenes have been identified (labeled) in the collocated data. Accurate labeling was ensured by providing a supervising-analyst access to appropriate diagnostics, including difference and ratio channels, 3.7-µm reflected and emissive components, spectral histograms, Coakley-Bretherton spatial coherence plots, mean, standard deviation, and gray-level difference (GLD) statistics. This analysis yielded 4323 cloud and no-cloud samples at a spatial resolution of 8 × 8 instantaneous fields of view (IFOV), from 257 NOAA-11 and NOAA-12 orbits.

Bayesian cloud discriminant functions calculated from the labeled samples and utilizing feature vectors including radiometric and GLD spatial characteristics successfully classified scenes into one of the seven cloud and no-cloud classes with significant skill (Kuipers’ performance index 0.63). Utilizing the posterior probability of the classified samples enabled some clouds that were classified erroneously to be identified (and discarded), improving the skill of the discriminant functions by an additional 10% or so. Removing the GLD statistics from the feature vector reduced the skill of the cloud discrimination by about 20% (relative to the nondiscarding discriminant function), while increasing the misclassification of midlevel clouds. However, some cloud classes can only be discriminated from their multispectral signatures. Day and night discriminant functions show similar skill.

Within raining cloud classes, rain rate has been related to the spatial and radiometric characteristics of the cloud. The skill of the rain-rate estimates is dependent on the cloud type. For nimbostratus and altocumulus classes 20%–25% of the rain-rate variation can be explained by predictors that measure the temperature, spatial texture, and degree of isotropy in the sampled clouds. Raining and nonraining samples of altocumulus, cumulus, cirrostratus, and nimbostratus can be delineated with at least 60% accuracy.

This approach, whereby cloud classes are identified then rain rates estimated as a function of cloud type, would seem to resolve some of the usual problems associated with rain-rate analyses from midlatitudes infrared and visible satellite data. It also extends rain-rate diagnosis to nonconvective (frontal) cloud systems.

Save