All Time Past Year Past 30 Days
Abstract Views 1 0 0
Full Text Views 220 56 6
PDF Downloads 77 29 1

A Pattern Recognition Technique for Distinguishing Surface and Cloud Types in the Polar Regions

View More View Less
  • 1 Department of Meteorology, University of Wisconsin, Madison, WI 53706
Full access

Abstract

Measurement of polar cloud cover is important because of its strong radiative influence on the energy balance of the snow and ice surface. Conventional satellite cloud detection schemes often fail in the polar regions because the visible and thermal contrasts between cloud and surface are typically small. Nevertheless, experts looking at satellite imagery can distinguish clouds from the surface by examining the textural characteristics of the scene.

This paper describes an automated pattern recognition algorithm winch identities regions of various surface and cloud types at high latitudes from visible, near-infrared, and infrared AVHRR satellite data. Five spectral features give information about the magnitude of albedos and brightness temperatures, while three textural features describe the variability and “bumpiness” in a scene. The maximum likelihood decision rule is used to classify that region into one of seven surface categories or 11 cloud categories.

The algorithm was able to classify 870 training samples with a skill of 84%. Eighteen hundred artificer samples created using a Monte Carlo technique were classified with a skill of 92%, which represents the theoretical limit of class separability using the given features. Both the near-infrared information and the textural information proved to be especially useful in detecting high-latitude cloudiness. The algorithm experienced some difficulty identifying thin stratus over snow and ice and thin cirrus over land and water, situations which also prove difficult for most other cloud detection schemes.

When tested on AVHRR imagery from a different date, the algorithm showed a skill of 83% as verified against the analyses of three independent experts. Significant variability was encountered among the experts, underlining the need for an objective routine. This algorithm performed more accurately thin others constructed with alternate feature sets corresponding to various existing cloud detection schemes.

Abstract

Measurement of polar cloud cover is important because of its strong radiative influence on the energy balance of the snow and ice surface. Conventional satellite cloud detection schemes often fail in the polar regions because the visible and thermal contrasts between cloud and surface are typically small. Nevertheless, experts looking at satellite imagery can distinguish clouds from the surface by examining the textural characteristics of the scene.

This paper describes an automated pattern recognition algorithm winch identities regions of various surface and cloud types at high latitudes from visible, near-infrared, and infrared AVHRR satellite data. Five spectral features give information about the magnitude of albedos and brightness temperatures, while three textural features describe the variability and “bumpiness” in a scene. The maximum likelihood decision rule is used to classify that region into one of seven surface categories or 11 cloud categories.

The algorithm was able to classify 870 training samples with a skill of 84%. Eighteen hundred artificer samples created using a Monte Carlo technique were classified with a skill of 92%, which represents the theoretical limit of class separability using the given features. Both the near-infrared information and the textural information proved to be especially useful in detecting high-latitude cloudiness. The algorithm experienced some difficulty identifying thin stratus over snow and ice and thin cirrus over land and water, situations which also prove difficult for most other cloud detection schemes.

When tested on AVHRR imagery from a different date, the algorithm showed a skill of 83% as verified against the analyses of three independent experts. Significant variability was encountered among the experts, underlining the need for an objective routine. This algorithm performed more accurately thin others constructed with alternate feature sets corresponding to various existing cloud detection schemes.

Save