Polar Cloud and Surface Classification Using AVHRR Imagery: An Intercomparison of Methods

R. M. Welch Naval Oceanographic and Atmospheric Research Laboratory, Monterey, California

Search for other papers by R. M. Welch in
Current site
Google Scholar
PubMed
Close
,
S. K. Sengupta Naval Oceanographic and Atmospheric Research Laboratory, Monterey, California

Search for other papers by S. K. Sengupta in
Current site
Google Scholar
PubMed
Close
,
A. K. Goroch Naval Oceanographic and Atmospheric Research Laboratory, Monterey, California

Search for other papers by A. K. Goroch in
Current site
Google Scholar
PubMed
Close
,
P. Rabindra Department of Mathematics and Computer Science, South Dakota School of Mines and Technology, Rapid City, South Dakota

Search for other papers by P. Rabindra in
Current site
Google Scholar
PubMed
Close
,
N. Rangaraj Department of Mathematics and Computer Science, South Dakota School of Mines and Technology, Rapid City, South Dakota

Search for other papers by N. Rangaraj in
Current site
Google Scholar
PubMed
Close
, and
M. S. Navar Department of Mathematics and Computer Science, South Dakota School of Mines and Technology, Rapid City, South Dakota

Search for other papers by M. S. Navar in
Current site
Google Scholar
PubMed
Close
Restricted access

Abstract

Six Advanced Very High-Resolution Radiometer local area coverage (AVHPR LAC) arctic scenes are classified into ten classes. These include water, solid sea ice, broken sea ice, snow-covered mountains, snow-free land, and five cloud types. Three different classifiers are examined: 1) the traditional stepwise discriminant analysis (SDA) method; 2) the feed-forward back-propagation (FFBP) neural network; and 3) the probabilistic neural network (PNN).

More than 200 spectral and textural measures are computed. These are reduced to 20 features using sequential forward selection. Theoretical accuracy of the classifiers is determined using the bootstrap approach. Overall accuracy is 85.6%, 87.6%, and 87.0% for the SDA, FFBP, and PNN classifiers, respectively, with standard deviations of approximately 1%. Thin cloud/fog over ice is the class with the lowest accuracy (≈75%) for all of the classifiers. The snow-covered mountains, the cirrus over ice, and the land classes are classified with the highest accuracy (⩾90%) by all of the classifiers.

Abstract

Six Advanced Very High-Resolution Radiometer local area coverage (AVHPR LAC) arctic scenes are classified into ten classes. These include water, solid sea ice, broken sea ice, snow-covered mountains, snow-free land, and five cloud types. Three different classifiers are examined: 1) the traditional stepwise discriminant analysis (SDA) method; 2) the feed-forward back-propagation (FFBP) neural network; and 3) the probabilistic neural network (PNN).

More than 200 spectral and textural measures are computed. These are reduced to 20 features using sequential forward selection. Theoretical accuracy of the classifiers is determined using the bootstrap approach. Overall accuracy is 85.6%, 87.6%, and 87.0% for the SDA, FFBP, and PNN classifiers, respectively, with standard deviations of approximately 1%. Thin cloud/fog over ice is the class with the lowest accuracy (≈75%) for all of the classifiers. The snow-covered mountains, the cirrus over ice, and the land classes are classified with the highest accuracy (⩾90%) by all of the classifiers.

Save