A Two-Dimensional Hydrometeor Machine Classifier Derived from Observed Data

View More View Less
  • 1 ADAPT Service Corporation, Reading MA 01867
  • | 2 Air Force Geophysics Laboratory, Hanscom AFB, MA 01731
© Get Permissions
Full access

Abstract

Classification algorithms have been developed to distinguish six categories of cloud ice particles. These algorithms have been incorporated in schema which, when applied to shadowgraph images produced by the Precision Measurement System laser scanning device, have demonstrated the capability of classifying with more consistency than human classifiers, and with almost no sensitivity to particle orientation.

The data used to derive the algorithms consisted of observations obtained on four separate aircraft flights. Two human classifiers, interacting with a preliminary machine classification, defined the correct answers for this training data set. The algorithms were then tested against arbitrarily selected segments from two additional flights. The ADAPT Service Corporations eigenvector, or empirical orthogonal function (EOF) technique, defined the features objectively, and the ADAPT independent eigenscreening algorithm development program related these features to the particle type.

Analysis of the performance suggests that considerable variation is to be expected, based on the set-to-set variation of the distribution of particle types between real data sets. The classification schema have been developed to allow the user to change key parameters in order to compensate for this variation.

It was concluded that the machine classification was superior to manual classification for the identification of large numbers of particles in terms of speed and consistency.

Abstract

Classification algorithms have been developed to distinguish six categories of cloud ice particles. These algorithms have been incorporated in schema which, when applied to shadowgraph images produced by the Precision Measurement System laser scanning device, have demonstrated the capability of classifying with more consistency than human classifiers, and with almost no sensitivity to particle orientation.

The data used to derive the algorithms consisted of observations obtained on four separate aircraft flights. Two human classifiers, interacting with a preliminary machine classification, defined the correct answers for this training data set. The algorithms were then tested against arbitrarily selected segments from two additional flights. The ADAPT Service Corporations eigenvector, or empirical orthogonal function (EOF) technique, defined the features objectively, and the ADAPT independent eigenscreening algorithm development program related these features to the particle type.

Analysis of the performance suggests that considerable variation is to be expected, based on the set-to-set variation of the distribution of particle types between real data sets. The classification schema have been developed to allow the user to change key parameters in order to compensate for this variation.

It was concluded that the machine classification was superior to manual classification for the identification of large numbers of particles in terms of speed and consistency.

Save