Two-Dimensional Hydrometeor Image Classification by Statistical Pattern Recognition Algorithms

View More View Less
  • 1 University of Wyoming, Laramie, WY 82071
© Get Permissions Rent on DeepDyve
Restricted access

Abstract

The investigation reported here involves the automatic classification of binary (black and white) images of hydrometeors (ice particles and raindrops) taken from cloud samples. The goal is to classify such images (both complete and fractional) into the seven most common classes of hydrometeors by statistical pattern recognition techniques. Detailed investigation about the data acquisition system and preprocessing is made. Four moment invariants which yield good class separation were used as features for the classification process. A Bayes decision function which minimizes the probability of misclassification is used for classification.

Bayes theorem is employed to update mean vectors and covariance matrices involved in the decision function. A discrete Kalman filtering algorithm is developed for the on-line estimation of the probability of occurrence of each class. For such estimation a discrete adaptive Kalman filtering algorithm is also developed which adjusts the filter gain matrix such as to whiten the innovations sequence. These techniques were shown to work well but the adaptive algorithm was found to converge to the correct probability more rapidly.

The classification algorithm was modified to classify incomplete or fractional images and two metrics were successfully developed to detect the unclassifiable images. The adaptive Kalman filter with the Bayes decision function was employed to classify about 2000 images per minute of CPU time with about 10% error.

Abstract

The investigation reported here involves the automatic classification of binary (black and white) images of hydrometeors (ice particles and raindrops) taken from cloud samples. The goal is to classify such images (both complete and fractional) into the seven most common classes of hydrometeors by statistical pattern recognition techniques. Detailed investigation about the data acquisition system and preprocessing is made. Four moment invariants which yield good class separation were used as features for the classification process. A Bayes decision function which minimizes the probability of misclassification is used for classification.

Bayes theorem is employed to update mean vectors and covariance matrices involved in the decision function. A discrete Kalman filtering algorithm is developed for the on-line estimation of the probability of occurrence of each class. For such estimation a discrete adaptive Kalman filtering algorithm is also developed which adjusts the filter gain matrix such as to whiten the innovations sequence. These techniques were shown to work well but the adaptive algorithm was found to converge to the correct probability more rapidly.

The classification algorithm was modified to classify incomplete or fractional images and two metrics were successfully developed to detect the unclassifiable images. The adaptive Kalman filter with the Bayes decision function was employed to classify about 2000 images per minute of CPU time with about 10% error.

Save