The Classification of Ambiguous Ice Particle Shadowgraphs by Consensus

Abstract

A major impediment to the development of computer algorithms for the automatic classification of ice particle types found in the atmosphere as measured by a Particle Measuring System two-dimensional probe is the difficulty of obtaining training data. This is especially true when, as is usually the case, the particle shapes do not correspond to any of the pure crystal types found in textbooks. This paper presents the results of testing such a training set. Sources of bias among human observers include the effect of training and previous familiarity with the data, fatigue, and particle orientation, as well as subjective differences among observers. The deviation of individual human observers from the classifications arrived at by consensus indicates an upper bound to the accuracy possible in automated classification schemes.