Ice crystal images from optical array probes: classification with convolutional neural networks

Abstract

Although airborne optical array probes (OAPs) have existed for decades, our ability to maximize extraction of meaningful morphological information from the images produced by these probes has been limited by the lack of automatic, unbiased, and reliable classification tools. The present study describes a methodology for automatic ice crystal recognition using innovative machine learning. Convolutional neural networks (CNNs) have recently been perfected for computer vision and have been chosen as the method to achieve the best results together with the use of finely tuned dropout layers. For the purposes of this study, The CNN has been adapted for the Precipitation Imaging Probe (PIP) and the 2DS Stereo Probe (2DS), two commonly used probes that differ in pixel resolution and measurable maximum size range for hydrometeors. Six morphological crystal classes have been defined for the PIP and eight crystal classes and an artifact class for the 2DS. The PIP and 2DS classifications have five common classes. In total more than 8000 images from both instruments have been manually labeled, thus allowing for the initial training. For each probe the classification design tries to account for the three primary ice crystal growth processes: vapor deposition, riming, and aggregation. We included classes such as fragile aggregates and rimed aggregates with high intra-class shape variability that are commonly found in convective clouds. The trained network is finally tested through human random inspections of actual data to show its real performance in comparison to what humans can achieve.