Walter Scott, Jr. College of Engineering

Graduate Exam Abstract

Abstract: Recent developments in machine learning are applied to in-situ data collected by a Multi-Angle Snowflake Camera (MASC), incorporating convolutional and residual networks in big data environments. These networks provide the following benefits: require little initial preparation and automatic feature extraction, high accuracy and through transfer learning techniques, and relatively small training sets. The networks have large supporting communities and are popular for image processing and classification tasks specifically. In this paper, a convolutional neural network (CNN) is adapted and tasked with classifying images captured from two storm events in December 2014 and February 2015 in Greeley, Colorado. A training data set containing 1400 MASC images was developed by visual inspection of recognizable snowflake geometries and sorted into six distinct classes. The network trained on this data set achieved a mean accuracy of 93.4% and displayed excellent generality. A separate training data set was developed sorting flakes into three classes showcasing distinct degrees of riming. The network was then tasked with classifying images and estimating where flakes fell within this riming scale. The riming degree estimator yields promising initial results but would benefit from larger training sets. Future applications are discussed.
Adviser: Branislav Notaros
Co-Adviser: N/A
Non-ECE Member: Christine Chiu (Atmospheric Science)
Member 3: Ali Pezeshki (Electrical and Computer Engineering)
Addional Members: N/A
Program of Study: