Development of a classification algorithm for ice crystal habit by using deep learning

• Main Author: Zhang, Huiying
• Format: Master Thesis
• Language: English
• Published: 2021
• Subjects: Norway; Ny-Ålesund; Ny Ålesund
• Online Access: http://hdl.handle.net/10852/89265; http://urn.nb.no/URN:NBN:no-91875

Ice crystals, as an important component of clouds, have a strong influence on cloud radiative properties and precipitation formation. Moreover, ice crystal habits are controlled by the environment (temperature and humidity) in which they grow in and as such, are excellent tracers of in-cloud conditions. Therefore, ice crystal habit classification is an excellent tool to better understand the microphysical processes in clouds and thus, cloud radiative properties and precipitation formation. Over the past 50 years, researchers have improved the ability of algorithms to automatically and efficiently classify ice crystal habits. The most recent attempts have utilized machine learning and more specifically, a Convolutional Neural Network (CNN), due to its ability to catch the main features that describe ice crystal habits and recognize patterns between images. However, the CNNs trained on standard ice crystal habit images are difficult to apply in reality, due to the complexity of ice crystals in nature, which are generally a combination of different habits, rimed, or aggregates, and the difference between training dataset and real-world dataset. Therefore, in this thesis, a CNN is trained using images of ideal and complex ice crystals recorded by the HoloBalloon instrument during the NASCENT campaign in Fall 2019, in Ny-Ålesund, Norway. The dataset includes 16,259 images that were hand-labeled into 9 ice crystal habit classes. The best performing classification model ensemble (including 10 members), BestIce, achieved an overall accuracy of 87.55% and a class-wise accuracy of 91.72%. The models performed best when classifying plates and lollipops and frozen droplets and small ice with per-class accuracies of around 99.5% and 98%, respectively. To validate BestIce in a real-world application, the model is used to predict the ice crystals observed on a different day. When the prediction probability of BestIce was 99% or higher, which made up approximately 40% of the entire dataset, the global accuracy of the prediction was approximately 80%. However, when all of the ice crystals were classified with BestIce, the global accuracy fell to 63.31%. Nevertheless, the ability of BestIce to predict approximately 40% of the new dataset with such a high accuracy shows that the method developed in this thesis can be used to effectively classify ice crystals in a real-world setting.