| Summary: | Ice crystals play a crucial role in cloud optical properties and precipitation formation, as their shape affects their radiative properties, diffusional growth rate, fall speed, and collision efficiency. While the habit of ice crystals is determined by the ambient environment (i.e., temperature and humidity) in which they grow, it is also shaped by microphysical processes such as riming and aggregation. However, existing single-label classification algorithms face limitations when it comes to assigning multiple labels to a single ice crystal (i.e., a rimed column) or classifying the various components of an aggregated ice crystal.To overcome these limitations, this study introduces a novel multi-label classification system that considers both basic habits and physical processes leading to the observed ice crystal shapes. An object detection algorithm is presented that classifies each component of an aggregated ice crystal individually (including both basic habits and physical processes). The algorithm was trained on 18’000 ice crystal images and tested on 2300 ice crystal images captured by a holographic imager during the NASCENT campaign in Ny-Alesund, Svalbard. The algorithm offers a classification of both basic habits and microphysical processes with an accuracy of 86.5% and 81.4% respectively. The study results provide a deeper understanding of ice crystal shapes, which can improve precipitation and radiation estimates, and further contribute to advances in weather forecasting and climate research. Additionally, the algorithm's performance has shown a better generalization ability to predict ice crystal habits in new datasets compared to traditional deep learning algorithms.
|
|---|