IceDetectNet: Classifying Individual Ice Components in Aggregates

Previous field studies have shown that aggregated ice crystals often make up a large proportion (38%-54%) of the total cloud ice, especially in precipitating clouds. However, current classification algorithms face several challenges in precisely classifying aggregated ice crystals: (1) they assign a single label to each ice crystal, which is inadequate for cases where multiple labels are required (i.e., a rimed column); (2) they treat aggregated ice crystals as a single ice, regardless of the number of components involved; (3) the process of model training and fine-tuning requires extensive manual labeling, which is not only time-consuming but also limits the ability to use more diverse and larger datasets.

To overcome these limitations, this study presents a rotated object detection algorithm (IceDetectNet) that can classify ice crystals down to the aggregate component scale, providing detailed information on the number of components and the ice categories of the individual components. In addition, a novel contrastive semi-supervised learning (CSSL) algorithm is presented that reduces the need for manual labeling while maintaining high classification performance. Both algorithms have been trained and evaluated on holographic imager data from the NASCENT campaign in Ny-Ålesund, Svalbard and the CLOUDLAB campaign in central Switzerland. Based on the number and category of components from IceDetectNet, we quantified the number of collisions assuming a sticking efficiency of 1, calculated as the number of aggregated components minus one. This detailed component information allows the quantification of aggregation rates and collision kernels (the rate at which ice crystals collide). In addition, we quantified the aggregation fraction (the percentage of ice crystals that are aggregated) and the mean number of components per ice crystal, which allowed us to estimate the number of ice nucleation events (where each aggregated component represents one nucleation event).

IceDetectNet enables accurate classification of aggregated ice components, while CSSL enables faster, more adaptable training of ice crystal classification models over larger datasets. Together, these advances enhance our ability to quantify ice nucleation and aggregation processes, providing deeper insights into cloud microphysics and their influence on weather and climate.