Answers are grounded in this paper and cite their sources.
Machine learning (ML) has become a promising tool for image analysis, particularly for datasets with instrument-specific characteristics and variable image quality. The balloon-borne ice cloud particle imager (B-ICI) collects and images ice particles on a film strip during ascent through the troposphere. Although the instrument provides high optical and pixel resolution, the background features of the film and oil coating introduce noise that complicates image analysis. To process these raw data obtained from B-ICI, a set of tools, BICI-NET, has been created. It consists of two convolutional neural network (CNN) models: BICI-NET segmentation, which identifies ice particles in the images, and BICI-NET classification, which categorizes the particles by shape. A visual inspection step is included to ensure the quality of the segmentation results, allowing users to accept, correct, or reject individual mask images. These masks are then used to extract particle properties such as maximum dimension, cross-sectional area, and aspect ratio. The segmentation model was initially trained on data from 2013 to 2018, and its performance improves iteratively as new, inspected data are added to the training set. Using this model, the duration of the segmentation process is reduced by approximately 90%. Particles are classified into four different categories: columns, compacts, plates, and rosettes. These shape categories have been chosen based on current needs and because they had been used in previous studies with B-ICI data. We present the BICI-NET processing pipeline, its validation, and performance results. Comparisons with a manually analyzed test dataset show good agreement, especially for particles larger than ≳100 μ m. Minor biases affect smaller particles, and most misclassifications occur in the plates category, which is underrepresented in the training data. Significance Statement Processing data from cloud imaging instruments presents unique challenges. Depositional instruments, such as the B-ICI, are particularly difficult to automate in terms of image analysis due to the presence of artifacts and nonuniform backgrounds. The model presented in this paper is specifically designed for the B-ICI—and potentially for other deposition-based cloud particle imagers (CPIs), with a primary focus on image segmentation, i.e., identifying ice particles among various artifacts. Although this work does not propose a general model applicable across different instruments, it may represent a step toward developing more broadly applicable solutions.
No takes yet. Share an insight, caveat, or question.
Stenszky et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: