Abstract
The first hydrometeor classification technique based on two-dimensional video disdrometer (2DVD) data is presented. The method provides an estimate of the dominant hydrometeor type falling over time intervals of 60 s during precipitation, using the statistical behavior of a set of particle descriptors as input, calculated for each particle image. The employed supervised algorithm is a support vector machine (SVM), trained over 60 s precipitation time steps labeled by visual inspection. In this way, eight dominant hydrometeor classes can be discriminated. The algorithm achieved high classification performances, with median overall accuracies (Cohen's K) of 90% (0.88), and with accuracies higher than 84% for each hydrometeor class.
Introduction
Accurate identification of hydrometeor types—ranging from liquid rain to complex solid forms like graupel, aggregates, and dendrites—is critical for improving meteorological forecasts and validating polarimetric radar data. While remote sensing provides broad coverage, its classifications are often indirect. Ground-based instruments like the 2D Video Disdrometer (2DVD) offer high-resolution imagery and velocity data, serving as a potential "gold standard" for ground truth. However, the lack of automated classification methods for the high-volume data streams produced by 2DVDs has historically limited their utility in long-term precipitation research. This study introduces a supervised machine learning approach to bridge this gap.
Methodology
The proposed classification pipeline utilizes data from the 2DVD, an instrument that employs two orthogonal high-speed cameras to capture dual-view images of falling particles. This configuration allows for precise measurement of particle shape, size, and fall velocity.
Feature Extraction and Classification
For each detected particle, 13 descriptors are calculated, including equivolumetric diameter, fall velocity, roundness, and fractal dimension. To reduce noise and computational overhead, these descriptors are aggregated over 60-second intervals. Statistical moments (mean, median, quantiles) of these descriptors form a 115-dimensional feature vector for each interval.
A Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel was selected as the primary classifier. To optimize the model and identify key physical drivers, Multiple Kernel Learning (MKL) was used to rank feature importance. The model was trained on 400 expert-labeled intervals from diverse field campaigns in Davos (Switzerland), Remoray (France), and Ontario (Canada).
Results and Discussion
The SVM-based system demonstrated exceptional performance, achieving a median overall accuracy of 90% and a Cohen’s Kappa of 0.88. The model successfully discriminates between eight distinct hydrometeor categories.
Key Performance Drivers
Analysis via MKL revealed that three descriptors—pixel fraction (compactness), fall velocity, and particle diameter—accounted for approximately 70% of the model's decision weight. This highlights the necessity of combining geometric and physical parameters for robust classification. Furthermore, the model's performance was found to be sensitive to particle count, with accuracy dropping significantly in intervals containing fewer than 20 particles.
Physical Consistency
While air temperature was not an input feature, the resulting classifications showed strong correlation with expected temperature profiles. For instance, "melting snow-like" particles were correctly identified at the 0°C transition layer, validating the physical consistency of the machine learning approach.
Conclusion
This research provides the first automated framework for bulk hydrometeor classification using 2DVD data. By identifying the dominant precipitation "flavor" every minute, the system offers a reliable tool for long-term meteorological studies. Future work will aim to transition from bulk interval classification to particle-by-particle analysis, enabling a more granular understanding of mixed-phase precipitation events.
Summary updated