Abstract
A data-driven approach to the classification of hydrometeors from measurements collected with polarimetric weather radars is proposed. In a first step, the optimal number of hydrometeor classes ($n_{opt}$) that can be reliably identified from a large set of polarimetric data is determined. This is done by means of an unsupervised clustering technique guided by criteria related both to data similarity and to spatial smoothness of the classified images. In a second step, the $n_{opt}$ clusters are assigned to the appropriate hydrometeor class by means of human interpretation and comparisons with the output of other classification techniques. The main innovation in the proposed method is the unsupervised part: the hydrometeor classes are not defined a priori, but they are learned from data. The approach is applied to data collected by an X-band polarimetric weather radar during two field campaigns. Seven hydrometeor classes ($n_{opt} = 7$) have been found in the data set, and they have been identified as light rain (LR), rain (RN), heavy rain (HR), melting snow (MS), ice crystals/small aggregates (CR), aggregates (AG), and rimed-ice particles (RI).
Introduction: The Challenge of Hydrometeor Classification
Hydrometeor classification (HC) is essential for identifying precipitation types—such as rain, snow, or hail—within a radar's sampling volume. Traditionally, this process relies on "fuzzy logic" classifiers built upon theoretical scattering simulations. However, these simulations often struggle to accurately model the complex geometries of ice-phase particles like snowflakes. Furthermore, they frequently fail to account for the unique "fingerprints" or calibration nuances of specific radar instruments.
To address these limitations, this research introduces a more objective, instrument-tailored classification system. By moving away from a priori assumptions, the proposed method "learns" the most distinct particle categories directly from the radar data while ensuring that the resulting weather maps remain spatially consistent.
Methods: An Unsupervised Clustering Framework
The researchers developed a two-step unsupervised framework designed to prioritize both data similarity and physical realism.
1. Clustering with Spatial Constraints
The core of the method involves Agglomerative Hierarchical Clustering (AHC) applied to a representative subset of radar observations. The key innovation is the "merging rule." Rather than relying solely on mathematical similarity (e.g., Euclidean distance), the algorithm incorporates spatial smoothness. It operates on the physical assumption that hydrometeor types do not change erratically between adjacent pixels. If a cluster appears spatially "noisy" or fragmented, the algorithm forces it to merge with its most similar neighbor, resulting in more coherent classification maps.
2. Optimal Partition Selection
Determining the correct number of categories is a critical step. The authors analyzed several quality indices, including the SD index, Cohen’s kappa, and Accuracy Spread (AS).
As shown in the analysis of these indices, seven clusters ($n_{opt} = 7$) provided the optimal balance between data similarity and spatial coherence. Once these seven clusters were mathematically defined, human experts labeled them based on physical properties and comparisons with existing models.
Datasets and Validation
The approach was tested using data from an X-band dual-polarization Doppler radar (MXPol) deployed during two distinct field campaigns: one in the Swiss Alps (Davos) and another in the French Mediterranean (Ardèche). The dataset comprised approximately 50 precipitation events.
To validate the radar's classifications, the team used a Two-Dimensional Video Disdrometer (2DVD). This ground-based instrument captures high-speed photographs of falling particles, providing a "ground truth" for comparison. The results were also benchmarked against the DR2009 algorithm, a standard fuzzy-logic classifier.
Key Results and Insights
- Superior Performance: Validation against disdrometer data showed that the clustering method achieved an accuracy of 49% (Cohen’s kappa 0.23) for ice-phase particles. This significantly outperformed the traditional DR2009 model, which scored 38% (kappa 0.08).
- Physical Consistency: The model successfully identified the vertical "melting layer"—the bright band where snow transitions to rain—and correctly mapped the progression from light crystals at high altitudes to heavier, rimed particles at lower levels.
- Calibration Robustness: Because the method is data-driven, it is inherently more resilient to radar miscalibration than simulation-based methods, as it adapts to the specific sensitivities of the hardware used.
- Natural Classification: The data naturally settled into seven distinct categories: Light Rain (LR), Rain (RN), Heavy Rain (HR), Melting Snow (MS), Ice Crystals/Small Aggregates (CR), Aggregates (AG), and Rimed-Ice Particles (RI).
Limitations and Future Directions
While the clustering process is unsupervised, the final labeling stage still requires expert human interpretation, meaning the system is not yet fully "plug-and-play" for new climatic regions. Additionally, the classifier's performance is tied to its training database; rare events not captured in the initial 50 samples may not be recognized.
Future work will explore incorporating temporal consistency—ensuring classifications do not fluctuate rapidly over time—and developing more sophisticated physical contiguity rules to further refine spatial constraints.
Summary updated