Abstract: The phase state of cloud particles refers to the physical state or form of clouds, typically encompassing characteristics such as structure, composition, and appearance. It is regarded as a crucial parameter in cloud microphysics. The study of cloud particle phase states is regarded as crucial for a deep understanding of the impact of clouds on the earth’s climate, environment, and sustainable development, providing a scientific foundation for addressing climate change and improving environmental quality. Millimeter-wave cloud radar is widely used for identifying phase states of cloud particles due to its ability to detect internal cloud structures and microphysical characteristics. However, the cloud radar detects composite information about multiphase particles and lacks temperature data within cloud layers, leading to significant aliasing of phase state characteristics of cloud particles. Results in a high rate of misclassification when employing machine learning techniques for identifying the phase state of cloud particles.
To accurately identify phases of cloud particles, a Bayesian Random Forest Method is employed, utilizing co-located millimeter-wave cloud radar and microwave radiometer observations. Models are established separately for rainy and non-rainy weather conditions, utilizing the cloud radar’s reflectivity factor, spectral width of velocity, radial velocity, and depolarization ratio, as well as the microwave radiometer’s vertical temperature profile and the altitude of each parameter. These features are utilized to train models for the identification of snowflakes, ice crystals, liquid cloud droplets, and raindrops. Results indicate that the model achieves accuracies of 97%, 95%, 89%, and 96% in recognizing snowflakes, ice crystals, liquid cloud droplets, and raindrops on the test set, respectively. The precision, recall, and F1 scores for snowflakes, ice crystals, and raindrops all exceed 0.96, while the performance metrics for liquid cloud droplets are relatively poorer, which may be related to their vertical spatial distribution and sample size. By analyzing the feature importance of the model, it is found that the temperature feature carries the greatest weight in phase state recognition, followed by radial velocity, while the depolarization ratio contributes the least. Through case analysis of the evolution of observed factors such as reflectivity factor, radial velocity, and depolarization ratio, model recognition results show to be consistent with cloud microphysical processes. Statistical analysis shows that Bayesian Random Forest Method achieves a cloud particle phase state recognition accuracy of 96% under both rainy and non-rainy weather conditions. This algorithm can be applied to decision-making services for weather processes such as snowfall, enabling minute-level monitoring of precipitation types, phase transition timings, and evolution processes, thereby addressing the issue of automatic precipitation identification in operational settings.