| The cloud particle phase refers to various forms of hydrometeor in the cloud,mainly in cluding rain,snow,supercooled water,mixed-phase state,ice crystal,etc.As one of the main cloud microphysical parameters and the main focus of cloud micro-detection,the distribution and evolution of cloud particle phase have exerted a crucial influence on such research as weather modification and aircraft icing.At present,the most commonly used cloud phase products are mainly based on weather radar and fuzzy logic algorithms.Weather radar has many polarizations that can describe particles finely,however,it is not sensitive to small particles due to wavelength.Millimeter wave cloud radar,whose wavelength is similar to cloud particles,is the best means to detect cloud microphysics at present.The fuzzy logic method is simple and convenient,but has a drawback: a large number of parameters need to be determined on the basis of expertise and experience.It is subjective and makes it inconvenient to get multi-source data involved.The neural network can obtain knowledge from samples,which reduce subjectivity and makes multi-source data addition more easily.Based on the data of millimeter wave cloud radar and other new ground-based remote sensing equipment,this paper uses a neural network to study the hydrometeor classification algorithm of cloud particles,which mainly includes:1.Based on the data of cloud radar and microwave radiometer from May 20,2021 to April1,2022 in Beijing Weather Observatory(Station No:54511),the dataset of cloud particle phase is established by combining the fuzzy logic method,ground station data and expert judgment.2.In order to mitigate the misidentification caused by the temperature jump of the microwave radiometer in the rain,two DNN(Deep Neural Networks)models,respectively for rainfall and non-rainfall,are established according to the weather conditions and put into use.Compared with the fuzzy logic method and DNN model,this method,according to experimental results,show,can effectively alleviate the phase misidentification caused by temperature jumps at rainfall.3.An improved Tri-training algorithm is proposed to solve the difficulty of obtaining labeled data.The algorithm uses the physical threshold in the field of the cloud particle phase to reduce noise.Experimental results show that this method is better than traditional Tri-training.4.In view of the lack of linear depolarization ratio in single-polarization radar,the recognition of complex hydrometeor,such as mixed phase,is not good.From the perspective of data correlation,the height sequence of a time point is taken as input,and the correlation is learned with Transformer.Experimental results show that compared with DNN,the accuracy of the algorithm proposed in this paper of the mixed phase increases by 4%.It reflects the effectiveness of learning data correlation in complex phase recognition.Finally,a series of models proposed in this paper are integrated by the Stacking method to get a hybrid model with better performance.5.A cloud microphysical analysis of a typical mixed stratiform cloud precipitation and snowfall process in Beijing Weather Observatory(Station No:54511)is performed using the algorithm proposed in this paper and the traditional fuzzy logic method.The analysis results show that the recognition results of the algorithm proposed in this paper are generally conform to the cloud microphysical processes,and are more robust in terms of data quality than the fuzzy logic method. |
