An Automatic Rainfall‐Type Classification Algorithm Combining Diverse 3‐D Features of Radar Echoes

Abstract

This paper proposes an improved algorithm which combines neural networks with diverse 3-D structural features to partition radar reflectivity into convective and stratiform precipitation types. Radar data used in this work were obtained from three networked X-band Doppler radars located in Chengdu, China. The proposed algorithm consists of the two sections: six high-resolution features, which could be extracted from radar volume scanning and expressed the characteristics of the target in many ways, are selected as input to the neural network; systematic self-diagnosis is implemented; and the optimized model is determined according to analysis of bias and variance of classifier. Three state-of-art classification algorithms were implemented as references for algorithm evaluation. Both subjective comparison and statistical results convince that performance of the proposed algorithm is better than performance of traditional classification algorithms in various weather systems. The statistical results show that the proposed algorithm F-score values are 3%, 24%, and 30% higher than the fuzzy logic, SHY95, and BL algorithms and the recognition speed of the proposed algorithm is 54 times, two times, and four times that of fuzzy logic, SHY95, and BL, respectively. Considering network training is an offline procedure, classification by proposed neural network algorithm has great potential in real-time weather analysis for the precipitation classification. Plain Language SummaryIn recent years, networked radar become the popular way to overcome some of the inherent shortcomings of single radar. There are different scanning strategies for convective and stratiform clouds, and it is difficult to accurately distinguish between convective clouds and stratiform clouds, and when the accuracy is high, there will be a problem of low-recognition rate. Our algorithm utilizes a variety of features derived from radar products that have a good distinction between the two categories. At the same time, combined with neural network, its powerful nonlinear fitting and high speed make the recognition speed fast and the recognition result is high. The recognition result of our algorithm is higher than the recognition result of the popular fuzzy logic algorithm, and its recognition speed is even several times that of the fuzzy logic algorithm.

An Automatic Framework of Region‐of‐Interest Detection and Classification for Networked X‐Band Weather Radar System

“…Real‐time reflectivity data are introduced into the first block. In the first block, a back propagation neural network (BPNN) based classification algorithm (Lei et al, 2019) is applied to divide the precipitation cloud into convective rainfall and stratiform rainfall pixel by pixel. In the second block, density‐based spatial clustering algorithm (DBSCAN) is used to cluster the convective pixels into several convective cells.…”

“…In our previous study (Lei et al, 2019), six distinguished features are extracted from radar product, and a back propagation neural network (BPNN) model is carefully trained to identify convective pixels from the reflectivity image. This model is modified to realize cloud pixel classification in this block.…”

“…The modified features consist of cloud information at all the altitudes; hence, they are expected to better describe the horizontal structure of the target. For more details about BPNN, the reader is referred to Lei et al (2019).…”

“…This model is modified to realize cloud pixel classification in this block. Different from the original idea, all the horizontal features, including mean background reflectivity, horizontal gradient 10.1029/2020EA001181 of reflectivity, and area of region whose reflectivity is higher than 35 dBZ (shorted as F2, F4 and F5 in Lei et al, 2019), are calculated from the composed reflectivity of the radars, rather than the 3-km constant altitude plan position indicator puzzle (CAPPIPZ). The modified features consist of cloud information at all the altitudes; hence, they are expected to better describe the horizontal structure of the target.…”

“…Since the development of machine learning algorithms, precipitation cloud classification based on supervised machine learning has been published in the literature. These algorithms introduced hand‐made meteorological features into supervised machine learning methods such as fuzzy logic (Wang et al, 2012; Yang et al, 2013), decision tree (Gagne et al, 2009), neutral network (Lei et al, 2019; Xu et al, 2017), and others (Baldwin et al, 2005; McGovern et al, 2017). Performance of these algorithms is generally better than traditional algorithms like SHY95 and BL.…”