利用卷积神经网络开展偏振雷达定量降水估测研究

Research on Quantitative Precipitation Estimation by Polarized Radar Using CNN

关键词:  定量降水估测  卷积神经网络  S波段双偏振雷达  测雨精度

Key words:  Quantitative Precipitation Estimation(QPE), convolutional neural network, S-band dual polarization, measurement accuracy

摘要:

利用偏振升级改造后的广州新一代天气雷达（CINRAD/SAD）水平反射率ZH、差分传播相移率KDP、差分反射率因子ZDR和广东佛山219个地面气象自动站雨量数据，形成不同偏振量组合的8个数据集。基于卷积神经网络（CNN），建立雷达定量降水估测网络架构QPEnet， 并将该架构用于雷达定量降水估测（QPE），评估结果表明：数据集通道数N的增加可降低QPEnet的定量降雨估测的均方根误差（RMSE），并提高相关系数（CORR）；对于由ZH形成的数据集Z、Z_1~3 km和Z_6 min，随着通道数N的增加，数据集Z、Z_1~3 km和Z_6 min的性能逐步得到提高，数据集Z_1~3 km和Z_6 min的均方根误差（RMSE）分别是4.71和3.78，比数值集Z分别降低了1.3%和18.7%；数据集Z_1~3 km和Z_6 min的CORR分别是0.82和0.88，比数据集Z分别提高了2.5%和10.0%；对于ZH、KDP和ZDR偏振量组成的数据集里面，数据集Z_ZDR_KDP的拟合性能最好，RMSE为3.97，比数据集Z的RMSE降低了14.6%，CORR是0.86，比数据集Z提高了7.5%；分别对0.6～5 mm、5～10 mm、10～20 mm、20～30 mm、30～40 mm、40～50 mm和50 mm以上的7个降水量级的均方根误差（RMSE）、平均偏差比（MBR）、平均误差（AE）和相对误差（RE）等的统计结果表明，数据集Z_6 min降雨精度最高。

Abstract:

The ZH , ZDR and KDP of Guangzhou S-band dual polarization radar and rainfall data of 219 automatic meteorological stations in Foshan are used to form 8 datasets. Based on the convolutional neural network CNN, a radar quantitative precipitation estimation model is established, which will be used for ground precipitation estimation. The evaluating results of 8 datasets applied to the same precipitation estimation model are compared to each other. The results show that: The increase in the number of channels(N) of the datasets is beneficial to reduce the RMSE and improve CORR of the quantitative rainfall estimation results; For the datasets Z, Z_1~3 km and Z_6 min formed by ZH, as the number of channels increases, the performance of the data sets Z, Z_1~3 km and Z_6 min are gradually improved, and the RMSE of Z_1~3 km and Z_6 min are 4.71 and 3.78, which are -1.3% and 18.7% lower than that of dataset Z; the CORR of Z_1~3 km and Z_6 min are 0.82 and 0.88, which are 2.5% and 10% higher than that of dataset Z; Among other datasets composed of KDP and ZDR , the dataset Z_ZDR_KDP has the best fitting performance. The RMSE is 3.97, which is 14.6% lower than that of dataset Z, and the CORR is 0.86, which is 7.5% higher than that of dataset Z; The statistical results of RMSE, MBR, AE and RE for seven precipitation levels of 0.6～5 mm, 5～10 mm, 10～20 mm, 20～30 mm, 30～40 mm, 40～50 mm and above 50 mm respectively, show that dataset Z_6 min has the highest rainfall accuracy.