APPLICATION OF MLP IN RADAR QUANTITATIVE PRECIPITATION ESTIMATION
-
摘要: 通过临近时次、临近空间降水回波性质相似假设,以人工智能技术结合快速动态分级法(Fast Dynamic Categorical method,FDC)为核心,设计广西区域分钟级雷达定量降水估测产品模型。在单部雷达估测降水时,分两层三次使用感知器寻求合理的降水类型分类Z-I关系,首先将FDC看为一种多分类算法,根据单站雷达各级回波的降水估测结果误差将回波区内的站点分为强站点和弱站点两类,然后分别对这两类站点再次使用FDC建立新的强、弱两类Z-I关系。在多部雷达组网联合估测定量降水时,将各雷达估测值等权重看待,将单部雷达估测作为一个分支,通过连结方式构建一个多层感知器(Multilayer Perceptron,MLP)。无站点回波区采用K近邻算法(K-Nearest Neighbor,KNN)选择合适的MLP求得的Z-I关系估算降水量。对2019年3—10月试验产品进行检验分析,结果表明以区域站组成的训练组小时降水相关系数达0.973 7,以国家级气象台站组成的测试组相关系数达0.825 6。Abstract: Based on the assumption of similar properties of precipitation radar echoes in adjacent time and space, a minute-level radar quantitative precipitation estimation(QPE) product model in Guangxi is designed according to a core method which combins AI with the fast dynamic classification method (FDC). When precipitation is estimated by a single radar, a reasonable Z-I relationship is calculated by using a two-layer perceptron for three times. First, FDC is considered d as a multi-classification algorithm. According to precipitation estimation error ofsingle radar echoes at all levels, the stations in the echo area are divided into two categories: strong stations and weak stations. And then, by means of FDC applied to these two types of stations respectively, new strong and weak Z-I relationships are found. As for multiple joint radars used in QPE, the estimated values of each radar are treated with equal weights, and the single radar estimation is considered as a branch. A Multilayer Perceptron(MLP) is constructed by connection. In the no-station echo area, the KNN algorithm is used to select the appropriate MLP to obtain the Z-I relationship for precipitation estimation. The test and analysis of the model products from March to October in 2019 showed that the correlation coefficient of hourly precipitation of the train group composed of regional stations was 0.973 7, and the correlation coefficient of the test group composed of national meteorological stations was 0.825 6.
-
图 7 时间同图 4,强弱站点空间分布(a)、KNN分类区域(b)、第一次QPE(c)、第二次QPE(d)
表 1 几种单雷达QPE方法,单时次统计指标
统计指标 固定关系法Z=300I1.4 最优法Z=3I2.3 FDC 两次分类FDC r 0.410 8 0.428 4 0.428 4 0.708 7 ME/(mm/(6 min)) -0.118 7 0.353 9 2.4×10-17 -1.9×10-16 RMSE/(mm/(6 min)) 0.369 4 0.580 3 0.338 9 0.260 3 表 2 Train/Test组统计指标
估测值 r ME/(mm/h) RMSE/(mm/h) Train组 0.973 7 0.002 7 0.394 9 Test组 0.825 6 -0.016 9 0.905 9 表 3 2019年Train/Test组r站数分布
r Train组 Test组 站数 比例 站数 比例 (0.9, 1.0] 2 602 98.75% 16 17.58% (0.8, 0.9] 22 0.83% 51 56.04% (0.7, 0.8] 7 0.27% 19 20.88% (0.6, 0.7] 1 0.04% 3 3.30% (0.5, 0.6] 1 0.04% 2 2.20% (0.4, 0.5] 1 0.04% 0 0.00% 表 4 Test组分级检验统计指标(分级阈值)
阈值/(mm/h) [0.1, 2) [2, 5) [5, 10) [10, 20) [20,) ME/(mm/h) 0.160 8 -0.076 6 -1.415 9 -4.596 8 -12.873 3 RMSE/(mm/h) 0.959 4 2.294 8 7.322 5 7.322 5 16.942 1 -
[1] 谷军霞, 师春香, 潘旸. 天气雷达定量估测降水研究进展[J]. 气象科技进展, 2018, 8(1): 71-78. [2] 潘旸, 谷军霞, 徐宾, 等. 多源降水数据融合研究及应用进展[J]. 气象科技进展, 2018, 8(1): 143-152. [3] ZHANG J, HOWARD K, LANGSTON C, et al. Multi-Radar Multi-Sensor(MRMS)quantitative precipitation estimation: Initial operating capabilities[J]. Bull Amer Meteor Soc, 2016, 97(4): 621-638. [4] 叶飞, 梁海河, 文浩, 等. 新一代天气雷达均一性评估[J]. 气象科技, 2020, 48(3): 322-330. [5] 叶飞, 梁海河, 文浩, 等. 基于单站CAPPI格点数据的相邻天气雷达均一性评估系统研究[J]. 气象, 2020, 46(1): 50-62. [6] 张林, 杨洪平, 裴翀, 等. 相邻雷达回波差异检测方法研究与应用[J]. 气象, 2018, 44(2): 268-276. [7] 韩静, 楚志刚, 王振会, 等. 苏南三部地基雷达反射率因子一致性和偏差订正个例研究[J]. 高原气象, 2017, 36(6): 1 665-1 673. [8] 张诚忠, 薛纪善, 冯业荣, 等. 基于贝叶斯方案的雷达反射率反演水汽及其同化试验[J]. 热带气象学报, 2019, 35(2): 145-153. [9] 刘陈帅, 张阿思, 陈生. 基于S波段双极化雷达的变分法的定量降水估计算法[J]. 热带气象学报, 2022, 38(3): 422-432. [10] 冀春晓, 陈联寿, 徐祥德, 等. 多普勒雷达资料动态定量估测台风小时降水量的研究[J]. 热带气象学报, 2008, 24(2): 147-155. [11] 勾亚彬, 刘黎平, 王丹, 等. 基于云团的分组Z-R关系拟合方案及效果评估[J]. 暴雨灾害, 2015, 34(1): 1-8. [12] 杨杰, 刘黎平, 赵城城, 等. 雷达估测对流性降水的误差空间分布及Z-R关系的优化[J]. 高原气象, 2015, 34(6): 1 785-1 796. [13] 寇蕾蕾, 李应超, 楚志刚, 等. C波段双偏振多普勒天气雷达资料分析及在定量估计降水中的应用研究[J]. 热带气象学报, 2018, 34(4): 460-471. [14] MAESAKA T, IWANAMI K, MAKI M. Non-negative KDP estimation by monotone assumption below melting layer[C]//in Extended Abstracts, Seventh European Conf. on Radar in Meteorology and Hydrology, 2012. [15] ZHANG A, XIAO L, MIN C, et al. Evaluation of latest GPM-Era high-resolution satellite precipitation products during the May 2017 Guangdong extreme rainfall event[J]. Atmos Res, 2019, 216(1): 76-85. [16] XIAO R, CHANDRASEKAR V. Development of a neural network based algorithm for rainfall estimation from radar observations[J]. IEEE Transactions on Geoscience and Remote Sensing, 1997, 35(1): 160-171. [17] CHEN H, CHANDRASEKA V, TAN H, et al. Rainfall estimation from ground radar and TRMM precipitation radar using hybrid deep neural networks[J]. Geophys Res Lett, 2019, 46(16/17): 10 669-10 678. [18] YANG T H, FENG L, CHANG L Y. Improving radar estimates of rainfall using an input subset of artificial neural networks[J]. Journal of Applied Remote Sensing, 2016, 10(2): 026013. [19] 邵月红, 张万昌, 刘永和, 等. BP神经网络在多普勒雷达降水量的估测中的应用[J]. 高原气象, 2009, 28(4): 846-853. [20] 傅德胜, 肖晨, 谭畅, 等. RBF神经网络在雷达定量估测降水中的应用研究[J]. 气象科学, 2015, 35(2): 199-203. [21] 殷志远, 彭涛, 杨芳, 等. 基于QPE和QPF的遗传神经网络洪水预报试验[J]. 暴雨灾害, 2013, 32(4): 360-368. [22] 陈训来, 陈元昭, 赵春阳, 等. 梯度提升决策树在雷达定量降水估测中的应用研究[J]. 气象科技进展, 2019, 9(3): 132-137. [23] 张晨阳, 杨雪冰, 张文生. 气象大数据超短临精准降水机器学习与典型应用[J]. 农业大数据学报, 2019, 1(1): 78-87. [24] 汪瑛, 冯业荣, 蔡锦辉, 等. 雷达定量降水动态分级Z-I关系估算方法[J]. 热带气象学报, 2011, 27(4): 601-608. [25] 卓健, 陈少斌, 周冬静, 等. 快速动态分级法进行雷达定量估测降水方法研究[J]. 热带气象学报, 2018, 34(6): 856-864. [26] 梁维亮, 黄荣, 翟丽萍, 等. 广西短时强降水雷达定量降水估测方法试验[J]. 灾害学, 2016, 31(2): 59-62. [27] 潘旸, 谷军霞, 宇婧婧, 等. 中国区域高分辨率多源降水观测产品的融合方法试验[J]. 气象学报, 2018, 76(5): 755-766.