| 本文已被：浏览 1784次 下载 571次 |
码上扫一扫！ |
| 基于多尺度特征深度学习的短临降水预报 |
|
陈生1,2, 黄启桥1,2, 谭金凯2,3, 梁振清4, 吴翀5
|
|
1.中国科学院西北生态环境资源研究院/甘肃省遥感重点实验室/中国科学院黑河遥感试验研究站, 甘肃 兰州 730000;2. 南方海洋科学与工程广东省实验室（珠海），广东 珠海 519082;3.中山大学大气科学学院热带大气海洋系统科学教育部重点实验室，广东 珠海 519082;4.广西区气象技术装备中心，广西 南宁 530022;5.中国气象科学研究院，北京 100081
|
|
| 摘要: |
| 雷达回波外推是短临降水预报的一种重要方法。针对雷达回波外推中随着外推时间的增加而出现回波演变信息丢失这一问题，本文提出一种多尺度特征融合的深度学习短临降水预报模型(以下简称为MSF2)。首先，采用多尺度的卷积核对网络的浅层信息进行特征提取，弥补单一特征检测带来的不足。其次，将不同维度的特征信息进行拼接及通道混洗，进一步增强特征图通道之间的信息流通和信息表达能力。最后，将特征图中的多尺度信息进行融合，从而有效保留不同尺度的特征信息。利用华南雷达回波拼图资料数据，在3种不同降水强度（5 mm/h、10 mm/h和25 mm/h）下进行降水预报研究，并与光流法和ConvLSTM两种主流算法进行了对比。结果显示，在3种不同降水强度条件下，MSF2在所有评价指标（命中率POD、临界成功指数CSI、误报率FAR）中表现最优， 这表明引入多尺度机制能改善模型的特征提取能力。相比于目前主流的光流法和ConvLSTM，本文提出的模型对于短临降水预报具有较好的适用性和较高的预报精度,而且实现了业务化运行。 |
| 关键词: 短临预报 深度学习 多尺度特征 光流法 ConvLSTM |
| DOI：10.16032/j.issn.1004-4965.2023.069 |
| 分类号: |
| 基金项目: |
|
| SHORT-TERM PRECIPITATION NOWCASTING BASED ON MULTI-SCALE FEATURE DEEP LEARNING |
|
CHEN Sheng1,2, HUANG Qiqiao1,2, TAN Jinkai2,3, LIANG Zhenqing4, WU Chong5
|
|
1. Key Laboratory of Remote Sensing of Gansu Province, Heihe Remote Sensing Experimental Research Station，Northwest Institute of Eco-Environment and Resources, Chinese Academy of Sciences，Lanzhou 730000，China;2. Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai)，Zhuhai，Guangdong 519082，China;3. School of Atmospheric Sciences (Sun Yat-sen University)，Key Laboratory of Tropical Atmosphere-Ocean (Ministry of Education)，Zhuhai，Guangdong 519082，China;4. Meteorological Technical Equipment Center of Guangxi Zhuang Autonomous Region，Nanning 530022，China;5. Chinese Academy of Meteorological Sciences, Beijing 100081，China
|
| Abstract: |
| Radar echo extrapolation is an important method precipitation nowcasting. To address the problem of the loss of characteristic evolution information in echo extrapolation prediction with the increase of echo intensity and prediction time, this paper proposes a deep learning model for precipitation nowcasting based on multi-scale feature fusion (MSF2). Firstly, the multi-scale convolution kernel is used to extract the features of the shallow information of the network to offset the shortcomings caused by the single feature detection. Secondly, the feature information of different dimensions is spliced and the channels are shuffled to further enhance the information circulation and information expression capabilities between the feature map channels. Finally, the multi-scale information in the feature map is fused in order to effectively keep the channel information after the fusion of the feature map. With the South China radar echo data, the fusion experiment was carried out under three different precipitation intensities, and compared with two mainstream algorithms, i.e., ConvLSTM and optical flow. The experimental results show that MSF2 performs best in terms of all evaluation indexes under the conditions of precipitation rates 5 mm/h, 10 mm/h and 25 mm/h. It can be concluded that the introduction of a multi-scale mechanism can improve the feature extraction ability of the nowcasting model. Compared with the current radar echo extrapolation algorithm ConvLSTM and optical flow, the proposed model MSF2 has better potentials in operational applications and higher forecast accuracy for precipitation nowcasting. |
| Key words: nowcasting deep learning multi-scale features optical flow convLSTM |