Application of Deep Learning Model TAGAN in Nowcasting of Strong Convective Echo
-
摘要: 近年来深度学习模型在解决对防灾减灾影响巨大且极具挑战性的临近预报问题的应用中日益增多。本文中，我们把临近预报作为一个时空序列预测的任务，将雷达反射率因子作为试验对象，使用基于对抗神经网络（GAN）优化构建的TAGAN深度学习模型预测未来1小时的雷达回波图像，并且与Rover光流法、基于卷积神经网络的3D U-Net模型进行对比试验。选取2018年全球气象AI挑战赛雷达回波数据集进行训练与测试，检验结果表明TAGAN模型在命中率（POD），虚警率（FAR），临界成功指数（CSI）以及相关系数等多种评分上要优于传统的光流法和对比的3D U-Net深度学习模型，TAGAN模型在以上的检验评分表现出色，并且随预测时间的增加较之传统光流模型效果更优，这为拓展和提升深度学习模型在临近天气预报中的应用提供了参考依据。Abstract: In recent years, deep learning models have been increasingly used in solving nowcasting problems that largely affect disaster prevention and mitigation. In this study, we take nowcasting as a spatio-temporal sequence prediction task and use the radar reflectivity factor as the test object. We use the two-stream attention generative adversarial network (TAGAN) deep learning model based on the Generative Adversarial Network frame to predict the radar echo image of the future 1 h and compare it with the Rover optical flow method and the 3D U-Net model based on the convolutional neural network. The radar echo data set of the 2018 Global Weather AI Challenge is selected for training and testing. The test results show that the TAGAN model has advances by multiple scores, including hit rate, false alarm rate, critical success index, and correlation coefficient. The TAGAN model performs well in these test scores and increases with the prediction time compared to the traditional optical flow method and the comparative deep learning model. Moreover, compared with that of the traditional optical flow model, the improvement effect of the TAGAN model is more significant. The results may shed some light on the expansion and improvement of the application of deep learning models in near-weather forecasting.
-
Key words:
- Nowcasting /
- Spatiotemporal prediction /
- Deep learning /
- Radar echo
-
图 3 CBAM模块结构图，包含通道注意力模块（左部）和空间注意力模型（右部）
Figure 3. CBAM module structure diagram, including the channel attention module (left part) and the spatial attention model (right part)
图 4 Self-attention模块结构图，
$\otimes $ 表示矩阵乘法运算Figure 4. Self-attention module structure diagram (
$\otimes $ : matrix multiplication operation)
图 6 30 dBZ阈值下三种模型FAR和HSS测试集平均得分随时间变化（折线），阴影上下界代表测试集所有样本得分的上下四分位数
Figure 6. Average scores of the False Alarm Rate and Heidke Skill Score test sets of the three models under 30 dBZ threshold change over time (broken line): the upper and lower bounds of the shade represent the upper and lower quartiles, respectively, of the scores of all samples in the test set
图 7 四组预测个例最后一帧对比。（a–d）分别为四组回波过程，每一列分别为真实过程、Rover、TAGAN、3DUnet在四组预测个例中的最后一帧
Figure 7. Comparison of the last frame of the four sets of prediction cases, where (a–d) are the four sets of echo processes, and each column is the last frame of the real process, Rover, TAGAN, and 3DUnet in the four sets of process predictions
图 8 四组预测个例三种模型（红色：3DUnet，绿色：Rover，蓝色：TAGAN）6～60 min 30 dBZ阈值的FAR（实线）和HSS（虚线）技巧评分
Figure 8. Four groups of prediction cases and three models (red: 3DUnet; green: Rover; and blue: TAGAN) under 6–60 min of 30 dBZ threshold FAR (solid line) and HSS (dashed line) skill scores
图 9 四组预测个例第一帧回波图真实值，其中红框为计算质心的范围（覆盖未来9帧回波主体范围），蓝线轮廓为回波值大于30 dBZ的回波主体
Figure 9. True value of the first frame of the four groups of prediction cases: the red box represents the range of the calculated centroid covering the range of the echo subject of the next nine frames), while the blue line outline is the echo subject with an echo value greater than 30 dBZ
图 10 四组预测个例10个时次回波值大于30 dBZ回波主体的质心移动轨迹。其中，红色：真实回波；棕色：3DUnet；灰色：Rover；绿色：TAGAN
Figure 10. Four groups of prediction cases during the 10 times the echo value is greater than 30 dBZ. Center of mass movement trajectory of the echo subject: real echo (red); 3DUnet (brown); Rover (gray); TAGAN (green)
表 2 测试集检验对比
Table 2. Test set comparison
阈值30 min预测 阈值60 min预测 模型 10 dBZ 20 dBZ 30 dBZ 40 dBZ 10 dBZ 20 dBZ 30 dBZ 40 dBZ CSI TAGAN 0.71 0.66 0.48 0.16 0.61 0.55 0.34 0.08 3DUnet 0.66 0.60 0.42 0.11 0.54 0.47 0.24 0.04 Rover 0.63 0.57 0.40 0.15 0.52 0.44 0.27 0.06 FAR TAGAN 0.20 0.23 0.30 0.39 0.27 0.31 0.40 0.51 3DUnet 0.27 0.29 0.38 0.56 0.38 0.40 0.53 0.75 Rover 0.23 0.29 0.43 0.66 0.33 0.41 0.60 0.82 POD TAGAN 0.86 0.81 0.60 0.19 0.80 0.71 0.44 0.10 3DUnet 0.87 0.79 0.54 0.13 0.81 0.66 0.31 0.05 Rover 0.78 0.73 0.57 0.19 0.70 0.63 0.43 0.09 HSS TAGAN 0.78 0.75 0.61 0.25 0.69 0.65 0.46 0.12 3DUnet 0.71 0.67 0.53 0.18 0.61 0.56 0.33 0.06 Rover 0.72 0.70 0.54 0.24 0.60 0.54 0.38 0.11 注：加粗的数字表示对比模型中检验评分最优 
下载: 导出CSV
表 3 测试集检验对比
Table 3. Test set comparison
模型 均方误差（MSE） 平均绝对误差（MAE） 相关系数
（COR）TAGAN 2046.30 8513.19 0.853 3DUnet 2610.69 10372.13 0.812 Rover 2922.36 10427.86 0.793 注：加粗的数字表示对比模型中检验评分最优
下载: 导出CSV
-
[1] Ballas N, Yao L, Pal C, et al. 2015. Delving deeper into convolutional networks for learning video representations [J]. https://arxiv.org/abs/1511.06432. doi: 48550/arXiv. 1511.06432 [2] 曹春燕, 陈元昭, 刘东华, 等. 2015. 光流法及其在临近预报中的应用 [J]. 气象学报, 73(3): 471−480. doi: 10.11676/qxxb2015.034Cao Chunyan, Chen Yuanzhao, Liu Donghua, et al. 2015. The optical flow method and its application to nowcasting [J]. Acta Meteorologica Sinica (in Chinese), 73(3): 471−480. doi: 10.11676/qxxb2015.034 [3] 陈明轩, 王迎春, 俞小鼎. 2007. 交叉相关外推算法的改进及其在对流临近预报中的应用 [J]. 应用气象学报, 18(5): 690−701. doi: 10.3969/j.issn.1001-7313.2007.05.014Chen Mingxuan, Wang Yingchun, Yu Xiaoding. 2007. Improvement and application test of TREC algorithm for convective storm nowcast [J]. Journal of Applied Meteorological Science (in Chinese), 18(5): 690−701. doi: 10.3969/j.issn.1001-7313.2007.05.014 [4] Cao Y, Li Q Y, Shan H M, et al. 2019. Precipitation nowcasting with star–bridge networks [J]. https://arxiv.org/abs/1907.08069v2 doi: 10.48550/arXiv.1907.08069 [5] Çiçek Ö, Abdulkadir A, Lienkamp S S, et al. 2016. 3D U–Net: Learning dense volumetric segmentation from sparse annotation [C]//19th International Conference on Medical Image Computing and Computer-Assisted Intervention. Athens, Greece: Springer, 424–432. doi: 10.1007/978-3-319-46723-8_49 [6] Goodfellow I J, Pouget-Abadie J, Mirza M, et al. 2014. Generative adversarial nets [C]//Proceedings of the 27th International Conference on Neural Information Processing Systems. Montreal, Canada: MIT Press, 2672–2680. [7] Graves A. 2013. Generating sequences with recurrent neural networks [J].https://arxiv.org/abs/1308.0850v3. doi: 10.48550/arXiv.1308.0850 [8] Gulrajani I, Ahmed F, Arjovsky M, et al. 2017. Improved training of wasserstein GANs [C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates Inc. , 5769–5779. [9] 郭瀚阳, 陈明轩, 韩雷, 等. 2019. 基于深度学习的强对流高分辨率临近预报试验 [J]. 气象学报, 77(4): 715−727. doi: 10.11676/qxxb2019.036Guo Hanyang, Chen Mingxuan, Han Lei, et al. 2019. High resolution nowcasting experiment of severe convections based on deep learning [J]. Acta Meteorologica Sinica (in Chinese), 77(4): 715−727. doi: 10.11676/qxxb2019.036 [10] 韩雷, 王洪庆, 谭晓光, 等. 2007. 基于雷达数据的风暴体识别、追踪及预警的研究进展 [J]. 气象, 33(1): 3−10. doi: 10.3969/j.issn.1000-0526.2007.01.001Han Lei, Wang Hongqing, Tan Xiaoguang, et al. 2007. Review on development of radar-based storm identification, tracking and forecasting [J]. Meteorological Monthly (in Chinese), 33(1): 3−10. doi: 10.3969/j.issn.1000-0526.2007.01.001 [11] 韩雷, 王洪庆, 林隐静. 2008. 光流法在强对流天气临近预报中的应用 [J]. 北京大学学报(自然科学版), 44(5): 751−755. doi: 10.3321/j.issn:0479-8023.2008.05.014Han Lei, Wang Hongqing, Lin Yinjing. 2008. Application of optical flow method to nowcasting convective weather [J]. Acta Scientiarum Naturalium Universitatis Pekinensis (in Chinese), 44(5): 751−755. doi: 10.3321/j.issn:0479-8023.2008.05.014 [12] 韩丰, 龙明盛, 李月安, 等. 2019. 循环神经网络在雷达临近预报中的应用 [J]. 应用气象学报, 30(1): 61−69. doi: 10.11898/1001-7313.20190106Han Feng, Long Mingsheng, Li Yuean, et al. 2019. The application of recurrent neural network to nowcasting [J]. Journal of Applied Meteorological Science (in Chinese), 30(1): 61−69. doi: 10.11898/1001-7313.20190106 [13] He K M, Zhang X Y, Ren S Q, et al. 2016. Deep residual learning for image recognition [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 770–778. doi: 10.1109/CVPR.2016.90 [14] Hogan R J, Ferro C A T, Jolliffe I T, et al. 2010. Equitability revisited: Why the “equitable threat score” is not equitable [J]. Wea. Forecasting, 25(2): 710−726. doi: 10.1175/2009WAF2222350.1 [15] Isola P, Zhu J Y, Zhou T H, et al. 2017. Image-to-image translation with conditional adversarial networks [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 5967–5976. doi: 10.1109/CVPR.2017.632 [16] Mathieu M, Couprie C, LeCun Y. 2016. Deep multi-scale video prediction beyond mean square error [C]. ICLR. [17] Oprea S, Martinez-Gonzalez P, Garcia-Garcia A, et al. 2020. A review on deep learning techniques for video prediction [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence. [18] Radford A, Metz L, Chintala S. 2015. Unsupervised representation learning with deep convolutional generative adversarial networks [C]. ICLR. [19] Shi X J, Chen Z R, Wang H, et al. 2015. Convolutional LSTM network: A machine learning approach for precipitation nowcasting [C]//Proceedings of the 28th International Conference on Neural Information Processing Systems. Montreal, Canada: MIT Press, 802–810. [20] Shi X J, Gao Z H, Lausen L, et al. 2017. Deep learning for precipitation nowcasting: A benchmark and a new model [C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates Inc. , 5622–5632. [21] 施恩, 李骞, 顾大权, 等. 2018. 基于卷积神经网络的雷达回波外推方法[J]. 计算机应用, 38(3): 661–665, 676Shi En, Li Qian, Gu Daquan, et al. Weather radar echo extrapolation method based on convolutional neural networks [J]. Journal of Computer Applications (in Chinese), 38(3): 661–665, 676. doi: 10.11772/j.issn.1001-9081.2017082098 [22] Wang Y B, Gao Z F, Long M S, et al. 2018. PredRNN++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning [C]//International Conference on Machine Learning. PMLR, 5123-5132. [23] Wang Y B, Zhang J J, Zhu H Y, et al. 2019. Memory in memory: A predictive neural network for learning higher-order non-stationarity from spatiotemporal dynamics [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 9146–9154. doi: 10.1109/CVPR.2019.00937 [24] 王国荣, 平凡, 翟亮. 2019. 一种基于数值模式诊断自适应的北京地区对流性降水临近集合预报新方法 [J]. 大气科学, 43(4): 895−914. doi: 10.3878/j.issn.1006-9895.1810.18177Wang Guorong, Ping Fan, Zhai Liang. 2019. A new method of adaptive convective precipitation ensemble nowcasting based on the numerical model diagnosis over Beijing [J]. Chinese Journal of Atmospheric Sciences (in Chinese), 43(4): 895−914. doi: 10.3878/j.issn.1006-9895.1810.18177 [25] Woo W C, Wong W K. 2017. Operational application of optical flow techniques to radar-based rainfall nowcasting [J]. Atmosphere, 8(3): 48. doi: 10.3390/atmos8030048 [26] Woo S, Park J, Lee J Y, et al. 2018. CBAM: Convolutional block attention module [C]//Proceedings of the 15th European Conference on Computer Vision (ECCV). Munich, Germany: Springer, 3–19. doi: 10.1007/978-3-030-01234-2_1 [27] 吴剑坤, 陈明轩, 秦睿, 等. 2019. 变分回波跟踪算法及其在对流临近预报中的应用试验 [J]. 气象学报, 77(6): 999−1014. doi: 10.11676/qxxb2019.062Wu Jiankun, Chen Mingxuan, Qin Rui, et al. 2019. The variational echo tracking method and its application in convective storm nowcasting [J]. Acta Meteorologica Sinica (in Chinese), 77(6): 999−1014. doi: 10.11676/qxxb2019.062 [28] 俞小鼎, 周小刚, 王秀明. 2012. 雷暴与强对流临近天气预报技术进展 [J]. 气象学报, 70(3): 311−337. doi: 10.11676/qxxb2012.030Yu Xiaoding, Zhou Xiaogang, Wang Xiuming. 2012. The advances in the nowcasting techniques on thunderstorms and severe convection [J]. Acta Meteorologica Sinica (in Chinese), 70(3): 311−337. doi: 10.11676/qxxb2012.030 [29] Zhang H, Goodfellow I, Metaxas D, et al. 2019. Self-attention generative adversarial networks [C]//Proceedings of the 36th International Conference on Machine Learning. Long Beach, California, USA: PMLR, 7354–7363. [30] Zhao H, Gallo O, Frosio I, et al. 2017. Loss functions for image restoration with neural networks [J]. IEEE Transactions on Computational Imaging, 3(1): 47−57. doi: 10.1109/TCI.2016.2644865 -
点击查看大图
计量
- 文章访问数: 851
- HTML全文浏览量: 125
- PDF下载量: 257
- 被引次数: 0
