光流融合方法在复杂地形区分钟级降水临近预报中的应用

Application of optical-flow fusion in minute-scale precipitation nowcasting over complex terrain

摘要

为提高复杂地形区 0—2 h 降水临近预报精度，本文融合循环全对场变换 RAFT (Recurrent All-pairs Field Transforms) 光流与经典的 Farneback 稠密光流方法，并提出“物理–数据双驱动”公里–分钟级光流融合方法 (Fused)，该方法将 RAFT 数据驱动光流场与 Farneback 稠密光流场进行加权融合，采用四阶龙格–库塔法（RK4）四步积分进行高精度平流外推，并构建残差 AR(2) 模型与 Kalman 误差反馈机制抑制强度漂移，得到逐 10 min 外推降水场。模型以 2022—2024 年 5—9 月中国区域 5 km 多源融合分析产品为训练样本，以 2025 年大别山地区 45 次独立降水过程为检验样本，从强度、空间结构、客体形态三个方面进行评估。结果表明：在 120 min 预报时效内，Fused 方法的均方根误差为 0.55 mm，较单独使用 RAFT 方法与 Farneback 方法分别降低 19% 与 34%；雨区相关系数达 0.51；基于对象的诊断评估显示，对象相似度为 0.73，面积膨胀小于 10%，质心漂移低于 5 km；1 h 雨区临界成功指数达 0.42，2 h 预报仍保持业务可用水平。融合光流场平均绝对散度为 2.8×10⁻⁵ s⁻¹，相对误差小于 5%，满足质量守恒近似要求，验证了外推雨带在强度演变上的物理合理性以及该降水临近预报的可靠性。

Abstract

To improve the accuracy of 0–2 h precipitation nowcasting over complex terrain, a kilometer- and minute-scale optical-flow fusion framework (Fused) driven by both physics and data principles is proposed, which integrates the Recurrent All-Pairs Field Transforms (RAFT) optical flow with the classical Farneback dense optical flow. This method performs a weighted fusion of the RAFT data-driven optical flow field and the Farneback dense optical flow field, employs a four-step fourth-order Runge-Kutta (RK4) integration scheme for high-accuracy advection extrapolation, and constructs a residual AR(2) model with Kalman error feedback to suppress intensity drift, thereby generating 10-minute-interval extrapolated precipitation fields. The model is trained on the 5-km multi-source merged precipitation analysis product over China from May to September during 2022–2024, and independently validated against 45 precipitation events in 2025 over the Dabie Mountain region, with evaluations conducted from the perspectives of intensity, spatial structure, and object morphology. Verification results show that, within the 120 min forecast lead time, the Fused method achieves an RMSE of 0.55 mm, representing reductions of 19% and 34% compared with the RAFT and Farneback methods, respectively; the rain-area correlation coefficient reaches 0.51. The object-based diagnostic evaluation reveals an object similarity of 0.73, area expansion below 10%, and centroid displacement less than 5 km. The critical success index (CSI) for 1-h rain areas reaches 0.42, while the 2-h forecast remains operationally usable. The mean absolute divergence of the fused optical flow field is 2.8×10⁻⁵ s⁻¹, with a relative error below 5%, satisfying the mass conservation approximation, which ensures the physical rationality of intensity evolution and the reliability of the nowcast.

关键词

复杂地形 / 光流融合 / RAFT / 分钟级临近预报

Key words

omplex terrainc / optical-flow fusion / RAFT / minute-scale nowcasting

参考文献

[1] 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.

[2] Cao C Y,Chen Y S, Liu D H, et al. 2015. The optical flow method and its application to nowcasting [J].Acta Meteorologica Sinica,73(3):471-480 (in Chinese). doi:10.11676/qxxb2015.034

[3] 曹子月,姚成,李致家,等. 2017.基于 DEM 的大别-皖南山区平均洪峰滞时定量分析[J].湖泊科学,29(3):765-774.

[4] Cao Z Y, Yao C, Li Z J, et al. 2017. DEM-based quantitative analysis of average peak time lag of Dabie-South Anhui mountain area [J].Journal of Lake Sciences,29(3):765-774 (in Chinese). doi:10.18307/2017.0326

[5] 陈明轩,肖现,高峰,等. 2016.基于雷达四维变分分析系统的强对流高分辨率模拟个例分析和批量检验[J].气象学报,74(3):421-441.

[6] Chen M X, Xiao X, Gao F, et al. 2016. A case study and batch verification on high resolution numerical simulations of sever convective events using an analysis system based on rapid-refresh 4-D variational rada data assimilation [J].Acta Meteorologica Sinica,74(3):421-441 (in Chinese). doi:10.11676/qxxb2016.031

[7] 陈永仁. 2024.四川极端性暴雨的研究进展及新认识[J].高原山地气象研究,44(1):1-9.

[8] Chen Y R. 2024. Research progress and new understanding of extreme rainstorms in Sichuan [J].Plateau and Mountain Meteorology Research,44(1):1-9. (in Chinese). doi:10.3969/j.issn.1674−2184.2024.01.001

[9] 程丛兰,陈明轩,王建捷,等. 2013.基于雷达外推临近预报和中尺度数值预报融合技术的短时定量降水预报试验[J].气象学报,71(3):397-415.

[10] Cheng C L, Chen M X, Wang J J, et al. 2013. Short-term quantitative precipitation fprecast experiments based on blending of noe-casting with numerical weather prediction [J].Acta MeteorologicaS-inica,71(3):397-415 (in Chinese). doi:10.11676/qxxb2013.041

[11] 方巍,沈亮,邹立尧,等. 2023.基于 GCA-ConvLSTM 预测网格的短临降水雷达回波外推方法[J].暴雨灾害,42(4):427-436.

[12] Fang W, Shen L, Zou L Y, et al. 2023. Extrapolation method of precipitation nowcasting radar echo based on GCA-ConvLSTM prediction network [J].Torrential Rain and Disasters,42(4):427-436 (in Chinese). doi:10.12406/byzh.2021−245

[13] 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.

[14] Guo H Y, Chen M X, Han L, et al. 2019. High resolution nowcasting experiment of severe convections based on deep learning [J].Acta Meteorologica Sinica,77(4):715-727. (in Chinese). doi:10.11676/qxxb2019.036

[15] 黄兴友,王捷,刘俊杰,等. 2018.基于模糊逻辑的地物回波识别方法及效果检验[J].热带气象学报,34(3):305-313.

[16] Huang X Y, Wang J, Liu J J, et al. 2018. The algorithm and verification of ground clutter identification based on fuzzy logic [J].Journal of tropical meteorology,34(3):305-313 (in Chinese). doi:10.16032/j.issn.1004−4965.2018.03.005

[17] 孔德璇. 2023.基于深度学习的贵州地区分钟级临近降水预报研究[D].南京:南京信息工程大学.

[18] Kong D X. 2023. Study in minute-level near-term precipitation forecasting in Guizhou based on deep learning [D]. Nanjing:Nanjing University of Information Science & Technology (in Chinese)

[19] 马雷. 2018.双偏振雷达反射率因子数据质量控制及在拼图中的应用[D].南京: 南京信息工程大学.

[20] Ma L. 2018. Data quality control for weather radar reflectivity factor and its application in regional radar composite [D]. Nanjing:Nanjing University of Information Science & Technology (in Chinese)

[21] 史小康,程文聪,张文军,等. 2022.基于 FY-4 红外资料和光流法的不同空间尺度云团的外推可预报性研究[J].气象,48(9):1101-1112.

[22] Shi X K, Cheng W C, Zhang W J, et al. 2022. Study on extrapolation predictability of cloud clusters in different space scales based on FY-4 infrared data and optical flow method [J].Meteorological Monthly,48(9):1101-1112 (in Chinese). doi:10.7519/j.issn.1000−0526.2022.113001

[23] 田刚,陈良华,魏凡,等. 2021.基于光流法雷达外推的 2020 年长江致洪降水临近预报检验评估[J].暴雨灾害,40(3):316-325.

[24] Tian G, Chen L H, Wei F, et al. 2021. Evaluation of flood-producing rainfall nowcasting based on radar extrapolation with the variational optical flow method in the Yangtze River Basin in 2020 [J].Torrential Rain and Disasters,40(3):316-325 (in Chinese). doi:10.3969/j.issn.1004−9045.2021.03.010

[25] 吴奕霄,李劲,项瑛,等. 2023. Farneback 光流法在雷达回波追踪中的应用[J].气象科学,43(4):515-522.

[26] Wu Y X, Li J, Xiang Y, et al. 2023. Application of Farneback optical flow method in radar echo tracking [J].Journal of the Meteorological Sciences,43(4):515-522 (in Chinese). doi:10.12306/2023jms.0039

[27] 邢楠,仲跻芹,雷蕾,等. 2023.基于 CMA-BJ 的北京地区短时强降水预报试验[J].应用气象学报,34(6):641-654.

[28] Xing N, Zhong J Q, Lei L, et al. 2023. A probabilistic forecast experiment of short-duration heavy rainfall in Beijing based on CMA-BJ [J].Journal of Applied Meteorological Science,34(6):641-654 (in Chinese). doi:10.11898/1001−7313.20230306001

[29] 杨绚,代刊,朱跃建. 2022.深度学习技术在智能网格天气预报中的应用进展与挑战[J].气象学报,80(5):649-667.

[30] Yang X, Dai K, Zhu J J. 2022. Progress and challenges of deep learning techniques in intelligent grid weather forecasting [J].Acta Meteorologica Sinica,80(5):649-667 (in Chinese). doi:10.11676/qxxb2022.051

[31] 袁凯,庞晶,李武阶,等. 2022.深度学习模型对武汉地区雷达回波临近预报的检验评估[J].暴雨灾害,41(4):458-466.

[32] Yuan K, Pang J, Li W J, et al. 2022. Examination and evaluation of deep learning models for radar echo nowcasting in Wuhan area [J].Torrential Rain and Disasters,41(4):458-466 (in Chinese). doi:10.3969/j.issn.1004−9045.2022.04.010

[33] 张博,张芳华,李晓兰,等. 2024. "23·7"华北特大暴雨数值预报检验评估[J].应用气象学报,35(1):17-32.

[34] Zhang B, Zhang F H, Li X L, et al. 2024. Verification and assessment of “23·7”sever rainstorm numerical prediction in north China [J].Journal of Applied Meteorological Science,35(1):17-32(in Chinese). doi:10.11898/1001−7313.20240102002

[35] 张蕾,魏鸣,李南,等. 2014.改进的光流法在回波外推预报中的应用[J].科学技术与工程,14(32):133-137.

[36] Zhang L, Wei M, Li N, et al. 2014. The compatibility of astragalus membranaceus and salvia miltiorrhiza extract control preliminary study of lipid metabolism in rats [J].Science Technology and Engineering,14(32):133-137 (in Chinese). doi:10.3969/j.issn.1671−1815.2014.32.024

[37] 庄潇然,郑玉,王亚强,等. 2023.基于深度学习的融合降水临近预报方法及其在中国东部地区的应用研究[J].气象学报,81(2):286-303.

[38] Zhuang X R, Zheng Y, Wang Y Q, et al. 2023. A deep learning-based precipitation nowcast model and its application over East China [J].Acta Meteorologica Sinica,81(2): 286-303 (in Chinese). doi:10.11676/qxxb2023.035

[39] Davis C A, Brown B G, Bullock R. 2006. Object-based verification of precipitation forecasts. Part I: Methodology and application to mesoscale rain areas[J].Monthly Weather Review,134(6):1772-1784. doi:10.1175/MWR3145.1

[40] Janez D. 2006. Statistical comparisons of cassifiers over multiple data sets [J].Journal of Machine Learning Research,7:1-30. doi:10.5555/1248547.1248548

[41] Krishnamurti T N, Bounoua L. 1996. An Introduction to Numerical Weather Prediction Techniques [M]. Boca Raton: CRC Press

[42] Liu J, Qian X, Peng L, et al. 2024. TEDR: A spatiotemporal attention radar extrapolation network constrained by optical flow and distribution correction [J].Atmospheric Research:311. doi:10.1016/j.atmosres.2024.107702

[43] Teed Z, Deng J. 2020. Recurrent all-pairs field transforms for optical flow [C]//Proceedings of the European Conference on Computer Vision (ECCV). Lecture Notes in Computer Science,12347:402-419. doi:10.1007/978−3−030−58536−5_24

[44] Williamson J H. 1980. Low-storage Runge-Kutta schemes[J].Journal of Computational Physics, 35(1): 48-56. doi:10.1016/0021−9991(80)90033−9

[45] Zhu W, Lu Z, Zhang Y, et al. 2025. A differential transformer-based hybrid deep learning model for radar echo extrapolation [J].Remote Sensing,17(1976):1-30. doi:10.3390/rs17121976

基金资助

安徽省气象局创新发展专项(CXM202403)