Lightning nowcasting with missing multimodal information: three-dimensional sample construction and grouped adaptive fusion
-
摘要: 闪电临近预报对灾害预警及航空、电力安全至关重要，现有方法难以实现三维样本构建、多源异构特征融合和资料缺失条件下的稳定预报。针对上述问题，本文提出三维非极大值抑制(3D-NMS)的三维时空样本构建方法和闪电临近预报模型MG-SimVP。首先，提出3D-NMS方法，通过三维卷积计算候选区域的闪电密度，并利用非极大值抑制剔除低密度候选区域和高度重叠的冗余样本，以提取闪电聚集且主体完整的时空样本；其次，构建多源分组自适应融合编码器MGAF-Encoder，通过闪电、雷达和卫星特征的分组独立编码、全局通道响应感知、动态加权融合与共享特征回注，减弱强响应模态对稀疏闪电特征的掩盖，缓解多源异构资料混合造成的信息淹没；最后，面向多模态信息缺失场景，根据信息缺失引起的通道响应变化动态调整各模态融合权重，降低信息缺失的影响，提高模型预报稳定性。实验表明，训练样本中有效闪电信息更加集中，对流主体及其演变特征得到更完整保留，3D-NMS使基准模型的平均CSI提升13.0%，平均FAR降低20.2%；MG-SimVP在1小时预报窗口内的平均CSI达到0.670，较基准模型提升11.4%，平均FAR降低29.0%，且在信息缺失条件下CSI保持率不低于79.49%，表现出较好的预报精度与业务适应性。Abstract:
Objective Lightning nowcasting is essential for severe convective weather warning and the safety of aviation and power systems. However, existing methods face three major challenges. First, conventional fixed-window sampling tends to generate numerous non-lightning or weak-lightning background samples and may repeatedly divide or truncate the same convective system, resulting in incomplete representations of thunderstorm structures and temporal evolution. Second, substantial differences in sparsity, response intensity, and spatial distribution among lightning, radar, and satellite observations may cause strong-response modalities to obscure sparse lightning features during direct fusion. Third, multi-source observations may be delayed, missing, or abnormal in operational environments, leading to input distribution shifts and degraded forecasting performance. To address these problems, this paper proposes a three-dimensional Non-Maximum Suppression (3D-NMS) method for spatio-temporal sample construction and a Multi-source Grouping SimVP model, termed MG-SimVP for lightning nowcasting with missing multimodal information. Methods MG-SimVP is developed within the encoder–translator–decoder framework of SimVP. The model uses VLF-LLN lightning observations, weather radar reflectivity, and Himawari-8 satellite features as inputs. All observations are remapped to a 256×256 grid and temporally aligned at 10-minute intervals. Six consecutive historical frames are used to predict the lightning density distributions of the following six frames, corresponding to a 1-hour forecast window. construct spatio-temporal samples, the original lightning data are first organized into a 1024 ×1024 ×T spatio-temporal tensor. Candidate samples are extracted using a 256×256×12 Gaussian sliding window. Three-dimensional convolution is then employed to calculate the lightning density concentration response of each candidate. Background filtering removes samples with weak lightning activity or displaced convective centers, while non-maximum suppression eliminates highly overlapping candidates and retains samples with stronger responses. The proposed Multi-source Grouped Adaptive Fusion Encoder (MGAF-Encoder) separately encodes lightning, radar, and satellite features. Global average pooling and modality-specific Multi-Layer Perceptrons (MLPs) are used to estimate adaptive channel weights, after which the aligned modality features are dynamically fused. The shared fusion feature is fed back to each branch to enhance cross-modal information exchange. When a modality is missing, it is replaced with a zero-valued tensor. MGAF-Encoder adjusts the fusion weights according to changes in the global channel responses, reducing the contribution of the missing modality and strengthening the remaining valid observations. The fused features are further processed by Mid_Xnet for multi-scale spatio-temporal evolution modeling and decoded to generate future lightning density fields.Results and Discussions Compared with the unscreened sampling strategy, 3D-NMS increases the average CSI of the multi-source SimVP model from 0.5323 to0.6014 , corresponding to a relative improvement of 13.0%. Meanwhile, the average FAR decreases from0.4155 to0.3317 , a relative reduction of 20.2%. These results indicate that 3D-NMS effectively removes low-value background and redundant samples, concentrates valid lightning information, and preserves the principal thunderstorm structures and their continuous evolution. Under complete multi-source inputs, MG-SimVP outperforms the standard SimVP model at all lead times from 10 to 60 minutes. The average CSI increases from0.6014 to0.6701 , representing an improvement of 11.4%, while the average FAR decreases from0.3317 to0.2354 , representing a reduction of 29.0%. At the 60-minute lead time, MG-SimVP maintains a CSI of0.5570 , compared with0.4594 for the standard SimVP. This demonstrates that grouped encoding and adaptive fusion can reduce the masking effect of strong radar and satellite responses on sparse lightning features. Incomplete-input experiments further demonstrate the robustness of MG-SimVP. When satellite observations are missing, the model retains 88.24% of the complete-input CSI performance. When radar observations are missing, the retention rate is 84.10%. Even when only historical lightning observations are available, the model retains 79.49% of the complete-input performance. The model therefore avoids severe forecasting failure under missing-data conditions and maintains the identification of the principal lightning activity regions.Conclusions This paper proposes a 3D-NMS-based spatio-temporal sample construction method and a multi-source lightning nowcasting model, MG-SimVP, for missing multimodal information. The 3D-NMS method improves the concentration of valid lightning information and preserves complete convective structures and continuous evolution while removing low-density and highly redundant samples. MGAF-Encoder suppresses the masking effect of strong-response modalities on sparse lightning features through grouped independent encoding, global channel response perception, adaptive weighted fusion, and shared-feature feedback. Under missing observations, the encoder dynamically adjusts modality contributions according to changes in channel responses, thereby reducing the influence of missing data and strengthening the contribution of the remaining valid observations. Experimental results show that MG-SimVP achieves an average CSI of 0.6701 within the 1-hour forecast window, improving the standard multi-source SimVP baseline by 11.4%, while reducing the average FAR by 29.0%. Under missing-input conditions, the model retains at least 79.49% of its complete-input CSI performance. The proposed method therefore exhibits satisfactory lightning-region identification accuracy, false-alarm control, robustness, and operational adaptability. Future work will incorporate richer dynamic and thermodynamic variables, including three-dimensional radar echoes and numerical weather prediction fields, and will further optimize the sample construction strategy and loss function to improve forecasting performance for sudden, small-scale, and locally high-density lightning events. -
表 1 闪电输入SimVP
Lead Time(min) POD ↑ FAR ↓ CSI ↑ 10 0.8450 0.2214 0.6813 20 0.8083 0.2623 0.6279 30 0.7964 0.3651 0.5462 40 0.8027 0.4869 0.4557 50 0.8129 0.5877 0.3766 60 0.7755 0.5746 0.3788 
下载: 导出CSV
表 3 多源输入MG-SimVP
Lead Time(min) POD ↑ FAR ↓ CSI ↑ 10 0.8814 0.1028 0.8006 20 0.8695 0.1609 0.7453 30 0.8488 0.2141 0.6894 40 0.8366 0.2699 0.6390 50 0.8155 0.3198 0.5895 60 0.7883 0.3450 0.5570
下载: 导出CSV
表 2 多源输入SimVP
Lead Time(min) POD ↑ FAR ↓ CSI ↑ 10 0.8996 0.1745 0.7540 20 0.8884 0.2418 0.6923 30 0.8431 0.2826 0.6329 40 0.8408 0.3682 0.5643 50 0.8255 0.4339 0.5056 60 0.8210 0.4894 0.4594
下载: 导出CSV
表 4 未使用3D-NMS三维时空样本构建的多源输入SimVP
Lead Time(min) POD ↑ FAR ↓ CSI ↑ 10 0.8323 0.2014 0.6879 20 0.8810 0.3278 0.6163 30 0.8885 0.4422 0.5213 40 0.8638 0.4551 0.5018 50 0.8651 0.5156 0.4504 60 0.8505 0.5509 0.4162
下载: 导出CSV
表 5 VLF-LLN+雷达输入MG-SimVP
Lead Time(min) POD ↑ FAR ↓ CSI ↑ 10 0.9109 0.1903 0.7502 20 0.8973 0.2552 0.6864 30 0.8910 0.3402 0.6105 40 0.8757 0.3916 0.5601 50 0.8649 0.3198 0.5029 60 0.8594 0.5298 0.4380
下载: 导出CSV
表 7 仅VLF-LLN输入MG-SimVP
Lead Time(min) POD ↑ FAR ↓ CSI ↑ 10 0.8647 0.1957 0.7143 20 0.8472 0.2919 0.6279 30 0.8077 0.3543 0.5597 40 0.8207 0.4555 0.4866 50 0.7997 0.5348 0.4166 60 0.7982 0.5658 0.3912
下载: 导出CSV
表 6 VLF-LLN+卫星输入MG-SimVP
Lead Time(min) POD ↑ FAR ↓ CSI ↑ 10 0.8593 0.1161 0.7721 20 0.8771 0.2670 0.6647 30 0.8582 0.3815 0.5612 40 0.8334 0.4419 0.5021 50 0.8070 0.4918 0.4531 60 0.7774 0.5120 0.4282
下载: 导出CSV
-
[1] CHEN Zhiwei, ZHAO Xinjie, XIANG Nianwen, et al. Bipolar leader theory-based simulation analysis of aircraft lightning strike attachment zones[J]. Journal of Electrostatics, 2025, 138: 104199. doi: 10.1016/j.elstat.2025.104199. [2] HUO Hong, WANG Dezhi, CHEN Hao, et al. Considering the methods of lightning protection and early warning for power transmission lines based on lightning data analysis[J]. IEEE Access, 2024, 12: 54168–54181. doi: 10.1109/ACCESS.2024.3387968. [3] ZHANG Zhejia, TIAN Ye, WANG Guangyu, et al. A forest fire prediction method for lightning stroke based on remote sensing data[J]. Forests, 2024, 15(4): 647. doi: 10.3390/f15040647. [4] QIE Xiushu, LIU Dongxia, JIANG Rubin, et al. Detection, mechanism, and forecasting of lightning and thunderstorms[J]. Journal of Meteorological Research, 2025, 39(3): 741–761. doi: 10.1007/s13351-025-5033-1. [5] GUO Shuchang, WANG Jinyan, GAN Ruhui, et al. Experimental study of cloud-to-ground lightning nowcasting with multisource data based on a video prediction method[J]. Remote Sensing, 2022, 14(3): 604. doi: 10.3390/rs14030604. [6] SAUNDERS C. Charge separation mechanisms in clouds[J]. Space Science Reviews, 2008, 137(1): 335–353. doi: 10.1007/s11214-008-9345-0. [7] NEEL M S. Modeling the electric potential and surface charge density near charged thunderclouds[J]. The Physics Teacher, 2018, 56(3): 165–167. doi: 10.1119/1.5025296. [8] NICORA M, MESTRINER D, BRIGNONE M, et al. Assessment of the lightning performance of overhead distribution lines based on lightning location systems data[J]. International Journal of Electrical Power & Energy Systems, 2022, 142: 108230. doi: 10.1016/j.ijepes.2022.108230. [9] KUMAR K M and REDDY A R M. A fast DBSCAN clustering algorithm by accelerating neighbor searching using Groups method[J]. Pattern Recognition, 2016, 58: 39–48. doi: 10.1016/j.patcog.2016.03.008. [10] CHOI C and HONG S Y. MDST-DBSCAN: A density-based clustering method for multidimensional spatiotemporal data[J]. ISPRS International Journal of Geo-Information, 2021, 10(6): 391. doi: 10.3390/ijgi10060391. [11] GHANBARPOUR A and MINAEI B. EXDBSCAN: An extension of DBSCAN to detect clusters in multi-density datasets[C]. 2014 Iranian Conference on Intelligent Systems, Bam, Iran, 2014: 1–5. doi: 10.1109/IranianCIS.2014.6802561. [12] PARAMANIK M M R, RABBANI K M G, IMRAN A, et al. Prediction of lightning activity over Bangladesh using diagnostic and explicit lightning parameterizations of WRF model[J]. Natural Hazards, 2024, 120(5): 4399–4422. doi: 10.1007/s11069-023-06355-6. [13] LIU Dongxia, YU Han, and SUN Chunfa. Estimation of lightning activity of squall lines by different lightning parameterization schemes in the weather research and forecasting model[J]. Remote Sensing, 2023, 15(20): 5070. doi: 10.3390/rs15205070. [14] MANDÚ T B, ALVES L E R, VENDRASCO É P, et al. Development of vertical radar reflectivity profiles based on lightning density using the geostationary lightning mapper dataset in the subtropical region of Brazil[J]. Remote Sensing, 2024, 16(20): 3767. doi: 10.3390/rs16203767. [15] ZHOU Wei, LI Jinliang, WANG Hongjie, et al. Forecasting of local lightning using spatial–channel-enhanced recurrent convolutional neural network[J]. Atmosphere, 2024, 15(12): 1478. doi: 10.3390/atmos15121478. [16] ZHOU Xinyuan, GENG Yangli, YU Haomin, et al. LightNet+: A dual-source lightning forecasting network with bi-direction spatiotemporal transformation[J]. Applied Intelligence, 2022, 52(10): 11147–11159. doi: 10.1007/s10489-021-03089-5. [17] JING Jinrui, LI Qian, and PENG Xuan. MLC-LSTM: Exploiting the spatiotemporal correlation between multi-level weather radar echoes for echo sequence extrapolation[J]. Sensors, 2019, 19(18): 3988. doi: 10.3390/s19183988. [18] 宋毅, 张晗奕, 孙丰, 等. PPNet: 基于预先预测的降雨短时预测模型[J]. 电子与信息学报, 2024, 46(2): 492–502. doi: 10.11999/JEIT230547.SONG Yi, ZHANG Hanyi, SUN Feng, et al. PPNet: A precipitation nowcasting model based on pre-prediction[J]. Journal of Electronics & Information Technology, 2024, 46(2): 492–502. doi: 10.11999/JEIT230547. [19] GAO Zhangyang, TAN Cheng, WU Lirong, et al. SimVP: Simpler yet better video prediction[C]. Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 3160–3170. doi: 10.1109/CVPR52688.2022.00317. [20] TAN Cheng, GAO Zhangyang, LI Siyuan, et al. SimVPv2: Towards simple yet powerful spatiotemporal predictive learning[J]. IEEE Transactions on Multimedia, 2025, 27: 5170–5184. doi: 10.1109/tmm.2025.3543051. [21] WANG Xiaoying, ZHAO Haixiang, ZHANG Guojing, et al. Spatiotemporal predictive learning for radar-based precipitation nowcasting[J]. Atmosphere, 2024, 15(8): 914. doi: 10.3390/atmos15080914. [22] 唐志豪, 韩袁鹏, 张慧, 等. 时空聚焦感知与协同加权优化的多任务闪电临近预报方法[J]. 电子与信息学报, 2026, 48(12). doi: 10.11999/JEIT260234.TANG Zhihao, HAN Yuanpeng, ZHANG Hui, et al. Multi-task lightning nowcasting with spatio-temporal focal perception and synergistic weighted loss[J]. Journal of Electronics & Information Technology, 2026, 48(12). doi: 10.11999/JEIT260234. [23] LU Mingyue, JIN Chuanwei, YU Manzhu, et al. MCGLN: A multimodal ConvLSTM-GAN framework for lightning nowcasting utilizing multi-source spatiotemporal data[J]. Atmospheric Research, 2024, 297: 107093. doi: 10.1016/j.atmosres.2023.107093. [24] HAN D, CHOO M, IM J, et al. Precipitation nowcasting using ground radar data and simpler yet better video prediction deep learning[J]. GIScience & Remote Sensing, 2023, 60(1): 2203363. doi: 10.1080/15481603.2023.2203363. [25] 郑辉, 陈富, 何舒平, 等. 一种多尺度时空相关注意力与状态空间建模的降水临近预报方法[J]. 电子与信息学报, 2026, 48(3): 1219–1229. doi: 10.11999/JEIT250786.ZHENG Hui, CHEN Fu, HE Shuping, et al. A multi-scale spatiotemporal correlation attention and state space modeling-based approach for precipitation nowcasting[J]. Journal of Electronics & Information Technology, 2026, 48(3): 1219–1229. doi: 10.11999/JEIT250786. [26] JUMIANTI N, MARZUKI M, YUSNAINI H, et al. Prediction of extreme rain in Kototabang using Himawari-8 satellite based on differences in cloud brightness temperature[J]. Remote Sensing Applications: Society and Environment, 2024, 33: 101102. doi: 10.1016/j.rsase.2023.101102. [27] 高晓荣, 梁建茵, 李春晖, 等. 多平台(雷达、卫星、雨量计)降水信息的融合技术初探[J]. 高原气象, 2013, 32(2): 549–555. doi: 10.7522/j.issn.1000-0534.2012.00052.GAO Xiaorong, LIANG Jianyin, LI Chunhui, et al. Preliminary studies on merged techniques based on precipitation information from multiplatform (radar, satellite and rain gauge)[J]. Plateau Meteorology, 2013, 32(2): 549–555. doi: 10.7522/j.issn.1000-0534.2012.00052. [28] SHEN Yan, HONG Zhen, PAN Yang, et al. China’s 1 km merged gauge, radar and satellite experimental precipitation dataset[J]. Remote Sensing, 2018, 10(2): 264. doi: 10.3390/rs10020264. [29] DONG Jianzhi, CROW W T, CHEN Xi, et al. Statistical uncertainty analysis-based precipitation merging (SUPER): A new framework for improved global precipitation estimation[J]. Remote Sensing of Environment, 2022, 283: 113299. doi: 10.1016/j.rse.2022.113299. [30] GYASI-AGYEI Y. A framework for comparing two rainfields based on spatial structure: A case of radar against selected satellite precipitation products over southeast Queensland, Australia[J]. Journal of Hydrology, 2022, 613: 128356. doi: 10.1016/j.jhydrol.2022.128356. [31] LIU Yi, WANG Jialei, SONG Yang, et al. Lightning nowcasting based on high-density area and extrapolation utilizing long-range lightning location data[J]. Atmospheric Research, 2025, 321: 108070. doi: 10.1016/j.atmosres.2025.108070. [32] HUANG Yijun, FAN Yadong, CAI Li, et al. A new thunderstorm identification algorithm based on total lightning activity[J]. Earth and Space Science, 2022, 9(4): e2021EA002079. doi: 10.1029/2021EA002079. -
下载:
图(6) / 表(7)
计量
- 文章访问数: 20
- HTML全文浏览量: 3
- PDF下载量: 1
- 被引次数: 0
