• Ahijevych, D., J. O. Pinto, J. K. Williams, and M. Steiner, 2016: Probabilistic forecasts of mesoscale convective system initiation using the random forest data mining technique. Wea. Forecasting, 31, 581599, https://doi.org/10.1175/WAF-D-15-0113.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Audebert, N., B. Le Saux, and S. Lefèvre, 2016: Semantic segmentation of Earth observation data using multimodal and multi-scale deep networks. 13th Asian Conf. on Computer Vision, Taipei, Taiwan, Asian Federation of Computer Vision, 180–196.

    • Crossref
    • Export Citation
  • Badrinarayanan, V., A. Kendall, and R. Cipolla, 2017: SegNet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39, 24812495, https://doi.org/10.1109/TPAMI.2016.2644615.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bechini, R., and V. Chandrasekar, 2017: An enhanced optical flow technique for radar nowcasting of precipitation and winds. J. Atmos. Oceanic Technol., 34, 26372658, https://doi.org/10.1175/JTECH-D-17-0110.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bengio, Y., 2009: Learning deep architectures for AI. Found. Trends Mach. Learn., 2, 1127, https://doi.org/10.1561/2200000006.

  • Bessho, K., and et al. , 2016: An introduction to Himawari-8/9—Japan’s new-generation geostationary meteorological satellites. J. Meteor. Soc. Japan, 94, 151183, https://doi.org/10.2151/jmsj.2016-009.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Betz, H. D., K. Schmidt, W. P. Oettinger, and B. Montag, 2008: Cell-tracking with lightning data from LINET. Adv. Geosci., 17, 5561, https://doi.org/10.5194/adgeo-17-55-2008.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bonelli, P., and P. Marcacci, 2008: Thunderstorm nowcasting by means of lightning and radar data: Algorithms and applications in northern Italy. Nat. Hazards Earth Syst. Sci., 8, 11871198, https://doi.org/10.5194/nhess-8-1187-2008.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Buechler, D. E., and S. J. Goodman, 1990: Echo size and asymmetry: Impact on NEXRAD storm identification. J. Appl. Meteor., 29, 962969, https://doi.org/10.1175/1520-0450(1990)029<0962:ESAAIO>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Chen, L. C., G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, 2018: DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE Trans. Pattern Anal. Mach. Intell., 40, 834848, https://doi.org/10.1109/TPAMI.2017.2699184.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Chen, M., Y. Wang, F. Gao, and X. Xiao, 2012: Diurnal variations in convective storm activity over contiguous North China during the warm season based on radar mosaic climatology. J. Geophys. Res., 117, D20115, https://doi.org/10.1029/2012JD018158.

    • Search Google Scholar
    • Export Citation
  • Dabberdt, W. F., and et al. , 2005: Multifunctional mesoscale observing networks. Bull. Amer. Meteor. Soc., 86, 961982, https://doi.org/10.1175/BAMS-86-7-961.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm identification, tracking, analysis, and nowcasting—A radar-based methodology. J. Atmos. Oceanic Technol., 10, 785797, https://doi.org/10.1175/1520-0426(1993)010<0785:TTITAA>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Farnell, C., T. Rigo, and N. Pineda, 2017: Lightning jump as a nowcast predictor: Application to severe weather events in Catalonia. Atmos. Res., 183, 130141, https://doi.org/10.1016/j.atmosres.2016.08.021.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Garcia-Garcia, A., S. Orts-Escolano, S. Oprea, V. Villena-Martinez, and J. Garcia-Rodriguez, 2017: A review on deep learning techniques applied to semantic segmentation. arXiv, http://arxiv.org/abs/1704.06857.

  • Gatlin, P. N., and S. J. Goodman, 2010: A total lightning trending algorithm to identify severe thunderstorms. J. Atmos. Oceanic Technol., 27, 322, https://doi.org/10.1175/2009JTECHA1286.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Golik, P., P. Doetsch, and H. Ney, 2013: Cross-entropy vs. squared error training: A theoretical and experimental comparison. Proc. 14th Annual Conf., Lyon, France, International Speech Communication Association, 1756–1760.

  • Haberlie, A. M., and W. S. Ashley, 2018: A method for identifying midlatitude mesoscale convective systems in radar mosaics. Part I: Segmentation and classification. J. Appl. Meteor. Climatol., 57, 15751598, https://doi.org/10.1175/JAMC-D-17-0293.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Han, L., J. Sun, W. Zhang, Y. Xiu, H. Feng, and Y. Lin, 2017: A machine learning nowcasting method based on real-time reanalysis data. J. Geophys. Res. Atmos., 122, 40384051, https://doi.org/10.1002/2016JD025783.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Harris, R. J., J. R. Mecikalski, W. M. Mackenzie, P. A. Durkee, and K. E. Nielsen, 2010: The definition of GOES infrared lightning initiation interest fields. J. Appl. Meteor. Climatol., 49, 25272543, https://doi.org/10.1175/2010JAMC2575.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hou, A. Y., and et al. , 2014: The Global Precipitation Measurement Mission. Bull. Amer. Meteor. Soc., 95, 701722, https://doi.org/10.1175/BAMS-D-13-00164.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ioffe, S., and C. Szegedy, 2015: Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv, https://arxiv.org/abs/1502.03167.

  • Ji, S., M. Yang, and K. Yu, 2013: 3D convolutional neural networks for human action recognition. IEEE Trans. Pattern Anal. Mach. Intell., 35, 221231, https://doi.org/10.1109/TPAMI.2012.59.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Johnson, J. T., P. L. MacKeen, A. Witt, E. D. W. Mitchell, G. J. Stumpf, M. D. Eilts, and K. W. Thomas, 1998: The storm cell identification and tracking algorithm: An enhanced WSR-88D algorithm. Wea. Forecasting, 13, 263276, https://doi.org/10.1175/1520-0434(1998)013<0263:TSCIAT>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Karagiannidis, A., K. Lagouvardos, and V. Kotroni, 2016: The use of lightning data and Meteosat infrared imagery for the nowcasting of lightning activity. Atmos. Res., 168, 5769, https://doi.org/10.1016/j.atmosres.2015.08.011.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kingma, D., and J. Ba, 2015: ADAM: A method for stochastic optimization. Third Int. Conf. Learning Representations, San Diego, CA, International Machine Learning Society.

  • Kohn, M., E. Galanti, C. Price, K. Lagouvardos, and V. Kotroni, 2011: Nowcasting thunderstorms in the Mediterranean region using lightning data. Atmos. Res., 100, 489502, https://doi.org/10.1016/j.atmosres.2010.08.010.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • LeCun, Y., Y. Bengio, and G. Hinton, 2015: Deep learning. Nature, 521, 436444, https://doi.org/10.1038/nature14539.

  • Lee, S., H. Han, J. Im, E. Jang, and M. Lee, 2017: Detection of deterministic and probabilistic convection initiation using Himawari-8 Advanced Himawari Imager data. Atmos. Meas. Tech., 10, 18591874, https://doi.org/10.5194/amt-10-1859-2017.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Lynn, B. H., Y. Yair, C. Price, G. Kelman, and A. J. Clark, 2012: Predicting cloud-to-ground and intracloud lightning in weather forecast models. Wea. Forecasting, 27, 14701488, https://doi.org/10.1175/WAF-D-11-00144.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., and K. M. Bedka, 2006: Forecasting convective initiation by monitoring the evolution of moving cumulus in daytime GOES imagery. Mon. Wea. Rev., 134, 4978, https://doi.org/10.1175/MWR3062.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., W. M. Mackenzie, M. König, and S. Muller, 2010: Cloud-top properties of growing cumulus prior to convective initiation as measured by Meteosat Second Generation. Part II: Use of visible reflectance. J. Appl. Meteor. Climatol., 49, 25442558, https://doi.org/10.1175/2010JAMC2480.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., X. Li, L. D. Carey, E. W. McCaul, and T. A. Coleman, 2013: Regional comparison of GOES cloud-top properties and radar characteristics in advance of first-flash lightning initiation. Mon. Wea. Rev., 141, 5574, https://doi.org/10.1175/MWR-D-12-00120.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., J. K. Williams, C. P. Jewett, D. Ahijevych, A. LeRoy, and J. R. Walker, 2015: Probabilistic 0–1-h convective initiation nowcasts that combine geostationary satellite observations and numerical weather prediction model data. J. Appl. Meteor. Climatol., 54, 10391059, https://doi.org/10.1175/JAMC-D-14-0129.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Medina, B. L., L. D. Carey, C. G. Amiot, R. M. Mecikalski, W. P. Roeder, T. M. McNamara, and R. J. Blakeslee, 2019: A random forest method to forecast downbursts based on dual-polarization radar signatures. Remote Sens., 11, 826, https://doi.org/10.3390/rs11070826.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Metzger, E., and W. A. Nuss, 2013: The relationship between total cloud lightning behavior and radar-derived thunderstorm structure. Wea. Forecasting, 28, 237253, https://doi.org/10.1175/WAF-D-11-00157.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mosier, R. M., C. Schumacher, R. E. Orville, and L. D. Carey, 2011: Radar nowcasting of cloud-to-ground lightning over Houston, Texas. Wea. Forecasting, 26, 199212, https://doi.org/10.1175/2010WAF2222431.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Orlanski, I., 1975: A rational subdivision of scales for atmospheric processes. Bull. Amer. Meteor. Soc., 56, 527530, https://doi.org/10.1175/1520-0477-56.5.527.

    • Search Google Scholar
    • Export Citation
  • Perol, T., M. Gharbi, and M. Denolle, 2018: Convolutional neural network for earthquake detection and location. Sci. Adv., 4, e1700578, https://doi.org/10.1126/sciadv.1700578.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ray, P. S., 1986: Mesoscale Meteorology and Forecasting. Amer. Meteor. Soc., 793 pp.

    • Crossref
    • Export Citation
  • Rigo, T., N. Pineda, and J. Bech, 2010: Analysis of warm season thunderstorms using an object-oriented tracking method based on radar and total lightning data. Nat. Hazards Earth Syst. Sci., 10, 18811893, https://doi.org/10.5194/nhess-10-1881-2010.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ronneberger, O., P. Fischer, and T. Brox, 2015: U-Net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention, Vol. 9351, Springer, 234–241.

  • Salamon, J., and J. P. Bello, 2017: Deep convolutional neural networks and data augmentation for environmental sound classification. IEEE Signal Process. Lett., 24, 279283, https://doi.org/10.1109/LSP.2017.2657381.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sánchez, J. L., J. L. Marcos, M. T. De La Fuente, and A. Castro, 1998: A logistic regression model applied to short term forecast of hail risk. Phys. Chem. Earth, 23, 645648, https://doi.org/10.1016/S0079-1946(98)00102-5.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sánchez, J. L., E. García Ortega, and J. L. Marcos, 2001: Construction and assessment of a logistic regression model applied to short-term forecasting of thunderstorms in León (Spain). Atmos. Res., 56, 5771, https://doi.org/10.1016/S0169-8095(00)00089-2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schmidhuber, J., 2015: Deep learning in neural networks: An overview. Neural Networks, 61, 85117, https://doi.org/10.1016/j.neunet.2014.09.003.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schultz, C. J., L. D. Carey, E. V. Schultz, and R. J. Blakeslee, 2017: Kinematic and microphysical significance of lightning jumps versus nonjump increases in total flash rate. Wea. Forecasting, 32, 275288, https://doi.org/10.1175/WAF-D-15-0175.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Shelhamer, E., J. Long, and T. Darrell, 2017: Fully convolutional networks for semantic segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39, 640651, https://doi.org/10.1109/TPAMI.2016.2572683.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Shi, X., Z. Chen, H. Wang, D.-Y. Yeung, W. Wong, and W. Woo, 2015: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Proc. Advances in Neural Information Processing Systems, Montreal, QC, Canada, Curran Associates, 802–810, http://papers.nips.cc/paper/5955-convolutional-lstm-network-a-machine-learning-approach-for-precipitation-nowcasting.pdf.

  • Shi, X., Z. Gao, L. Lausen, H. Wang, D. Y. Yeung, W. K. Wong, and W. C. Woo, 2017: Deep learning for precipitation nowcasting: A benchmark and a new model. 31st Conf. on Neural Information Processing Systems, Long Beach, CA, Neural Information Processing Systems, 5617–5627.

  • Simonyan, K., and A. Zisserman, 2014: Very deep convolutional networks for large-scale image recognition. Third Int. Conf. on Learning Representations, San Diego, CA, International Machine Learning Society, https://arxiv.org/abs/1409.1556.

  • Stensrud, D. J., and et al. , 2013: Progress and challenges with warn-on-forecast. Atmos. Res., 123, 216, https://doi.org/10.1016/j.atmosres.2012.04.004.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sun, J., and et al. , 2014: Use of NWP for nowcasting convective precipitation: Recent progress and challenges. Bull. Amer. Meteor. Soc., 95, 409426, https://doi.org/10.1175/BAMS-D-11-00263.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Vincent, B. R., L. D. Carey, D. Schneider, K. Keeter, and R. Gonski, 2003: Using WSR-88D reflectivity for the prediction of cloud-to-ground lightning: A central North Carolina study. Natl. Wea. Dig., 27, 3544.

    • Search Google Scholar
    • Export Citation
  • Walker, J. R., W. M. Mackenzie, J. R. Mecikalski, and C. P. Jewett, 2012: An enhanced geostationary satellite-based convective initiation algorithm for 0–2-h nowcasting with object tracking. J. Appl. Meteor. Climatol., 51, 19311949, https://doi.org/10.1175/JAMC-D-11-0246.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wang, Y., M. Long, J. Wang, Z. Gao, and P. S. Yu, 2017: PredRNN: Recurrent neural networks for predictive learning using spatiotemporal LSTMs. 31st Conf. on Neural Information Processing Systems, Long Beach, CA, Neural Information Processing Systems, https://papers.nips.cc/paper/6689-predrnn-recurrent-neural-networks-for-predictive-learning-using-spatiotemporal-lstms.pdf.

  • Wilson, J. W., N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, 1998: Nowcasting thunderstorms: A status report. Bull. Amer. Meteor. Soc., 79, 20792099, https://doi.org/10.1175/1520-0477(1998)079<2079:NTASR>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilson, J. W., Y. Feng, M. Chen, and R. D. Roberts, 2010: Nowcasting challenges during the Beijing Olympics: Successes, failures, and implications for future nowcasting systems. Wea. Forecasting, 25, 16911714, https://doi.org/10.1175/2010WAF2222417.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Woo, W. C., and W. K. Wong, 2017: Operational application of optical flow techniques to radar-based rainfall nowcasting. Atmosphere, 8, 48, https://doi.org/10.3390/atmos8030048.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Xia, R., D. L. Zhang, and B. Wang, 2015: A 6-yr cloud-to-ground lightning climatology and its relationship to rainfall over central and eastern China. J. Appl. Meteor. Climatol., 54, 24432460, https://doi.org/10.1175/JAMC-D-15-0029.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yang, J., Z. Zhang, C. Wei, F. Lu, and Q. Guo, 2017: Introducing the new generation of Chinese geostationary weather satellites, Fengyun-4. Bull. Amer. Meteor. Soc., 98, 16371658, https://doi.org/10.1175/BAMS-D-16-0065.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yang, X., J. Sun, and W. Li, 2015: An analysis of cloud-to-ground lightning in China during 2010–13. Wea. Forecasting, 30, 15371550, https://doi.org/10.1175/WAF-D-14-00132.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yangli-Ao, G., and et al. , 2019: LightNet: A dual spatiotemporal encoder network model for lightning prediction. 25th Conf. on Knowledge Discovery and Data Mining, Anchorage, AK, Special Interest Group on Knowledge Discovery in Data, 2439–2447, https://doi.org/10.1145/3292500.3330717.

    • Crossref
    • Export Citation
  • Yumimoto, K., and et al. , 2016: Aerosol data assimilation using data from Himawari-8, a next-generation geostationary meteorological satellite. Geophys. Res. Lett., 43, 58865894, https://doi.org/10.1002/2016GL069298.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zhang, J., 2010: Multi-source remote sensing data fusion: Status and trends. Int. J. Image Data Fusion, 1, 524, https://doi.org/10.1080/19479830903561035.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zhang, W., L. Han, J. Sun, H. Guo, and J. Dai, 2017: Application of multi-channel 3D-cube successive convolution network for convective storm nowcasting. 2019 Int. Conf. on Big Data, Los Angeles, CA, IEEE, 1705–1710, https://doi.org/10.1109/BigData47090.2019.9005568.

    • Crossref
    • Export Citation
  • Zheng, Y. G., J. Chen, and P. J. Zhu, 2008: Climatological distribution and diurnal variation of mesoscale convective systems over China and its vicinity during summer. Chin. Sci. Bull., 53, 15741586, https://doi.org/10.1007/s11434-008-0116-9.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Distribution of topography (shaded; in units of km), NLDN sensors (red points), and radar stations (black triangles). The white rectangle denotes the selected area (20°–40°N, 100°–120°E).

  • View in gallery

    Flowchart of the lightning nowcasting algorithm with deep learning using multisource observation data.

  • View in gallery

    Illustration of the LightningNet architecture, which is a fully convolutional network with an encoder–decoder architecture. It consists of three-dimensional convolutional layers and normalization layers (blue), pooling layers (green), upsampling layers (red), and a softmax classifier (yellow). The output of LightningNet is an image with the class (lightning or not lightning) probability for each pixel.

  • View in gallery

    ROC and AUC of 0–1-h lightning nowcasts produced by LightningNet with different data types for August 2018. True positive rate = h/(h + m) and false positive rate = f/(c + f).

  • View in gallery

    The 0–1-h lightning nowcasting results from LightningNet and the corresponding observations from NLDN on 23 Aug 2018 (the probabilistic predictions are shaded, and black dots represent lightning observations; the skill scores are given in the top right corner of each subfigure).

  • View in gallery

    As in Fig. 5, but for 4 Aug 2018.

  • View in gallery

    Lightning predictions within the next 60 min at 1100 BJT 1 Aug 2018 [the lightning observations and lightning predictions are the same in (a)–(c); contours show the lightning probabilistic nowcasting, and the black dots denote lightning observations from 1100 to 1200 BJT, while the color-shaded areas are radar reflectivity at (a) 1100, (b) 1130, and (c) 1200].

All Time Past Year Past 30 Days
Abstract Views 358 358 45
Full Text Views 116 116 39
PDF Downloads 157 157 50

A Deep Learning Network for Cloud-to-Ground Lightning Nowcasting with Multisource Data

View More View Less
  • 1 Chinese Academy of Meteorological Science, and University of Chinese Academy of Science, and National Meteorological Center, Beijing, China
  • | 2 National Meteorological Center, Beijing, China
  • | 3 Chinese Academy of Meteorological Science, Beijing, China
  • | 4 National Meteorological Center, Beijing, China
© Get Permissions
Free access

Abstract

Precise and timely lightning nowcasting is still a great challenge for meteorologists. In this study, a new semantic segmentation deep learning network for cloud-to-ground (CG) lightning nowcasting, named LightningNet, has been developed. This network is based on multisource observation data, including data from a geostationary meteorological satellite, Doppler weather radar network, and CG lightning location system. LightningNet, with an encoder–decoder architecture, consists of 20 three-dimensional convolutional layers, pooling and upsampling layers, normalization layers, and a softmax classifier. The central–eastern and southern China was selected as the study area, with considerations given to the topography and spatial coverage of the weather radar and lightning observation networks. Brightness temperatures (TB) of six infrared bands from the Himawari-8 satellite, composite reflectivity mosaic, and CG lightning densities were used as the predictors because of their close relationships with lightning activity. The multisource data were first interpolated into a uniform spatial/temporal resolution of 0.05° × 0.05°/10 min, and then training and test datasets were constructed, respectively. LightningNet was trained to extract the features of lightning initiation, development, and dissipation. The evaluation results demonstrated that LightningNet was able to achieve good performance of 0–1-h lightning nowcasts using the multisource data. The probability of detection, the false alarm ratio, the area under relative operating characteristic curve, and the threat score (TS) of LightningNet with all three types of data reached 0.633, 0.386, 0.931, and 0.453, respectively. Because geostationary meteorological satellite and radar both possess the capability of capturing lightning initiation (LI) features, LightningNet also showed good performance in LI nowcasting. When all three types of data were used, more than 50% LI was predicted accurately and the TS exceeded 0.36. LightningNet’s nowcast performance using triple-source data was clearly superior to that using only single-source or dual-source data, and these findings indicate that LightningNet has good capability of combining multisource data effectively to produce more reliable lightning nowcasts.

© 2020 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Kanghui Zhou, zhoukh@cma.gov.cn

Abstract

Precise and timely lightning nowcasting is still a great challenge for meteorologists. In this study, a new semantic segmentation deep learning network for cloud-to-ground (CG) lightning nowcasting, named LightningNet, has been developed. This network is based on multisource observation data, including data from a geostationary meteorological satellite, Doppler weather radar network, and CG lightning location system. LightningNet, with an encoder–decoder architecture, consists of 20 three-dimensional convolutional layers, pooling and upsampling layers, normalization layers, and a softmax classifier. The central–eastern and southern China was selected as the study area, with considerations given to the topography and spatial coverage of the weather radar and lightning observation networks. Brightness temperatures (TB) of six infrared bands from the Himawari-8 satellite, composite reflectivity mosaic, and CG lightning densities were used as the predictors because of their close relationships with lightning activity. The multisource data were first interpolated into a uniform spatial/temporal resolution of 0.05° × 0.05°/10 min, and then training and test datasets were constructed, respectively. LightningNet was trained to extract the features of lightning initiation, development, and dissipation. The evaluation results demonstrated that LightningNet was able to achieve good performance of 0–1-h lightning nowcasts using the multisource data. The probability of detection, the false alarm ratio, the area under relative operating characteristic curve, and the threat score (TS) of LightningNet with all three types of data reached 0.633, 0.386, 0.931, and 0.453, respectively. Because geostationary meteorological satellite and radar both possess the capability of capturing lightning initiation (LI) features, LightningNet also showed good performance in LI nowcasting. When all three types of data were used, more than 50% LI was predicted accurately and the TS exceeded 0.36. LightningNet’s nowcast performance using triple-source data was clearly superior to that using only single-source or dual-source data, and these findings indicate that LightningNet has good capability of combining multisource data effectively to produce more reliable lightning nowcasts.

© 2020 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Kanghui Zhou, zhoukh@cma.gov.cn

1. Introduction

Lightning is a sudden electrostatic discharge that occurs during a thunderstorm, and it is characterized by both mesoscale and microscale characteristics, rapid evolution, and complicated interactions with the surrounding atmosphere. Lightning often accompanies severe weather phenomena such as hail, tornadoes, high winds, and heavy rainfall, and it poses a serious threat to life and property. Therefore, precise and timely lightning forecasts are needed. However, it is still a great challenge for meteorologists to predict lightning precisely in operational meteorology today (Ray 1986; Gatlin and Goodman 2010; Lynn et al. 2012; Stensrud et al. 2013; Sun et al. 2014; Schultz et al. 2017; Farnell et al. 2017).

From a physical perspective, charge separation before lightning happens most likely occurs during rebounding collisions between ice crystals and large ice hydrometeors (such as graupel and hail) that remain suspended in the mixed phase zone by the updraft of a growing thunderstorm (Vincent et al. 2003). Radar reflectivity data can be used to indirectly identify the electrification process within a developing thunderstorm because graupel and hail particles return large reflectivity echoes (Buechler and Goodman 1990; Vincent et al. 2003; Mosier et al. 2011). Hence, large reflectivity has been widely used as an indicator of the onset of cloud-to-ground (CG) lightning in studies of nowcasting (e.g., Mecikalski et al. 2013). The appearance of a large reflectivity (e.g., 30–40 dBZ) at temperatures below a certain threshold (e.g., from 0° to −20°C) was usually used to nowcast lightning (Buechler and Goodman 1990; Vincent et al. 2003; Mosier et al. 2011). Meanwhile, geostationary meteorological satellite is able to offer spectral information to identify physical processes. For instance, cloud-top glaciation is an important proxy indicator for places where the noninductive charging process and significant updrafts may be occurring in cumulus clouds (Mecikalski et al. 2013).

In the past few decades, various extrapolation methods based on high-resolution observation data from satellites, radar, or lightning location systems have been widely used to track and predict the movements of convective systems, and these are rather efficient and effective in terms of operation (Wilson et al. 1998, 2010).

Examples of extrapolation algorithms based on radar data include Thunderstorm Identification, Tracking, Analysis, and Nowcasting (TITAN) (Dixon and Wiener 1993), Storm Cell Identification and Tracking (SCIT) (Johnson et al. 1998), the optical flow method (Bechini and Chandrasekar 2017; Woo and Wong 2017), and the machine learning method (Wang et al. 2017). However, not all precipitation in radar echoes produces lightning, so it is obvious that radar echo extrapolation is not equivalent to lightning nowcasting.

Radar data have also been used as input to various machine learning techniques to monitor and/or to nowcast convective storms. Haberlie and Ashley (2018) applied different machine learning algorithms, for example, random forest, gradient boosting, to identify mesoscale convective systems (MCSs) in radar reflectivity images. Their results suggested that the algorithms can distinguish between MCS and non-MCS samples with a high probability of detection and a low probability of false alarms. Medina et al. (2019) used random forest (RF) method to identify thunderstorms’ downburst with eight dual-polarization radar signatures that have been hypothesized to possess physical implications for downburst.

Meanwhile, there are also thunderstorm tracking and extrapolation algorithms with lightning location data, which can be used to identify the locations of thunderstorms and then track and predict their movements and activity areas (Kohn et al. 2011; Betz et al. 2008). Furthermore, both radar and lightning data have been combined to nowcast thunderstorms, and these approaches showed better performances than those that used only a single type of data (Bonelli and Marcacci 2008; Rigo et al. 2010; Metzger and Nuss 2013).

Although the extrapolation methods are able to predict the movement trend of a storm for the next 0–2 h, they do not consider the physical laws of initiation, development, and dissipation of the storm, and thus, it is difficult for them to predict the evolution precisely. Besides, if only an extrapolation method is used, many isolated thunderstorms with a life cycle shorter than 1 h could be predicted as a false alarm when the prediction time is beyond their lifetime.

Geostationary meteorological satellite has become an effective tool for monitoring convective systems due to the great progress that has been made in the past few decades in terms of their capability to provide remote sensing images over a vast area with high spatial (~kilometers) and temporal (~minutes) resolutions (Bessho et al. 2016; Yang et al. 2017). Furthermore, the satellite has unique advantages in monitoring the characteristics of the initial stages of convective clouds, such as rapid increases in cloud thickness, sharp drops in cloud-top temperature, and transformations in the cloud-top hydrometeor phase (Mecikalski and Bedka 2006; Mecikalski et al. 2010). Therefore, satellite data can be used for convection initiation (CI) monitoring and its early warning, and researchers have carried out much work and have developed a number of methods and products for CI nowcasting based on satellite data (Mecikalski and Bedka 2006; Walker et al. 2012; Lee et al. 2017).

Until recently, traditional methods of CI nowcasting have relied on the knowledge and experience of researchers to design the primary “interest fields (IFs)” according to the physical characteristics of CI reflected in satellite data. Those approaches emphasize the utilization of critical thresholds in one or more IFs derived from satellite data (Harris et al. 2010; Karagiannidis et al. 2016). Of course, satellite data can be combined with other types of data such as numerical weather prediction (NWP) data to improve the CI nowcasting algorithms (Mecikalski et al. 2015).

When a geostationary meteorological satellite observes the development of CI, it can only obtain the characteristics of the cloud tops, such as the albedo and brightness temperatures (TB), but it is unable to detect the internal structure of the clouds. Fortunately, ground-based weather radar is capable of obtaining the height of ice crystals and large ice hydrometeors indicating charge separation during rebounding collisions. For instance, Mecikalski et al. (2013) found that radar echoes of 20–30 dBZ at the height of 2–4 km formed 30–60 min before lightning initiation (LI; defined as the time of the first lightning, of any kind, generated in a cumulonimbus cloud) in Oklahoma and Florida. Additionally, real-time observations of lightning are certain to provide important information for their nowcasting. Therefore, effective integration of the above-mentioned multisource data is expected to produce much better lightning nowcasts.

However, the effective utilization of a multisource dataset to its full potential is still a difficult task (Zhang 2010). On the one hand, data from satellite, radar, and other mesoscale observing networks are becoming increasingly available at an astonishing speed, which makes it impossible to manually extract the useful information for lightning nowcasting (Dabberdt et al. 2005; Hou et al. 2014). On the other hand, different observations differ significantly in the physical meanings, and thus at present, experienced meteorologists are still needed to explain and extract the relevant data. In addition, there may be some undiscovered features of each type of observation data for nowcasting lightning.

Machine learning techniques provide new possibilities for multisource data integrated applications in nowcasting convective activities. Sánchez et al. (1998, 2001) used logistic regression models to predict hail and thunderstorms based on radiosonde data, and their results showed that the forecasting models had a high probability of detection (POD) and a low false alarm ratio (FAR). Mecikalski et al. (2015) combined Geostationary Operational Environmental Satellite (GOES) data with NWP data using the RF and logistic regression (LR) methods, and they extracted 25 related elements (variables) for CI and achieved CI nowcasting results for North America that showed a better performance than that of satellite IFs used alone. Han et al. (2017) developed a method that automatically extracted the occurrence and development of thunderstorms by applying the support vector machine (SVM) algorithm to high-resolution reanalysis fields, and they obtained encouraging performance in CI nowcasting. Ahijevych et al. (2016) also employed the RF method to generate 2-h forecasts of the likelihood for initiation of MCS based on radar, satellite and NWP data. RF method in the study showed a good ability to combine the multisource data, and over 99% of the 550 observed initiation of MCS events were detected within 50 km.

Deep learning (DL) is a subset of machine learning algorithms that uses multilayer artificial neural networks to deliver state-of-the-art accuracy in many tasks (Bengio 2009; Schmidhuber 2015). Similar to traditional machine learning algorithms like artificial neural networks and SVM, DL networks can model complex nonlinear systems. Moreover, these networks perform better in extracting the advantageous features with deeper layers. DL has already been used successfully in a wide range of applications including computer vision, facial recognition, and medical diagnosis tasks, and it has yielded results comparable to and in some cases superior to those produced by human experts (LeCun et al. 2015; Schmidhuber 2015).

Preliminarily, DL has also been applied in nowcasting convective systems. Shi et al. (2015, 2017) and Wang et al. (2017) proposed DL networks that automatically learn the features of radar echo evolution of convective systems. Their results showed that their DL networks could predict the evolution of convective systems successfully, and the overall performance was superior to that of traditional extrapolation methods. Zhang et al. (2017) based on radar and analysis fields, constructed a deep convolutional network to achieve effective nowcasts of convective storms (radar reflectivity ≥ 35 dBZ), and also to predict the CI and evolution of convective systems. Yangli-Ao et al. (2019) used NWP data to predict lightning within the next six hours with DL method. Their DL network with dual encoders was used to extract spatiotemporal features of Weather Research and Forecasting (WRF) Model simulation data and recent lightning observations. Their results showed that DL achieved a threefold improvement in equitable threat score for the 6-h prediction compared with three other established forecasting methods.

Lightning nowcasting can be considered as a semantic pixel-wise segmentation task. Based on “images” consisting of satellite multichannel, radar, and lightning observation data, we used a DL network to segment the lightning and nonlightning area. Semantic pixel-wise segmentation is an active topic of research, and various DL networks have been developed, such as the fully convolutional network (FCN) (Shelhamer et al. 2017), U-Net (Ronneberger et al. 2015), SegNet (Badrinarayanan et al. 2017), and Deeplab (Chen et al. 2018). These networks have been applied to automatic driving, image search engines, cancer detection, and so forth (Garcia-Garcia et al. 2017).

The present work is devoted to the application of DL using multisource data to nowcast cloud-to-ground lightning (lightning hereafter) for the next whole hour. To the best of our knowledge, this is the first time that DL has been used to integrate multisource observation data in lightning nowcasting and to extract LI features.

The paper is organized as follows. Section 2 presents the data used and the study area, and section 3 introduces the methodology utilized to develop the new DL network for lightning nowcasting. Section 4 illustrates the verification results. Finally, section 5 summarizes the conclusions of this work and discusses the points deserving of further research.

2. Data

a. Satellite data

Himawari-8 is a new-generation geostationary meteorological satellite launched by Japan in October 2014, and it began operations in July 2015. Himawari-8 carries an Advanced Himawari Imager (AHI) instrument that observes three visible bands, two near-infrared bands, and ten infrared bands. Its observations have high spatial and temporal resolutions. The spatial resolution of visible band data reaches 0.5 km, and that of the infrared-band data reaches 2 km. The refresh time interval of its full disk observations is 10 min. In addition to the level 1 (L1) products, the Japan Meteorological Agency also provides abundant L2 and L3 products, including cloud phase, cloud type, and cloud-top height data (Yumimoto et al. 2016).

The L2 and L3 products have been developed based on the L1 products, and DL is good at feature extraction, in which it can directly extract the necessary convective features from the L1 products. Considering these aspects, we selected only the TB data from six infrared bands of Himawari-8 AHI (Table 1), from which we can capture not only convective systems, but also rapidly developing cumulus (Bessho et al. 2016). Note that since it is only available during daytime, the visible band data were abandoned in this study.

Table 1.

Selected satellite, radar, and lightning predictors for lightning nowcasting.

Table 1.

In this study, we downloaded the Himawari-8 L1 gridded data (netCDF4 format) with a resolution of 0.05° latitude × 0.05° longitude (the distance of 0.05° latitude is 5.55 km, while that of 0.05° in longitude is varies from 4.25 km at 20°N to 5.20 km at 40°N.) from the Japan Aerospace Exploration Agency’s (JAXA) P-Tree system.

b. Radar data

For operational consideration, we use the composite maximum reflectivity mosaic data provided by the Meteorological Observation Center of the China Meteorological Administration (CMA). It has a horizontal resolution of 0.01° latitude × 0.01° longitude at 6-min intervals and covers most of China. The data can be acquired from the internet website of CMA (http://data.cma.cn).

The number of radars in our study area is 144. The detection range of each radar is 150 km (C band) or 230 km (S band). The minimum, median, mean and maximum radar separation distances in the study area are 50.46, 112.43, 113.66, and 244.04 km, respectively. All the radars scan under the radar volume coverage pattern 21 (VCP21) mode, with an interval of approximately six minutes. The data quality control method consistent with Chen et al. (2012) is applied to remove abnormal propagation, ground clutter, and particle clutter.

In this study, the temporal resolution of all of the data was unified to 10 min, and we used linear interpolation in time to obtain the new radar reflectivity data sequences at 10-min intervals.

c. Lightning data

The lightning location data used in this study were obtained from the National Lightning Detection Network (NLDN) of China. This network is equipped with ground-based advanced time of arrival and direction systems CG lightning detection sensors, and 394 sensors were in operation in 2016. These sensors cover most of China except for the Tibetan Plateau and the western part of inner Mongolia. The location accuracy of NLDN is approximately 0.5 km, and its detection efficiency is larger than 80%; the average detected radius of a sensor is approximately 300 km (Xia et al. 2015; Yang et al. 2015).

Lightning location data are a type of discrete point data (which can also be acquired from http://data.cma.cn). The data were gridded by using the nearest interpolation method. If a lightning event was observed within a radius, R, of a grid point, then the number of lightning in this grid point added by 1, which indicates that the event occurred at this grid point. Otherwise, the grid point was marked by 0. Lightning data would be used both as predictor and label in our DL network. Note that when the lightning was used as predictor, the lightning density (i.e., the cumulative number of each grid point) was utilized. On the other hand, when the lightning was used as label, each grid was only marked 0 or 1 (the “1” indicating lightning event occurrence).

Considering that lightning typically occurs on a meso-γ scale (Orlanski 1975), R was set to 20 km in this study. Note that if R is set too small, for the lightning nowcasts produced by LightningNet, there will be too many missing forecasts, while if R is too large, there will be too many false alarms.

d. Study area

An area (20°–40°N, 100°–120°E) was selected for this study (Fig. 1), with good coverage in terms of radar and lightning observations in China, where convective systems frequently occur in summer (Zheng et al. 2008).

Fig. 1.
Fig. 1.

Distribution of topography (shaded; in units of km), NLDN sensors (red points), and radar stations (black triangles). The white rectangle denotes the selected area (20°–40°N, 100°–120°E).

Citation: Journal of Atmospheric and Oceanic Technology 37, 5; 10.1175/JTECH-D-19-0146.1

3. Methodology

a. Lightning nowcasting process

Development of the lightning nowcasting algorithm with DL consisted of three steps (Fig. 2). First, a dataset, including a training subset and a test subset, was constructed. Second, the new DL network, called LightningNet, was designed, trained, and tested with the dataset. Finally, LightningNet with optimal weights was implemented for lightning nowcasting with multisource data.

Fig. 2.
Fig. 2.

Flowchart of the lightning nowcasting algorithm with deep learning using multisource observation data.

Citation: Journal of Atmospheric and Oceanic Technology 37, 5; 10.1175/JTECH-D-19-0146.1

b. Construction of training/testing sets

The multisource data types used as input for LightningNet are listed in Table 1. The temporal and spatial resolutions of the various types of data were very different. Hence, it was necessary to unify the resolutions of the various types of data first. All of the data were interpolated to temporal resolution of 10-min interval and to spatial resolutions of 0.05° latitude × 0.05° longitude.

To enable LightningNet to fully extract the temporal evolution features of convective systems, an image sequence with time steps Nt must be included in each sample. After testing, we found that when Nt was taken as three, a good balance was achieved between the prediction performance and computational efficiency. Therefore, the multisource data of the past 30 min were used to nowcast lightning activity within the next 0–60 min in this network.

The spatial size of satellite, radar, and lightning grid data covering the study area, with a horizontal resolution of 0.05° × 0.05°, was 400 × 400. Thus, the training sample was a matrix of Nt × 400 × 400 × Np (Np is the number of predictors), which was labeled with lightning observations. The labeled matrix, with a size of 400 × 400 pixels, was derived from lightning activities in the next 60 min. For each point of the labeled matrix, if lightning was detected less than 20 km from the grid point, that grid was marked as 1, and otherwise, it was marked as 0.

In this study, we selected 1530 samples from May to August in 2017 and 2018. We split the dataset into the following two independent subsets: a test subset and a training subset. A validation set, which was 20% of the training set, was randomly divided during the training. The test subset contained all of the samples for August 2018 (274 samples), whereas the training set included the remaining samples (1256 samples).

LightningNet, which is a DL network, has many trainable parameters, which is 29 128 577. Thus, large numbers of training samples are required to avoid overfitting. Consequently, training dataset augmentation was necessary (Salamon and Bello 2017; Perol et al. 2018). Four strategies, namely, horizontal flip, 90° rotation, 180° rotation, and 270° rotation, were applied to generate new samples. Specifically, we flipped or rotated the predictor array as well as the corresponding labeled array in terms of the spatial dimension to generate a new sample. With the above four strategies, fourfold transformations of the original samples were generated. Finally, we obtained the final training set with 6280 samples (i.e., 1256 × 5).

c. DL network architecture

Most of the existing semantic segmentation networks have been designed to segment images (Garcia-Garcia et al. 2017). However, in our study, we had to extract the temporal and spatial developing features of convective systems from image sequences with a size of Nt × height × width × Np, where Nt is the length of image sequence and Np is the number of predictors. Thus, on the basis of SegNet, we developed a new architecture—LightningNet—that is able to extract the temporal and spatial features of multisource data.

SegNet has an encoder–decoder architecture based on the two-dimensional (2D) convolutional layers of the Visual Geometry Group-16 network (Simonyan and Zisserman 2014). Compared with SegNet, we replaced the 2D convolutional layers with three-dimensional (3D) convolutional (Conv3D) layers, which have been found to be competent at feature extraction in temporal and spatial dimensions (Ji et al. 2013).

Similar to SegNet, LightningNet has also an encoder network and a corresponding decoder network, followed by a final pixel wise classification layer (Fig. 3). There are 26 convolutional layers in SegNet, 13 layers in the encoder and another 13 in the decoder. However, in LightningNet, because the size of input images was relatively small, the encoder and decoder networks consist of 10 Conv3D layers, respectively.

Fig. 3.
Fig. 3.

Illustration of the LightningNet architecture, which is a fully convolutional network with an encoder–decoder architecture. It consists of three-dimensional convolutional layers and normalization layers (blue), pooling layers (green), upsampling layers (red), and a softmax classifier (yellow). The output of LightningNet is an image with the class (lightning or not lightning) probability for each pixel.

Citation: Journal of Atmospheric and Oceanic Technology 37, 5; 10.1175/JTECH-D-19-0146.1

In the same manner as SegNet, feature maps produced by the convolutional layers were then batch normalized to accelerate the training process. We normalized the activations of the previous layer at each batch, that is, applied a transformation that maintained the mean activation close to 0 and the activation standard deviation near 1 (Ioffe and Szegedy 2015). Following that step was the construction of the max-pooling layer. In a manner different from SegNet, we replaced the 2D max-pooling layer with the 3D max-pooling layer in LightningNet. The final decoder output of LightningNet was fed to a binary-class softmax classifier to produce class probabilities for each pixel independently.

1) Convolutional-3D layers

Each channel of the 20 Conv3D layers was obtained by convolving the channels of the previous layer with a bank of linear 3D filters, such as ones for summing, adding a bias (this is a concept internal to the LightningNet network) term, and applying pointwise nonlinearity (Perol et al. 2018).

We represent the input of layer n as a set of K matrices Xn1(k), with k = 1, …, K (K is the number of convolution filters of each layer). Each of these matrices represents a different input feature map. A feature map Xn(l) can be obtained by computing the sum of k convolutions of the feature maps from the previous layer. The output feature maps Xn(l),l=1,,L (L is the number of convolutional layers, which is 20 in LightningNet) are represented as follows:
Xn(l)=ReLU[bn(l)+k=1KWn(k,l)*Xn1(k)]forl{1,,L},
where ReLU(x) = max(0, x) is the rectified linear unit activation function. The asterisk denotes the 3D convolution operation, the matrices Wn(k,l) represent the filters of layer n, and bn(l) is the bias for feature map l.

Each encoder in the encoder network performs convolution with a filter bank to produce a set of feature maps. A batch normalization layer, accelerating the deep network training process, is then applied to each of these maps (Badrinarayanan et al. 2017).

2) Encoder and decoder

LightningNet has an encoder–decoder architecture. The encoder is a succession of Conv3D layers followed by batch normalization and ReLU(x). The activations of the previous layer in each batch are normalized. Blocks of convolution are followed by a pooling layer. The decoder has the same number of convolutions and the same number of blocks. In place of pooling, the decoder performs upsampling using unpooling layers. The mirrored decoder in the decoder network upsamples its input feature map(s) by using the memorized max-pooling indices from their corresponding encoder feature map(s) (Audebert et al. 2016).

3) Pooling layers and upsampling layers

A 3D max-pooling layer takes a group of neighbors in the feature map and condenses them into a single output by computing the maximum of all incoming activations in the window. The 3D max pooling is used to achieve translation invariance over small spatial shifts in the input array.

Upsampling layer operates by relocating at the maximum index computed by the associated pooling layer. For example, the first pooling layer computes the mask of the maximum activations and passes it to the last unpooling layer, which will upsample the feature map to a full resolution by placing the activations on the mask indices and zeroes everywhere else (Audebert et al. 2016).

4) Output layer

The high-dimensional feature representation at the output of the final decoder is fed to a trainable softmax classifier. We applied the softmax function to the class scores to obtain a properly normalized probability distribution, which can be interpreted as a posterior distribution over the classes conditioned on the input X(0) and the network parameters W and b:
pc=P[class=c|X(0),W,b]=exp(xc)k=01exp(xk),wherec={0,1},
W = {W1, …, W20} is the set of all of the weights, and b = {b1, …, b20} is the set of all of the biases, and c is the number of classes (Perol et al. 2018). The xc and xk are the prediction vectors. Here we have 2 classes, 0 and 1, representing no lightning and lightning, respectively.

This softmax classifier predicts each pixel independently. The output of the softmax classifier is a two-channel image of probabilities. The predicted segmentation corresponds to the class with maximum probability at each pixel.

5) LightningNet training

We optimized LightningNet parameters by minimizing the loss function on a patch of N pixels, that is, 400 × 400 pixels in the study area, and let yi denote its label on pixel i, and c is the number of classes:
loss=1Ni=1Nc=01ycilog[pc(i)],wherec={0,1}.

We used the cross-entropy loss (Golik et al. 2013) as the objective function for training LightningNet. For optimization, we utilized the ADAM algorithm (Kingma and Ba 2015), and the learning rate was set to be 10−4; all other parameters were kept at the default value (Perol et al. 2018). The number of epochs, patch size, and iterations is 30, 2, 3140, respectively. Early stopping strategy was used during the training. When the loss on validation set no longer reduced in 5 epochs, the training process would terminate, and we saved the model weight with minimum loss on validation set.

Nvidia Compute Unified Device Architecture (CUDA) library and Nvidia TITAN graphics were used in the training and forecasting processes of LightningNet. The results showed that 1-h lightning nowcasting at a 0.05° latitude × 0.05° longitude resolution in the study area was completed within 2 min, which makes it feasible for operational applications.

4. Experimental results

Classical skill scores, including the POD, FAR, threat score (TS), accuracy, bias, and equitable threat score (ETS), are typically used for evaluating deterministic forecasts; however, they also can be used for evaluating the probabilistic forecast performance by thresholding the probabilistic forecasts and turning them into deterministic forecasts. After different thresholds of probability were tested, we determined that when the probabilistic threshold value was 0.5, we could obtain forecasts with the highest ETS, and TS, and the bias was closest to 1. Therefore, the value of 0.5 was taken as the threshold in the following evaluations.

a. Evaluation

The performance of LightningNet was assessed quantitatively by using an entire month of nowcasting results during August 2018.

In addition, in order to evaluate the results of different combinations of multisource data during the application of LightningNet, we trained LightningNet with the above training set by using various combinations of predictors from satellite, radar, and lightning data. The skill scores obtained are shown in Table 2.

Table 2.

Evaluation of lightning 0–1-h nowcasting results produced by LightningNet with different predictors and a traditional method for August 2018. The equations of POD, FAR, bias, accuracy, ETS, and TS are presented in the table, where h, m, f, and c indicate the hits, misses, false alarms, and correct negatives, respectively, and hrandom = (h + f) × (h + m)/(h + m + f + c).

Table 2.

We also evaluated a lightning nowcasting algorithm by utilizing only three satellite-derived IFs, namely, (i) TB10.8µm < −16°C, indicative of the cloud-top glaciation; (ii) TB10.8µm trend < −4°C (10 min)−1, indicative of the cloud depth; and (iii) TB6.2µmTB10.8µm > −25°C, indicative of the cloud growth rate, which were originally used with Meteosat Second Generation satellite imagery in the study of Karagiannidis et al. (2016). Because TB10.8µm data of Himawari-8 were not available, we replaced the TB10.8µm data with the TB10.4µm data. Meanwhile, we adjusted the threshold values by enumeration to determine the values with the highest TS and ETS in evaluations of lightning nowcasting. Finally, we evaluated the new combination of (i) TB10.4µm < −35°C, (ii) TB10.4µm trend < −5°C (10 min)−1, and (iii) TB6.2µmTB10.4µm > −7°C. In agreement with the results presented by Karagiannidis et al. (2016), we also found that when at least two IFs give a positive estimation, the performance is best. Thus, we present the evaluation results when more than two IFs criteria were satisfied below.

As shown in Table 2, lightning nowcasts from LightningNet showed significantly better performance than that of IFs in terms of all six evaluation indices. The more predictors LightningNet used, the better performance became. And more specifically, the method with two types of predictors showed better performance than that with only one type of predictor, whereas the performance using all three types of predictors was the best.

We also show the receiver operating characteristic (ROC) in Fig. 4. The closer the ROC curve is to the upper left corner of the graph, the more skillful the method is. As the same with Table 2, we can see that the performance of LightningNet with triple-source data is clearly better than that of with dual-source or single-source data. The area under curve (AUC) is also consistent with Table 2, showing the AUC of LightningNet with multisource data is the highest, reaching 0.931.

Fig. 4.
Fig. 4.

ROC and AUC of 0–1-h lightning nowcasts produced by LightningNet with different data types for August 2018. True positive rate = h/(h + m) and false positive rate = f/(c + f).

Citation: Journal of Atmospheric and Oceanic Technology 37, 5; 10.1175/JTECH-D-19-0146.1

The analyses above suggest that LightningNet is capable of effectively extracting the features of lightning activity from multisource data and generating much better nowcasting results than the traditional method. Furthermore, it is able to comprehensively apply the advantages of multisource data to produce more accurate nowcasting results of lightning activity. Note that the performance of LightningNet with satellite and lightning data is slightly better than that with radar and lightning data. This result can be closely related to the limited radar predictor number used in LightningNet, because some other radar predictors, such as vertically integrated ice (VII; Mosier et al. 2011), and 3D radar reflectivity were not used, which would likely lead to better performance, according to the studies of Vincent et al. (2003) and Mosier et al. (2011).

b. Case study—23 August 2018

On 23 August 2018, intense lightning activities occurred in South China. Convection started at about 1100 Beijing standard time (BJT), and then, it gradually strengthened and reached the most vigorous stage at 1600 BJT. After that, it began to weaken and dissipated at about 2300 BJT.

LightningNet’s 0–1-h nowcasts and lightning observations for this case are presented in Fig. 5. The results show that LightningNet gave an excellent performance in predicting lightning during the different development stages of convective storms, including the initiation, intensification, and dissipation stages. The average POD, FAR, bias, accuracy, ETS, and TS for this case amounted to 0.68, 0.38, 1.09, 0.96, 0.45, and 0.48, respectively. Note that higher probabilities of nowcasting correspond to denser lightning activities, which will be very useful in weather forecasting operations.

Fig. 5.
Fig. 5.

The 0–1-h lightning nowcasting results from LightningNet and the corresponding observations from NLDN on 23 Aug 2018 (the probabilistic predictions are shaded, and black dots represent lightning observations; the skill scores are given in the top right corner of each subfigure).

Citation: Journal of Atmospheric and Oceanic Technology 37, 5; 10.1175/JTECH-D-19-0146.1

The lightning nowcasting results in this case also demonstrate that the LightningNet method proposed in this study is capable of extracting the convective features from multisource data well, and it can produce reliable lightning nowcasts.

c. Case study—4 August 2018

A few thunderstorms occurred in North China on 4 August 2018, including a multicell thunderstorm in the north and two other small thunderstorms. We only present the predictions and lightning observations from 0000 to 1100 BJT 4 August 2018 in Fig. 6.

Fig. 6.
Fig. 6.

As in Fig. 5, but for 4 Aug 2018.

Citation: Journal of Atmospheric and Oceanic Technology 37, 5; 10.1175/JTECH-D-19-0146.1

Figure 6 shows that LightningNet had a good performance in prediction of the movement, merge and split of the multicell thunderstorm, showing the good capacity of predicting large-scale thunderstorm. For another two small thunderstorms, LightningNet also successfully predicted the initiation and dissipation of the thunderstorms, exhibiting the good nowcasting capability of small-scale thunderstorms. The average POD, FAR, bias, accuracy, ETS, and TS for this case amounted to 0.65, 0.27, 0.97, 0.97, 0.50, and 0.52, respectively.

d. Lightning initiation nowcasting and evaluation

Conventional nowcasting algorithms for convective systems, such as TITAN and optical flow algorithms, are based on the previous development of existing convective systems, and these methods predict their movement trends and activity areas by using a linear or nonlinear extrapolation method (Wilson et al. 1998). In comparison to conventional nowcasting algorithms, LightningNet, with its powerful feature extraction capability, is able to effectively extract CI features and then make good LI predictions. We selected another case to analyze and evaluate the LI nowcasting performance of LightningNet.

The selected convective weather case occurred in the southeastern part of China on 1 August 2018, in an area where thermal convection frequently generates in climatology (Zheng et al. 2008). As shown in Fig. 7, five isolated convective systems, labeled as A, B, C, D, and E, initiated or developed from 1100 to 1200 BJT.

Fig. 7.
Fig. 7.

Lightning predictions within the next 60 min at 1100 BJT 1 Aug 2018 [the lightning observations and lightning predictions are the same in (a)–(c); contours show the lightning probabilistic nowcasting, and the black dots denote lightning observations from 1100 to 1200 BJT, while the color-shaded areas are radar reflectivity at (a) 1100, (b) 1130, and (c) 1200].

Citation: Journal of Atmospheric and Oceanic Technology 37, 5; 10.1175/JTECH-D-19-0146.1

As we can see from the radar reflectivity image at 1100 BJT (Fig. 7a), convective systems C and E had already been initiated and were at their mature stage, while there was no indication that convective systems A, B, and D would form over the next 1 h. Radar reflectivities at 1130 BJT (Fig. 7b) showed that after half an hour, convective systems A and B formed and moved southward. Meanwhile, convective systems C and E were developing from single-cell systems to multicell systems, which led to significant increases in the convective area. There was still no indication in the radar reflectivity imagery of the initiation of convective system D. At 1200 BJT (Fig. 7c), convective systems A, B, C, and E continued to grow, and their regions of intense radar reflectivity were consistent with the lightning observations, in which data indicated that lightning might happen at that moment. We also found that convective system D was generated in the radar reflectivity image at 1200 BJT.

For the nowcasting performance of the five convective systems, LightningNet successfully predicted not only the movement of existing lightning activity, but also the LI in newborn convective systems, which was indicative of its good capability for LI nowcasting. In this case, we found that the lead time of LightningNet’s nowcasting had been more than 30 min. Both radar and satellite observations would provide information for LI nowcasting. According to the research of Mecikalski et al. (2013), radar observations can provide more information on the 10–20-min lightning nowcasting time scales, while satellite observations can provide more information beyond 20 min forecast period, especially in the 30–60 min time period.

To further evaluate the LI nowcasting capability of LightningNet, we also selected the time period of 24–30 August 2018, for analysis, and all of the LI areas were marked manually in the study area to present the evaluation results (Table 3).

Table 3.

Evaluation of LI nowcasting from LightningNet and one traditional method.

Table 3.

As shown in Table 3, LightningNet gave an excellent performance in LI nowcasting. When all three types of predictors were used, more than 50% LI was predicted accurately and the TS exceeded 0.36. Similar to the evaluation of nowcasting for all lightning, the LI nowcasting performance of LightningNet was much better than the traditional method of Karagiannidis et al. (2016). Similar to the results described in section 4a, when more predictors for LightningNet were used, the performance improved further. Furthermore, LightningNet achieved a better performance in LI nowcasting using satellite data than that using radar data, which was consistent with the previous study of Mecikalski et al. (2013). It is also possible that the radar data used were not applied optimally in this study due to lack of representativeness of radar data in some areas and/or poor choice of predictor variables.

5. Discussion and conclusions

Based on multisource data, we have developed a DL network for nowcasting lightning, named LightningNet.

LightningNet can effectively achieve 0–1-h lightning nowcasts. Both evaluations of selected cases and nowcasting results for an entire month showed that LightningNet has encouraging performance in lightning nowcasts, with the TS values exceeding 0.45 at the spatial resolution about 5 km (0.05° latitude × 0.05° longitude). Two nowcasting cases demonstrated that the higher the LightningNet prediction probability is, the higher the lightning activity will be, thus indicating the extra usefulness of the probabilistic nowcasts.

LightningNet possesses a strong feature learning ability and is able to effectively achieve the integration of multisource data. Therefore, compared to the traditional algorithm of Karagiannidis et al. (2016), LightningNet showed a much better performance for both lightning and LI nowcasting. As more data types are used with LightningNet, the performance improves further. The very strong feature learning capability of the DL network is likely the main reason why LightningNet can effectively achieve the comprehensive application of multisource data. Yet, additional details about the mechanisms need to be further explored in future. Furthermore, for operational consideration, the radar data used in the work are only the 2D composite reflectivity mosaic. The 3D radar reflectivity mosaic data, as well as other useful variables (e.g., vertically integrated ice) will be employed in future work.

This work is an attempt at achieving the comprehensive application of multisource data. DL has powerful feature extract capabilities, and so it will be a promising way for convective activity nowcasting.

Based on this work, we would like to combine more types of data, such as observations from automatic weather stations and high-resolution numerical weather prediction data, to further improve LightningNet in future. We will also try to obtain lightning density predictions that will provide more warning information. Furthermore, the currently implemented DL network is applicable for producing only 1-h nowcasts, and that implementing a DL network for producing nowcasts for any lead time between 0 and 2 h is a topic of future work.

Acknowledgments

This work was supported by the National Key R&D Program of China (Grants 2018YFC1507504 and 2017YFC1502003) and the National Natural Science Foundation of China (41875005). Research products consisting of Himawari-8 L1 gridded data used in this paper were supplied by the P-Tree System, Japan Aerospace Exploration Agency (JAXA). We also greatly thank the China Meteorological Administration (http://data.cma.cn) for providing the composite reflectivity mosaic data and cloud-to-ground lightning location data. Thanks are extended to the three anonymous reviewers for their greatly helpful and insightful suggestions.

REFERENCES

  • Ahijevych, D., J. O. Pinto, J. K. Williams, and M. Steiner, 2016: Probabilistic forecasts of mesoscale convective system initiation using the random forest data mining technique. Wea. Forecasting, 31, 581599, https://doi.org/10.1175/WAF-D-15-0113.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Audebert, N., B. Le Saux, and S. Lefèvre, 2016: Semantic segmentation of Earth observation data using multimodal and multi-scale deep networks. 13th Asian Conf. on Computer Vision, Taipei, Taiwan, Asian Federation of Computer Vision, 180–196.

    • Crossref
    • Export Citation
  • Badrinarayanan, V., A. Kendall, and R. Cipolla, 2017: SegNet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39, 24812495, https://doi.org/10.1109/TPAMI.2016.2644615.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bechini, R., and V. Chandrasekar, 2017: An enhanced optical flow technique for radar nowcasting of precipitation and winds. J. Atmos. Oceanic Technol., 34, 26372658, https://doi.org/10.1175/JTECH-D-17-0110.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bengio, Y., 2009: Learning deep architectures for AI. Found. Trends Mach. Learn., 2, 1127, https://doi.org/10.1561/2200000006.

  • Bessho, K., and et al. , 2016: An introduction to Himawari-8/9—Japan’s new-generation geostationary meteorological satellites. J. Meteor. Soc. Japan, 94, 151183, https://doi.org/10.2151/jmsj.2016-009.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Betz, H. D., K. Schmidt, W. P. Oettinger, and B. Montag, 2008: Cell-tracking with lightning data from LINET. Adv. Geosci., 17, 5561, https://doi.org/10.5194/adgeo-17-55-2008.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bonelli, P., and P. Marcacci, 2008: Thunderstorm nowcasting by means of lightning and radar data: Algorithms and applications in northern Italy. Nat. Hazards Earth Syst. Sci., 8, 11871198, https://doi.org/10.5194/nhess-8-1187-2008.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Buechler, D. E., and S. J. Goodman, 1990: Echo size and asymmetry: Impact on NEXRAD storm identification. J. Appl. Meteor., 29, 962969, https://doi.org/10.1175/1520-0450(1990)029<0962:ESAAIO>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Chen, L. C., G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, 2018: DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE Trans. Pattern Anal. Mach. Intell., 40, 834848, https://doi.org/10.1109/TPAMI.2017.2699184.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Chen, M., Y. Wang, F. Gao, and X. Xiao, 2012: Diurnal variations in convective storm activity over contiguous North China during the warm season based on radar mosaic climatology. J. Geophys. Res., 117, D20115, https://doi.org/10.1029/2012JD018158.

    • Search Google Scholar
    • Export Citation
  • Dabberdt, W. F., and et al. , 2005: Multifunctional mesoscale observing networks. Bull. Amer. Meteor. Soc., 86, 961982, https://doi.org/10.1175/BAMS-86-7-961.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm identification, tracking, analysis, and nowcasting—A radar-based methodology. J. Atmos. Oceanic Technol., 10, 785797, https://doi.org/10.1175/1520-0426(1993)010<0785:TTITAA>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Farnell, C., T. Rigo, and N. Pineda, 2017: Lightning jump as a nowcast predictor: Application to severe weather events in Catalonia. Atmos. Res., 183, 130141, https://doi.org/10.1016/j.atmosres.2016.08.021.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Garcia-Garcia, A., S. Orts-Escolano, S. Oprea, V. Villena-Martinez, and J. Garcia-Rodriguez, 2017: A review on deep learning techniques applied to semantic segmentation. arXiv, http://arxiv.org/abs/1704.06857.

  • Gatlin, P. N., and S. J. Goodman, 2010: A total lightning trending algorithm to identify severe thunderstorms. J. Atmos. Oceanic Technol., 27, 322, https://doi.org/10.1175/2009JTECHA1286.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Golik, P., P. Doetsch, and H. Ney, 2013: Cross-entropy vs. squared error training: A theoretical and experimental comparison. Proc. 14th Annual Conf., Lyon, France, International Speech Communication Association, 1756–1760.

  • Haberlie, A. M., and W. S. Ashley, 2018: A method for identifying midlatitude mesoscale convective systems in radar mosaics. Part I: Segmentation and classification. J. Appl. Meteor. Climatol., 57, 15751598, https://doi.org/10.1175/JAMC-D-17-0293.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Han, L., J. Sun, W. Zhang, Y. Xiu, H. Feng, and Y. Lin, 2017: A machine learning nowcasting method based on real-time reanalysis data. J. Geophys. Res. Atmos., 122, 40384051, https://doi.org/10.1002/2016JD025783.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Harris, R. J., J. R. Mecikalski, W. M. Mackenzie, P. A. Durkee, and K. E. Nielsen, 2010: The definition of GOES infrared lightning initiation interest fields. J. Appl. Meteor. Climatol., 49, 25272543, https://doi.org/10.1175/2010JAMC2575.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hou, A. Y., and et al. , 2014: The Global Precipitation Measurement Mission. Bull. Amer. Meteor. Soc., 95, 701722, https://doi.org/10.1175/BAMS-D-13-00164.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ioffe, S., and C. Szegedy, 2015: Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv, https://arxiv.org/abs/1502.03167.

  • Ji, S., M. Yang, and K. Yu, 2013: 3D convolutional neural networks for human action recognition. IEEE Trans. Pattern Anal. Mach. Intell., 35, 221231, https://doi.org/10.1109/TPAMI.2012.59.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Johnson, J. T., P. L. MacKeen, A. Witt, E. D. W. Mitchell, G. J. Stumpf, M. D. Eilts, and K. W. Thomas, 1998: The storm cell identification and tracking algorithm: An enhanced WSR-88D algorithm. Wea. Forecasting, 13, 263276, https://doi.org/10.1175/1520-0434(1998)013<0263:TSCIAT>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Karagiannidis, A., K. Lagouvardos, and V. Kotroni, 2016: The use of lightning data and Meteosat infrared imagery for the nowcasting of lightning activity. Atmos. Res., 168, 5769, https://doi.org/10.1016/j.atmosres.2015.08.011.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kingma, D., and J. Ba, 2015: ADAM: A method for stochastic optimization. Third Int. Conf. Learning Representations, San Diego, CA, International Machine Learning Society.

  • Kohn, M., E. Galanti, C. Price, K. Lagouvardos, and V. Kotroni, 2011: Nowcasting thunderstorms in the Mediterranean region using lightning data. Atmos. Res., 100, 489502, https://doi.org/10.1016/j.atmosres.2010.08.010.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • LeCun, Y., Y. Bengio, and G. Hinton, 2015: Deep learning. Nature, 521, 436444, https://doi.org/10.1038/nature14539.

  • Lee, S., H. Han, J. Im, E. Jang, and M. Lee, 2017: Detection of deterministic and probabilistic convection initiation using Himawari-8 Advanced Himawari Imager data. Atmos. Meas. Tech., 10, 18591874, https://doi.org/10.5194/amt-10-1859-2017.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Lynn, B. H., Y. Yair, C. Price, G. Kelman, and A. J. Clark, 2012: Predicting cloud-to-ground and intracloud lightning in weather forecast models. Wea. Forecasting, 27, 14701488, https://doi.org/10.1175/WAF-D-11-00144.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., and K. M. Bedka, 2006: Forecasting convective initiation by monitoring the evolution of moving cumulus in daytime GOES imagery. Mon. Wea. Rev., 134, 4978, https://doi.org/10.1175/MWR3062.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., W. M. Mackenzie, M. König, and S. Muller, 2010: Cloud-top properties of growing cumulus prior to convective initiation as measured by Meteosat Second Generation. Part II: Use of visible reflectance. J. Appl. Meteor. Climatol., 49, 25442558, https://doi.org/10.1175/2010JAMC2480.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., X. Li, L. D. Carey, E. W. McCaul, and T. A. Coleman, 2013: Regional comparison of GOES cloud-top properties and radar characteristics in advance of first-flash lightning initiation. Mon. Wea. Rev., 141, 5574, https://doi.org/10.1175/MWR-D-12-00120.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., J. K. Williams, C. P. Jewett, D. Ahijevych, A. LeRoy, and J. R. Walker, 2015: Probabilistic 0–1-h convective initiation nowcasts that combine geostationary satellite observations and numerical weather prediction model data. J. Appl. Meteor. Climatol., 54, 10391059, https://doi.org/10.1175/JAMC-D-14-0129.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Medina, B. L., L. D. Carey, C. G. Amiot, R. M. Mecikalski, W. P. Roeder, T. M. McNamara, and R. J. Blakeslee, 2019: A random forest method to forecast downbursts based on dual-polarization radar signatures. Remote Sens., 11, 826, https://doi.org/10.3390/rs11070826.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Metzger, E., and W. A. Nuss, 2013: The relationship between total cloud lightning behavior and radar-derived thunderstorm structure. Wea. Forecasting, 28, 237253, https://doi.org/10.1175/WAF-D-11-00157.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mosier, R. M., C. Schumacher, R. E. Orville, and L. D. Carey, 2011: Radar nowcasting of cloud-to-ground lightning over Houston, Texas. Wea. Forecasting, 26, 199212, https://doi.org/10.1175/2010WAF2222431.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Orlanski, I., 1975: A rational subdivision of scales for atmospheric processes. Bull. Amer. Meteor. Soc., 56, 527530, https://doi.org/10.1175/1520-0477-56.5.527.

    • Search Google Scholar
    • Export Citation
  • Perol, T., M. Gharbi, and M. Denolle, 2018: Convolutional neural network for earthquake detection and location. Sci. Adv., 4, e1700578, https://doi.org/10.1126/sciadv.1700578.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ray, P. S., 1986: Mesoscale Meteorology and Forecasting. Amer. Meteor. Soc., 793 pp.

    • Crossref
    • Export Citation
  • Rigo, T., N. Pineda, and J. Bech, 2010: Analysis of warm season thunderstorms using an object-oriented tracking method based on radar and total lightning data. Nat. Hazards Earth Syst. Sci., 10, 18811893, https://doi.org/10.5194/nhess-10-1881-2010.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ronneberger, O., P. Fischer, and T. Brox, 2015: U-Net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention, Vol. 9351, Springer, 234–241.

  • Salamon, J., and J. P. Bello, 2017: Deep convolutional neural networks and data augmentation for environmental sound classification. IEEE Signal Process. Lett., 24, 279283, https://doi.org/10.1109/LSP.2017.2657381.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sánchez, J. L., J. L. Marcos, M. T. De La Fuente, and A. Castro, 1998: A logistic regression model applied to short term forecast of hail risk. Phys. Chem. Earth, 23, 645648, https://doi.org/10.1016/S0079-1946(98)00102-5.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sánchez, J. L., E. García Ortega, and J. L. Marcos, 2001: Construction and assessment of a logistic regression model applied to short-term forecasting of thunderstorms in León (Spain). Atmos. Res., 56, 5771, https://doi.org/10.1016/S0169-8095(00)00089-2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schmidhuber, J., 2015: Deep learning in neural networks: An overview. Neural Networks, 61, 85117, https://doi.org/10.1016/j.neunet.2014.09.003.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schultz, C. J., L. D. Carey, E. V. Schultz, and R. J. Blakeslee, 2017: Kinematic and microphysical significance of lightning jumps versus nonjump increases in total flash rate. Wea. Forecasting, 32, 275288, https://doi.org/10.1175/WAF-D-15-0175.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Shelhamer, E., J. Long, and T. Darrell, 2017: Fully convolutional networks for semantic segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39, 640651, https://doi.org/10.1109/TPAMI.2016.2572683.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Shi, X., Z. Chen, H. Wang, D.-Y. Yeung, W. Wong, and W. Woo, 2015: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Proc. Advances in Neural Information Processing Systems, Montreal, QC, Canada, Curran Associates, 802–810, http://papers.nips.cc/paper/5955-convolutional-lstm-network-a-machine-learning-approach-for-precipitation-nowcasting.pdf.

  • Shi, X., Z. Gao, L. Lausen, H. Wang, D. Y. Yeung, W. K. Wong, and W. C. Woo, 2017: Deep learning for precipitation nowcasting: A benchmark and a new model. 31st Conf. on Neural Information Processing Systems, Long Beach, CA, Neural Information Processing Systems, 5617–5627.

  • Simonyan, K., and A. Zisserman, 2014: Very deep convolutional networks for large-scale image recognition. Third Int. Conf. on Learning Representations, San Diego, CA, International Machine Learning Society, https://arxiv.org/abs/1409.1556.

  • Stensrud, D. J., and et al. , 2013: Progress and challenges with warn-on-forecast. Atmos. Res., 123, 216, https://doi.org/10.1016/j.atmosres.2012.04.004.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sun, J., and et al. , 2014: Use of NWP for nowcasting convective precipitation: Recent progress and challenges. Bull. Amer. Meteor. Soc., 95, 409426, https://doi.org/10.1175/BAMS-D-11-00263.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Vincent, B. R., L. D. Carey, D. Schneider, K. Keeter, and R. Gonski, 2003: Using WSR-88D reflectivity for the prediction of cloud-to-ground lightning: A central North Carolina study. Natl. Wea. Dig., 27, 3544.

    • Search Google Scholar
    • Export Citation
  • Walker, J. R., W. M. Mackenzie, J. R. Mecikalski, and C. P. Jewett, 2012: An enhanced geostationary satellite-based convective initiation algorithm for 0–2-h nowcasting with object tracking. J. Appl. Meteor. Climatol., 51, 19311949, https://doi.org/10.1175/JAMC-D-11-0246.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wang, Y., M. Long, J. Wang, Z. Gao, and P. S. Yu, 2017: PredRNN: Recurrent neural networks for predictive learning using spatiotemporal LSTMs. 31st Conf. on Neural Information Processing Systems, Long Beach, CA, Neural Information Processing Systems, https://papers.nips.cc/paper/6689-predrnn-recurrent-neural-networks-for-predictive-learning-using-spatiotemporal-lstms.pdf.

  • Wilson, J. W., N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, 1998: Nowcasting thunderstorms: A status report. Bull. Amer. Meteor. Soc., 79, 20792099, https://doi.org/10.1175/1520-0477(1998)079<2079:NTASR>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilson, J. W., Y. Feng, M. Chen, and R. D. Roberts, 2010: Nowcasting challenges during the Beijing Olympics: Successes, failures, and implications for future nowcasting systems. Wea. Forecasting, 25, 16911714, https://doi.org/10.1175/2010WAF2222417.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Woo, W. C., and W. K. Wong, 2017: Operational application of optical flow techniques to radar-based rainfall nowcasting. Atmosphere, 8, 48, https://doi.org/10.3390/atmos8030048.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Xia, R., D. L. Zhang, and B. Wang, 2015: A 6-yr cloud-to-ground lightning climatology and its relationship to rainfall over central and eastern China. J. Appl. Meteor. Climatol., 54, 24432460, https://doi.org/10.1175/JAMC-D-15-0029.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yang, J., Z. Zhang, C. Wei, F. Lu, and Q. Guo, 2017: Introducing the new generation of Chinese geostationary weather satellites, Fengyun-4. Bull. Amer. Meteor. Soc., 98, 16371658, https://doi.org/10.1175/BAMS-D-16-0065.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yang, X., J. Sun, and W. Li, 2015: An analysis of cloud-to-ground lightning in China during 2010–13. Wea. Forecasting, 30, 15371550, https://doi.org/10.1175/WAF-D-14-00132.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yangli-Ao, G., and et al. , 2019: LightNet: A dual spatiotemporal encoder network model for lightning prediction. 25th Conf. on Knowledge Discovery and Data Mining, Anchorage, AK, Special Interest Group on Knowledge Discovery in Data, 2439–2447, https://doi.org/10.1145/3292500.3330717.

    • Crossref
    • Export Citation
  • Yumimoto, K., and et al. , 2016: Aerosol data assimilation using data from Himawari-8, a next-generation geostationary meteorological satellite. Geophys. Res. Lett., 43, 58865894, https://doi.org/10.1002/2016GL069298.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zhang, J., 2010: Multi-source remote sensing data fusion: Status and trends. Int. J. Image Data Fusion, 1, 524, https://doi.org/10.1080/19479830903561035.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zhang, W., L. Han, J. Sun, H. Guo, and J. Dai, 2017: Application of multi-channel 3D-cube successive convolution network for convective storm nowcasting. 2019 Int. Conf. on Big Data, Los Angeles, CA, IEEE, 1705–1710, https://doi.org/10.1109/BigData47090.2019.9005568.

    • Crossref
    • Export Citation
  • Zheng, Y. G., J. Chen, and P. J. Zhu, 2008: Climatological distribution and diurnal variation of mesoscale convective systems over China and its vicinity during summer. Chin. Sci. Bull., 53, 15741586, https://doi.org/10.1007/s11434-008-0116-9.

    • Search Google Scholar
    • Export Citation
Save