Lightning Nowcasting with Multisource Data: A Deep Learning Approach

Based on multi-source observation data such as geostationary meteorological satellite, Doppler weather radar, and cloud-ground lightning, a lightning nowcasting model with the deep learning is proposed in this study. Considering the characteristics of geographical distribution and spatial resolution of radar and lightning observations in China, the mid-eastern region of China was selected as the experimental analysis area of this paper, with a spatial resolution of 0.05° longitude×0.05° latitude. Twelve bands (including visible, infrared, water vapor, etc.) from Himawari satellite and radar mosaic products (including basic reflectivity, vertically integrated liquid, etc.) were used as the predictors, and they are marked with lightning data to establish a training set with millions of samples. According to the characteristics of lightning development, a deep three-dimensional(3D) convolutional neural network(CNN) including 3D convolutional layer, pooling layer, fully connected layer, softmax classifier, etc., was constructed and trained. The training results showed that the test set classification accuracy exceeded 94%. The performance of the trained model was evaluated. The results show that the Threat Score(TS), Probability of Detection(POD）and False Alarm Rate(FAR）of 0-1 hour lightning nowcasting reach 0.51 ,0.60 and 0.23, respectively, in August 2017. Because the softmax classifier is used in the deep learning network and the forecast result is a probabilistic forecast product. The nowcasting results show that the regions with larger probability values usually correspond to the area with the intense lightning activities, which means the probability is a good indicator for thunderstorms. Since satellites have a capability of convection initiation(CI) observation, this model has a good performance in CI nowcasting before it is generated. The lightning prediction model implemented in this paper using the satellite, radar and lightning data could effectively improve the prediction performance of thunderstorms compared to that based on only a single-source data.