| Severe convective weather(SCW),which includes lightning or thunderstorms,hail,convective gusts(CG),short-duration heavy rain(HR),and tornadoes,poses a serious threat to life and property in most of the world.Due to the rapid evolution of small-scale convective systems and their complicated interaction with environmental features,forecasting SCW is still a challenging topic in operational meteorology today.Nowadays,nowcasting of SCW is still mainly depends on the extrapolation of observation data and short-range forecasting is depending on the "ingredients method".As a result,the performance of the above prediction needs to be improved.With the fast development of the high spatial-temporal resolution observation data as well as the high-resolution Numerical Weather Prediction(HNWP),the capability of SCW prediction is enhancing stably.However,the information contained in the observation and SCW data is too much to analysis for the weather forecasters(WFs).Thus,subjective and automatic method should be developed to help the WFs to make a quick and effective decision during the SCW forecasting process.In this work,we have constructed SCW prediction products covering nowcasting,very short-range forecasting and short-range forecasting based on deep learning method.We have developed SCW nowcasting products with multi-source observation data,very short-range forecasting products with multi-source observation data and HNWP data,short-range forecasting products with global NWP data.Besides,we have tried to make the black box of deep learning more transparent with different model interpretation and visualization methods.For nowcasting(0-2 h)of SCW,a new semantic segmentation deep learning network for cloud-to-ground lightning nowcasting,named Lightning Net,has been developed.This network is based on multi-source observation data,including data from a geostationary meteorological satellite,Doppler weather radar network,and the cloudto-ground lightning location system in China.Lightning Net,with an encoder-decoder architecture,consists of 20 three-dimensional convolutional layers,pooling and upsampling layers,normalization layers,and a softmax classifier.Lightning Net was trained to extract the features of lightning initiation,development,and dissipation.The evaluation results demonstrated that Lightning Net is able to achieve good performance of 0-1-hr lightning nowcasts using the multi-source data.The probability of detection,the false alarm ratio,the area under relative operating characteristic curve,the threat score(TS)of Lightning Net with all three types of data reached 0.633,0.386,0.931,and0.453,respectively.Because geostationary meteorological satellite and radar both possess the capability of capturing lightning initiation(LI)features,Lightning Net 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.Lightning Net’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 Lightning Net has good capability of combining multi-source data effectively to produce more reliable lightning nowcasts.For the very short-range(VSR,2-6 h)SCW forecasting,a semantic segmentation deep learning network,named Lightning Net-NWP,was used to merge the multi-source observation data and HNWP data to get better VSR lightning forecasts.The predictors of Lightning Net-NWP including lightning density,radar reflectivity,six infrared bands of Himawari-8,as well as the radar composite reflectivity from GRAPES3 km.Because the observation and HNWP data differs a lot,we designed two encode-decode symmetry sub-networks to extract the futures in above two data sources.The experimental results show that Lightning Net-NWP could combine the observation and HNWP data effectively and make a good lightning prediction for next 0-6 hours.The performance of Lightning Net-NWP with observation and HNWP data is much better than observation or HNWP data was used alone.The longer is the prediction,the advantage of combinational use of observation and HNWP data is larger.For short-range(6-72 h)forecasting of SCW,a deep learning objective forecasting solution for SCW,including short-duration HR,hail,CG and thunderstorms,based on NWP data was developed.Five years of severe weather observations were utilized to label the National Centers for Environmental Prediction(NCEP)final(FNL)analysis data.A large number of labeled samples for each type of weather were then selected for model training.The local temperatures,pressures,humidities,and winds from 1000 h Pa to 200 h Pa,as well as dozens of convective physical parameters,were taken as predictors in our model.A six-layer convolutional neural network(CNN)model was then built and trained to obtain optimal model weights.After that,the trained model was used to predict SCW based on the GFS forecast data as input.The objective forecasts of the deep learning algorithm also showed better forecasting skills than the subjective forecasts from the meteorologist,by increasing the threat scores(TSs)of thunderstorm,HR,hail and CG by 16.1%,33.2%,178%,and 55.7%,respectively.In order to understand the above deep learning networks better,we have used various methods,including impurity importance,permutation importance,convolutional layer output visualization,to explain how deep learning making predictions.The 144 predictors for short-range SCW forecasting is ranked in term of importance.The results show that the understanding of machine learning algorithm is similar to the meteorologists,accompanying by some heuristic points.The permutation importance of Lightning Net-NWP shows that the longer is the prediction,the more important is the HNWP predictor.The satellite data is most important for prediction of next 0-2 hours,while the HNWP predictor is most important for prediction of next 3-6hours.The convolutional layers of Lihgnting Net were visualized to analysis the process of lightning nowcasting. |
