Using Deep Learning For Severe Convective Nowcasting

Severe convective storms typically occur with heavy rain,hail,lightning,and strong winds,characterized by sudden,local and non-linear evolution.Therefore,severe convection nowcasting is a very important and still challenging topic.Although traditional extrapolationbased techniques could capture the rough trend of convection movement,it is difficult to predict the initiation and dissipation of storms using weather radar reflectivity and/or satellite images.With the development of next-generation weather radar,geostationary satellites and automatic weather stations,the understanding and nowcasting of severe convective,in particular convective initiation(CI),is remarkably improved.However,there are still some troubles in traditional extrapolation-based techniques due to the rapid increase of high spatial-temporal resolution data volume.Traditional CI prediction methods have high false alert rate,and it is difficult to accurately predict the timing and location of CI occurrence.Data-driven deep learning,which can automatically extract features from a large amount of data and have powerful non-linear mapping capabilities,presents new opportunities for severe convective,in particular CI,nowcasting.The spatial-temporal inconsistency of observations and single-task learning method(i.e.,either precipitation or lightning prediction each time)without consideration of the connection between disaster weather are two major problems based on multi-source observations for severe convective nowcasting using deep learning.Combined with geostationary satellite,weather radar,and ground-based automatic stations observations,this study focuses on the spatial-temporal distributions of severe convective and CI,and the improvement of CI,and convective storms precipitation and lightning nowcasting employing deep learning.The main results are as follows:(1)Investigated the spatial-temporal distribution of severe convection and convective initiation under different topographic conditions in ChinaBased on Himawari-8 infrared brightness temperature observations and a multi-threshold convective identification and tracking method,the spatial-temporal distribution of severe convection and CI during the warm season from 2016 to 2020 were systematically analyzed under different topographic conditions in China.The results showed that three hot spots of severe convection and initiation include the Qinghai-Tibet Plateau,Yunnan-Guizhou Plateau,and South China.During the daytime,CI mainly occurs near inland mountainous,tropical islands,and coastal mountainous areas.By contrast,CI mainly occurs in basins,plains,and coastal areas at night.There exist significant seasonal variations in convection activity and CI over the regions of land and ocean,presenting an increasing trend of the occurrence frequency of convection activity and CI from April to August with peak values between July and August.Differing from the unobvious diurnal variations in CI occurrence over the ocean,a significant diurnal evolution of CI is found in the land region.CI occurrence usually peaks between 11-14LST(Local Solar Time),but the peak hour of CI occurrence exhibits obvious regional differences under different topographic regions over land.(2)Developed a probabilistic forecasting method in timing and location of CI occurrence based on multi-channel infrared brightness temperature observations of geostationary satelliteA new CI identification algorithm is proposed to identify and filter CI using Himawari-8multi-channel infrared brightness temperature observations and NASA GPM(Global Precipitation Measurement Mission)precipitation products.Based on CI events identified by the algorithm,an event-based patch method is used to create the CI dataset.The deep learning CIUNet model is developed for predicting the timing and location of CI with a lead time of 30 min using the CI dataset.The permutation test and layer-wise relevance propagation method are employed to improve the understanding of how and why the CIUNet model makes its decisions.The verification metrics based on an independent test set of 3164 CI events demonstrate that the POD(probability of detection)and FAR(false alarm rate)achieve 93.3%and 18.3%,respectively.The CIUNet model can accurately predict the timing and location of CI derived from a typical convective outbreak with lead times of at least 30 min.According to the sensitivity experiments and permutation test results,the difference between brightness temperature channels,which represents the cloud-top height and cloud-top glaciation,plays a critical role in reducing FAR and improving POD.In addition,the terrain height has a positive effect on CI nowcasting over complex terrain.The layer-wise relevance propagation analysis of the typical CI events reveals that the CIUNet model can learn the regional brightness temperature characterization contributing to CI forecasting accuracy,which is consistent with the results of satellite infrared brightness thresholds-based.The results indicate that explainable deep learning can enhance the understanding of CI forcing factors and related mechanisms.(3)Constructed a precipitation nowcasting method coupling with weather radar,rain gauge,and geostationary satellite observationsThis method is presented for 0-2 h precipitation nowcasting using multiple deep learning models based on the Himawari-8 geostationary satellite,weather radar,and rain gauge precipitation product.The impact of deep learning model structure on precipitation nowcasting is investigated.The results demonstrate that the Conv LSTM and 3D UNet models have similar skill scores but are obviously better than the 2D UNet,Phy DNet,and Earth Former models.In addition,the effects of multi-source observations on precipitation nowcasting using the Conv LSTM model are explored,and the results reveal rain gauges made the largest contributions for precipitation nowcasting,followed by weather radar reflectivity and satellite infrared brightness temperatures.A two-branch Bi-Conv LSTM model is developed to address the problem of spatial-temporal inconsistencies in multi-source observations.The results demonstrate that the performance of precipitation nowcasting using different channels infrared temperature observations had significant differences.Compared to using only weather radar reflectivity and rain gauge precipitation observations,the skill scores improve by more than3.5% when adding the 10.4 channel brightness temperature observations as additional input for precipitation nowcasting.However,the model performance degrades when using all infrared channel brightness temperature observations for precipitation nowcasting,which indicates that the deep learning models are difficult to solve information redundancy amongst multi-source observations.In addition,the model ensemble and bias correction methods are used to improve the precipitation forecast skill scores.It alleviates the decay of forecast skill scores with lead time and precipitation under-forecasting.(4)Proposed a multi-task learning method for VIL and lightning nowcasting based on GOES satellite brightness temperature,weather radar,and GLM measurementsThe method is proposed for simultaneously forecasting 0-90 min VIL(vertically integrated liquid)and lightning using a multi-task learning model with a spatiotemporal power-weighted loss function based on the GOES-16 geostationary satellite,MRMS(Multi-Radar/Multi-Sensor System)VIL product and GLM(Geostationary Lightning Mapper).The impact of single-task learning and multi-task learning,and infrared brightness temperature observations of geostationary satellite on VIL and lightning nowcasting are explored.The results demonstrate that both multi-task and single-task learning models outperform the persistence model and optical flow method for VIL and lightning nowcasting.Multi-task learning is better than singletask learning,indicating that simultaneous prediction of VIL and lightning can more effectively extract feature information to further improve VIL and lightning nowcasting.Using infrared brightness temperature observations significantly improves lightning nowcasting skill scores,but has a relatively small impact on VIL forecasting.The verification metrics of organized and isolated convection nowcasting show that both multi-task and single-task learning models perform better for organized convection nowcasting than for isolated convection.