A deep learning objective forecasting solution for severe convective weather (SCW) including short-duration heavy rain (HR), hail, convective gusts (CG), and thunderstorms based on numerical weather prediction (NWP) data was developed. We first established the training datasets as follows. Five years of severe weather observations were utilized to label the NCEP final (FNL) analysis data. A large number of labeled samples for each type of weather were then selected for model training. The local temperature, pressure, humidity, and winds from 1000 to 200 hPa, 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 Global Forecast System (GFS) forecast data as input. The performances of the CNN model and other traditional methods were compared. The results show that the deep learning algorithm had a higher classification accuracy on HR and hail than support vector machine, random forests, and other traditional machine learning algorithms. The objective forecasts by use of the deep learning algorithm also showed better forecasting skills than the subjective forecasts by the forecasters. The threat scores (TSs) of thunderstorm, HR, hail, and CG were increased by 16.1%, 33.2%, 178%, and 55.7%, respectively. The deep learning forecast model is currently used in the National Meteorological Center of China to provide guidance for the operational SCW forecasting over China.

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 (T_B) of six infrared bands from theHimawari-8satellite, 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.

Tornadoes, as highly destructive weather events, require accurate detection for effective decision‐making. Traditional radar‐based tornado detection algorithms (TDA) face challenges with limited tornado feature extraction capabilities, leading to high false alarm rates and low detection probabilities. This study introduces the Multi‐Task Identification Network (MTI‐Net), leveraging Doppler radar data to enhance tornado recognition. MTI‐Net integrates tornado detection and estimation tasks to acquire comprehensive spatial and locational information. As part of MTI‐Net, we introduce a novel backbone network of Multi‐Head Convolutional Block (MHCB), which incorporates Spatial and Channel Attention Units (SAU and CAU). SAU optimizes local tornado feature extraction, while CAU reduces false alarms by enhancing dependencies among input variables. Experiments demonstrate the superiority of MTI‐Net over TDA, with a decrease in false alarm rates from 0.94 to 0.46 and an increase in hit rates from 0.23 to 0.81, highlighting the effectiveness of MTI‐Net in handling small‐scale tornado events.

It is argued here that even with the development of objective algorithms, convection-allowing numerical models, and artificial intelligence/machine learning, conceptual models will still be useful for forecasters until all these methods can fully satisfy the forecast requirements in the future. Conceptual models can help forecasters form forecast ideas quickly. They also can make up for the deficiencies of the numerical model and other objective methods. Furthermore, they can help forecasters understand the weather, and then help the forecasters lock in on the key features affecting the forecast as soon as possible. Ultimately, conceptual models can help the forecaster serve the end users faster, and better understand the forecast results during the service process. Based on the above considerations, construction of new conceptual models should have the following characteristics: 1) be guided by purpose, 2) focus on improving the ability of forecasters, 3) have multiangle consideration, 4) have multiscale fusion, and 5) need to be tested and corrected continuously. The traditional conceptual models used for forecasts of severe convective weather should be replaced gradually by new models that incorporate these principles.

Tornadoes are highly destructive weather events, and their timely detection is crucial for mitigating damage and saving lives. Doppler weather radar serves as the primary operational tool for tornado detection. However, existing detection methods mainly struggle with high false alarms. In addition, the potential mismatch between tornado reports and the corresponding radar signatures further degrade the detection accuracy. To tackle these issues, we propose TorDet, a novel two-stage deep learning approach designed for tornado detection using Doppler weather radar data. TorDet introduces a quantitative tornado labeling scheme that integrates the prominence of tornado-related radar features and tornado reports, explicitly categorizing samples (TOR, WRN, WEK, and NUL) to reduce ambiguity in both manual labeling and model training. TorDet’s two-stage design divides the detection process into distinct yet complementary tasks. In the Detection stage, it generates low-resolution feature maps to identify potential tornado-related regions by capturing large-scale information. The Positioning stage refines these detections, accurately locating tornado centers and classifying their categories. To better exploit limited data, we introduce deep supervision at intermediate layers in both stages, improving performance by enhancing the loss function and enriching gradient for backpropagation. We conduct experiments using radar data collected between 2017 and 2024. Experimental results demonstrate that TorDet, benefiting from its unique design and training strategy, reduces false alarms and maintains high correct detections compared with baseline deep learning models. On the test set, it improves the Probability of Detection (POD) from 0.511 to 0.647, reduces the False Alarm Ratio (FAR) from 0.795 to 0.697, and increases the Critical Success Index (CSI) from 0.171 to 0.260. Additional evaluation on an open-source dataset further supports its robustness and generalization capability across diverse radar data sources.