In recent years, China has suffered from frequent extreme precipitation events, and predicting their future trends has become an essential part of the current research on this issue. Because of the inevitable uncertainties associated with individual models for climate prediction, this study uses a machine learning approach to integrate and fit multiple models. The results show that the use of several evaluation metrics provides better results than the traditional ensemble median method. The correlation coefficients with the actual observations were found to improve from about 0.8 to 0.9, while the correlation coefficients of the precipitation amount (PRCPTOT), very heavy precipitation days (R20mm), and extreme precipitation intensity (SDII95) reached 0.95. Based on this, the precipitation simulations of moderate forced scenario for sharing socio-economic path (SSP2-4.5) from 27 coupled models in the Coupled Model Intercomparison Project Phase 6 (CMIP6) were used to explore potential changes in future extreme precipitation events in China and to calculate the distribution and trends of the PRCPTOT, extreme precipitation amount (R95pTOT), maximum consecutive 5-day precipitation (Rx5day), precipitation intensity (SDII), SDII95, and R20mm for the early 21st century (2023–2050), mid-21st century (2051–2075), and late 21st century (2076–2100), respectively. The results showed that the most significant increase in extreme precipitation indices is expected to occur by the end of the century, with the R95pTOT, Rx5day, and SDII95 increasing by 13.73%, 9.43%, and 9.34%, respectively, from the base period. The remaining three precipitation indexes, the PRCPTOT, SDII, and R20mm, also showed increases of 8.77%, 6.84%, and 4.02%, respectively. Additionally, there were apparent differences in the spatial variation of extreme precipitation. There were significant increasing trends of extreme precipitation indexes in central China and northeast China in the three periods, among which the total annual precipitation showed an increasing trend in central and northern China and a decreasing trend in western and south China. An increasing trend of annual precipitation intensity was found to be mainly concentrated in central China and south China, and the annual precipitation frequency showed a larger increasing trend at the beginning of this century. The annual precipitation frequency showed an increasing trend in the early part of this century. In general, all the indices showed an overall increasing trend in the future period, with the PRCPTOT, Rx5day, and SDII95 showing the most significant overall increasing trends.

Precipitation nowcasting is an important tool for nowcasting weather. In recent years, progress has been achieved in some models based on deep learning for precipitation nowcasting. However, these models do not consider the contextual relationships between the input data and the output of a network and their deficiency in capturing the information of prediction objects. To overcome these shortcomings, in this study, we propose a model that performs convolution operation on input data and the output of a Long short-term memory (LSTM) networks. Second, a self-attention operation is added to capture the local and global dependencies of the hidden state of LSTM. The proposed network structure is inserted in an encoding–forecasting network framework and applied to spatiotemporal sequence forecasting. Third, the outputs of the precede sequence are also regarded as the inputs of according LSTM layer and this operation effectively captures temporal feature of sequence data. Comprehensive experiments are conducted on the KTH action dataset and Hong Kong observation 07 radar echo maps dataset. The visual and quantitative prediction results demonstrate the accuracy and efficacy of the proposed model.

Improving the resolution of degraded radar echo images of weather radar systems can aid severe weather forecasting and disaster prevention. Previous approaches to this problem include classical super-resolution (SR) algorithms such as iterative back-projection (IBP) and a recent nonlocal self-similarity sparse representation (NSSR) that exploits the data redundancy of radar echo data, etc. However, since radar echoes tend to have rich edge information and contour textures, the textural detail in the reconstructed echoes of traditional approaches is typically absent. Inspired by the recent advances of faster and deeper neural networks, especially the generative adversarial networks (GAN), which are capable of pushing SR solutions to the natural image manifold, we propose using GAN to tackle the problem of weather radar echo super-resolution to achieve better reconstruction performance (measured in peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM)). Using authentic weather radar echo data, we present the experimental results and compare its reconstruction performance with the above-mentioned methods. The experimental results showed that the GAN-based method is capable of generating perceptually superior solutions while achieving higher PSNR/SSIM results.

This study developed a satellite, reanalysis, and gauge data merging model for daily-scale analysis using a random forest algorithm in Sichuan province, characterized by complex terrain. A high-precision daily precipitation merging dataset (MSMP) with a spatial resolution of 0.1° was successfully generated. Through a comprehensive evaluation of the MSMP dataset using various indices across different periods and regions, the following findings were obtained: (1) GPM-IMERG satellite observation data exhibited the highest performance in the region and proved suitable for inclusion as the initial background field in the merging experiment; (2) the merging experiment significantly enhanced dataset accuracy, resulting in a spatiotemporal distribution of precipitation that better aligned with gauge data; (3) topographic factors exerted certain influences on the merging test, with greater accuracy improvements observed in the plain region, while the merging test demonstrated unstable effects in higher elevated areas. The results of this study present a practical approach for merging multi-source precipitation data and provide a novel research perspective to address the challenge of constructing high-precision daily precipitation datasets in regions characterized by complex terrain and limited observational coverage.

Early warning and forecasting of tornadoes began to combine artificial intelligence (AI) and machine learning (ML) algorithms to improve identification efficiency in the past few years. Applying machine learning algorithms to detect tornadoes usually encounters class imbalance problems because tornadoes are rare events in weather processes. The ADASYN-LOF algorithm (ALA) was proposed to solve the imbalance problem of tornado sample sets based on radar data. The adaptive synthetic (ADASYN) sampling algorithm is used to solve the imbalance problem by increasing the number of minority class samples, combined with the local outlier factor (LOF) algorithm to denoise the synthetic samples. The performance of the ALA algorithm is tested by using the supporting vector machine (SVM), artificial neural network (ANN), and random forest (RF) models. The results show that the ALA algorithm can improve the performance and noise immunity of the models, significantly increase the tornado recognition rate, and have the potential to increase the early tornado warning time. ALA is more effective in preprocessing imbalanced data of SVM and ANN, compared with ADASYN, Synthetic Minority Oversampling Technique (SMOTE), SMOTE-LOF algorithms.

Nowcasting of severe convective precipitation is of great importance in meteorological disaster prevention. Radar echo extrapolation is an effective method for short-term precipitation nowcasting. The traditional radar echo extrapolation methods lack the utilization of radar historical data as well as overlooking the nonlinear motion of small- to medium-sized convective systems in radar echoes. To solve this, we propose a deep-learning model combining CNN and RNN. The model T-UNet proposed in this paper uses an efficient convolutional neural network of UNet architecture with a residual network, where the encoder and decoder networks are connected by nested dense skip paths, while a TrajGRU recurrent neural network is added at each layer, to achieve the perceptual capability of time series. The quantitative statistical evaluation showed that the use of T-UNet could improve the nowcasting skill (CSI score, HSS score) by a maximum of 10.57% and 7.80% over a 60 min prediction cycle. Further evaluation showed that T-UNet also improved the prediction accuracy and prediction performance in the strong echo region.