Short-Term Intense Precipitation And Hail Nowcasting: A Machine Learning Perspective

Short-term intense precipitation(SIP)and hail are the most prevalent convective weather in China.They occur suddenly,evolve violently,and usually lead to severe hazards.Weather radar could provide atmospheric observation with high spatial and temporal resolution and is also the cornerstone of convective weather nowcasting.The radar-based SIP and hail nowcasting research is of great significance for disaster prevention and control.Recently,machine learning has brought progress to the meteorology field from a statistical standpoint.However,data-oriented modeling strategies also pose distinctive challenges.In this dissertation,we focus on three routes from the machine learning perspective: mining novel paradigms for traditional meteorological problems,model design guided by meteorological priors,and new settings brought by data science.Simultaneously,we attempt to solve three specific meteorological problems: radar extrapolation,SIP estimation,and hail identification.The radar extrapolation problem could be defined as an image sequence prediction problem.Typical loss functions usually blur output images,which makes models difficult to predict convective systems with rich structural details.In this dissertation,we introduce the conditional generative adversarial network(CGAN)to improve this phenomenon.Additionally,a "pre-extrapolation and post-processing" paradigm is also designed to accelerate model convergence under high spatiotemporal resolution radar data.Finally,a two-stage CGAN-based extrapolation model is constructed.Compared with traditional methods,our model could provide more precise radar echo extrapolations,especially for convective systems.The challenge of SIP estimation is caused by the mismatch between pixel-level estimators and the meteorological concept of convection.Convective systems are areas consisting of multiple associated pixels.It is difficult for pixel-by-pixel methods to characterize these regions comprehensively.In this dissertation,we use machine learning methods to accommodate data in different formats,and explore the point-wise estimation(SIP events)or field-wise(precipitation grids)estimation problems of multiple irregular regions(convective cells)in three stages.(1)Based on convective conception,we design a series of physical features to represent cells;then,guided by the graph model,we complete the transition from cells to convective systems;finally,a machine learning algorithm is used to predict SIP events.(2)A spatiotemporal attention model is proposed and used to focus on convective systems as an alternative to graph models.Our method achieves the extension from point estimation of SIP events to field estimation of standard precipitation grid.(3)Drawing on the advantages of the two stages,we combine the graph neural network and the convolutional neural network,and design a correction and fusion strategy with the meteorological principle.Additionally,we also build more cell features.Finally,a quantitative SIP estimation method with delicate representation is proposed.In experiments,the three stages could estimate the SIP more accurately than traditional methods.As the stage progresses,the advantages become more significant.The key problem to hail identification is the model performance deterioration caused by lack of modeling samples.It is determined by the data distribution and the observation process.In this dissertation,we make improvements from the perspective of transfer learning.Supported by the large-scale off-site samples with different distributions,we employ the distribution alignment strategy to obtain auxiliary information,and build a hail identification model with small-scale target samples.Furthermore,the transfer and classification loss functions are modified according to the data characteristics.Finally,the model computational efficiency is upgraded,and the performance is also comparable to the model supported by large-scale data.