Research On Intelligent Nowcasting Method Of Hail And Heavy Rain Based On Weather Radar

Hail and heavy rain are two kinds of severe convective disastrous weather that are often occurred in China.Since they evolve rapidly and have a small spatial scale,their monitoring and forecasting rely on weather radars.The weather radar network has a wide coverage and high spatial and temporal resolution and could generate a large amount of data during operation.Therefore,radar-based nowcasting has a strong dependence on automatic forecasting algorithm.However,the existing nowcasting algorithms are poor-intelligent and cannot fully utilize the information provided by the radar data to make comprehensive forecasting decisions.The rapid development of artificial intelligence and related technologies has provided new tools for improving the radar-based nowcasting algorithm of convective storms.This paper studies the application of machine learning and image processing methods in weather-radar-based hail and heavy rain nowcasting,and specifically do the following work:(1)In order to introduce a radar morphological signature,the weak echo region,which has a strong correlation with hail,into automatic forecasting systems,the weak echo region identification algorithm and quantification algorithm are proposed.Then we analyze to find the correlation between the obtained quantity and hail occurrence.Besides,based on the weak echo region identification algorithm,an automatic crosssection making algorithm is proposed.This algorithm could generate cross-sections that display apparent weak echo region structures,which can help forecasters get more information about the storm internal.(2)In order to solve the problem of weather type decision of convective cells,relevant features are extracted from radar and sounding data based on convective cloud physics mechanism and operational forecasting experiences.Then based on this,a machine learning based hail versus heavy rain classification method is proposed.The method includes two machine learning models,the hail judgment classifier,and the precipitation prediction classifier,which can judge the weather type with certain forecast lead time.In addition,since the hail case data are difficult to be collected,a hail versus non-hail classification model based on positive-unlabeled learning is proposed.The model can be trained with only partially labeled hail samples and the remaining unlabeled samples,which can save many sample labeling costs.(3)Aiming at the problem that the high temporal and spatial variability of strong convective precipitation will affect the accuracy of quantitative precipitation estimation,a radar-gauge combined quantitative precipitation estimation algorithm based on adaptive segmentation is proposed.The algorithm uses the measurement information of the radar and rain gauges to cluster the observation points,and then divides the radar precipitation area into multiple sub-regions through clustering information,and makes independent precipitation estimation in each sub-region to avoid the interaction between different types of precipitations.