| The precipitation is a key factor that triggered the meteorological disasters,which has an important impact on human life and property.The subject of precipitation nowcasting is rainfall,especially convective weather such as short-term heavy rainfall.The precipitation nowcasting can offer the precipitation forecast products with high spatial and temporal resolution within two hours for region,which improves the ability to disaster monitoring and early warning.Therefore,the precipitation nowcasting is very important for the safety of people's property and economic development in the region.Dual polarimetric weather radar belongs to microwave remote sensing,which is one of the key ways to meteorological detection.Radar echo can penetrate the clouds,rain and snow with the ability of working all day,all-weather,and plays an irreplaceable role in the disaster weather monitoring,early warning and other aspects.The weather radar is widely used in China,providing high-precision and high coverage observation information.Weather radar data has become a powerful tool for people to study the meteorological detect and disasters forecast.Extrapolation technique based on radar data is the main technique in the precipitation nowcasting.In terms of data utilization,the traditional method didn't fully minning the information of historical precipitatation data,and just relied on the limited radar data for forecasting.In terms of forecast effect,the traditional method didn't consider the evolution of rainfall system,such as the occurrence,development and extinction of rainfall system,which led to the short timeliness of forecast.In terms of product refinement,the traditional method was not satisfying the requirement of the region,especially in the coastal areas with variable weather.By analyzing the existing relevant research,this study proposed three nowcasting methods which aim at offsetting the deficiency above.The research content mainly includes studying the innovative theoretical framework of precipitation nowcasting,the model construction and the practical application in Beibu Gulf area,which using new methods and technologies such as computer vision and machine learning based on the radar data.The purpose of this study was to build the system of theoretical and technical with high accuracy and refinement,and improve the ability of regional response to heavy rain.The result can provide strong support for the prevention and treatment of local meteorological disasters.To sum up,this paper has done the following innovative research:(1)An innovative theoretical framework for precipitation nowcasting was proposed.On the basis of summing up the definition and modeling of precipitation nowcasting,the theory framework including the data source and processing of precipitation nowcasting,the technology and the output of precipitation nowcasting were constructed based on the spatial information technology.This study further improved the study of precipitation nowcasting theoretically.(2)In this study,three precipitation nowcasting models were proposed from different angles,including the traditional extrapolation method based on subpixel optical flow(SPLK),the method based on random forest merge with extrapolation model(SPLK-RF)and the method based on deep learning(TSCNN-XGBoost).These new nowcasting models had special features.The traditional extrapolation method based on sub-pixel optical flow(SPLK)focused on the tracking of rainy pixels,which had more advantages in predicting small-scale and fast moving precipitation.The SPLK improved the predictability of storms about 4-10% compared to PPLK and TREC according the index CORR.The method based on random forest merging with extrapolation model(SPLK-RF)focused on solving the problem which was discontinuity in the forecast image by extrapolation technique,and had a good effect in the wide range of precipitation.The indexes of CORR and POD improved about 10-20%,and the FAR decreased about 5%,comparing with the SPLK.The method based on deep learning(TSCNN-XGBoost)refered to the method of temporal and spatial sequence prediction,and obtained the relevant rules by mining a large number of historical data.This method had a good effect on most rainfall forecasting because use of similarity modeling method.The index of CORR was higher than the other algorithms obviously.The accuracy in small-scale storm such as storm3 improved about 50%,and improved about 10% in the convective event such as storm2,which compared with the SPLK-RF.These three new nowcasting methods all can get a good effect on different types of precipitation and improve the accuracy.This study has enriched the method system of precipitation nowcasting.(3)The paper studied the precipitation nowcasting in Beibu Gulf area.The new nowcasting methods were used to forecasting the precipitation in Beibu Gulf area systematically,and got some innovative conclusions.Among them,the deep learning-based nowcasting method TSCNN-XGBoost has the best effect in the study area.The value of CORR increased by nearly 20% comparing with other nowcasting methods,while CSI increased by nearly 10%.The second performance was the traditional extrapolation method based on sub-pixel(SPLK).The SPLK-RF had a poor performance in the study area which its average CSI index also reached 0.4.The results established foundation to nowcasting systems in Beibu Gulf area.This study innovated in practice and provided a good example for similar regional precipitation nowcasting.From the above research,it is also found that the precipitation nowcasting was short in timeliness.This is because the extrapolation technique based on radar cannot obtain the internal driving force of the evolution of rainy system,which is lead to the timeliness and accuracy are not high.In order to get a good effect,it is necessary add various meteorological elements to the nowcasting and combine with the NWP,which can solve the problem of extrapolation fundamentally.This is one of hot spot and difficult questions of in the field of precipitation nowcasting,which needs further study. |
PDF Full Text Request