| Precipitation factor plays an important role in meteorological operation,precise precipitation data is important for disaster early warning of related departments.The real precipitation comes from the measurement of the rain gauge.Due to the uneven distribution of the rain gauge,it is impossible to obtain the high resolution precipitation data directly;radar echo data has a strong correlation with precipitation elements,and has the advantages of wide detection range and high spatial and temporal resolution.Therefore,radar quantitative estimation of precipitation has become the main important means to obtain high resolution precipitation,but it is still a challenging problem to achieve accurate estimation effect.There are mainly the following problems in the calculation of high-resolution precipitation data.Firstly,the estimation of precipitation by radar is mainly based on the Z-I relationship between reflectivity data and precipitation data and the correction of rain gauge.However,the Z-I relationship is limited by many assumptions,and the specific relationship is difficult to determine.Secondly,the formation of precipitation involves a complex evolution process,so it is difficult to reflect the real precipitation by using only the spatial information of radar echo.Finally,radar estimation of precipitation uses point-to-point method to convert reflectivity to precipitation,which ignores the spatial continuity of precipitation.Aiming at the above problems,this paper proposes a super resolution algorithm of precipitation elements on the basis of summarizing the previous work.The main research work includes the following contents:1.In this paper,the optical flow information of radar echo is applied to the super resolution process of precipitation elements.The change of radar echoes in time has a lot of relationships with the evolution of the precipitation,in order to calculate the radar echo flow accurately,this paper puts forward a kind of unsupervised learning model based on convolution neural network computing optical flow between radar echo,model uses the multi-layer convolution and residual learning,which effectively combine high-level features and underlying features,calculate accurate optical flow information.In order to solve the problem of lack of real optical flow field in echo data,this paper designs the appropriate loss function and integrates the method of data enhancement into the training process to enhance the generalization of network.2.In view of the estimation of precipitation by radar did not consider the problem of the spatial distribution of precipitation,this paper proposes a precipitation super-resolution algorithm based on generative adversarial networks,make the space of the generated data distribution accords with the objective law of real precipitation,and based on the relationship between the precipitation and radar echo data.Inspired by conditional generative adversarial networks,the radar echo and optical flow of echo are used as auxiliary condition variables for precipitation super-resolution,finally we use of a way called channel rearranging,combine all the resolution precipitation data to the 1 km resolution precipitation data.3.The channel attention mechanism is used to assist the network to focus on the effective information on the channel,and the spectral normalization method is used to constrain the discriminant to meet the Lipschitz constraint,so as to solve the problem of unstable training and difficult convergence of generative adversarial networks. |
PDF Full Text Request