| In recent years,with the development of science and technology,the meteorological department has obtained more and more data benefit by the improved meteorological equipment.The traditional methods are incapable of achieving high spatial and temporal resolution,and the performance of observation equipment is limited,resulting in insufficient stability of the measured rainfall results.The data obtained by the observation equipment cannot directly reveal the characteristics of regional climate change,which faces many limitations on practical applications.The reflectivity factor of doppler weather radar has a strong correlation with precipitation intensity,using radar to estimate precipitation has become one of the important tasks for many designs and applications with meteorological purposes for the advantage of a wide estimation range,high spatiotemporal resolution,and access to large-area precipitation data.On the other hand,rainfall has the characteristics of strong spatiotemporal characteristics and process changes,so radar cannot efficiently make fast and adaptive feedback with rainfall intensity.How to solve these problems is the key to radar inversion rainfall.Thus,this paper mainly uses deep learning technology to study the radar echo retrieval rainfall problem,and designs a deep learning algorithm suitable for the meteorological field.It uses the high spatiotemporal resolution of radar echo to get a refined rainfall estimation.The purpose is to reveal the evolution and movement of radar echoes,and to retrieve the current rainfall situation based on radar echoes.The main work of this paper are as follows:(1)Unlike data types for video task,radar echo data is highly imbalanced and there is currently no benchmark for radar echo retrieval rainfall issues.We established a baseline dataset with radar echo data and rainfall data collected by the Hunan Meteorological Observatory from February to October 2018.First,the radar data and rainfall data were analyzed and preprocessed to eliminate outliers,and the minimum and maximum values of each of the first six layers of the radar data were normalized to make the data scale to a reasonable range;Then,a set of baseline radar rainfall data set is constructed according to the organization form of radar and rainfall file.Through the self-defined hierarchical cache method,the original data is transformed into a specific matrix format that can be quickly loaded and processed by memory and video memory,so as to optimize the data loading speed.(2)since the Z-R relation between radar and rain has a number of parameters on different areas,and the rainfall varies with seasons,it cannot reflect the space-time characteristics of radar data and rainfall data.This paper proposes a radar quantitative precipitation estimation algorithm based on the spatiotemporal network model(ST-QPE),which designs a convolutional time-series network QPE-Net8 and a multi-scale feature fusion time-series network QPE-Net22 based on convolutional neural network and recurrent neural network.The error and accuracy characteristics of radar inversion rainfall effect are verified by using statistical and meteorological methods.The experimental results show that the two spatiotemporal networks based on the ST-QPE algorithm can effectively learn the characteristics of radar data.The inverted rainfall map is consistent with the live rainfall map,and the QPE-Net22 network model has good stability.The inverted rainfall map is consistent with the current rainfall map,and the QPE-Net22 network model has good stability.In the comparison and evaluation of various indicators,the QPE-Net22 network results of multi-scale feature fusion timing are closer to the true rainfall value than the QPE-Net8 network and the classic GRU network. |
PDF Full Text Request