Research On Wind Retrieval Technologies Of Spaceborne Laser Doppler Radar Based On Machine Learning

Wind field is an important meteorological element field for studying atmospheric dynamics and climate change,and it is also one of the important observations for atmospheric sounding.Although there are currently a variety of atmospheric wind field observation methods,they still cannot meet the urgent needs of meteorological forecasting practical work and high-resolution numerical weather forecasting systems for global three-dimensional,high-quality observation wind field information.Although conventional observation data such as radiosonde,small ball wind measurement and wind profiler can be directly assimilated and utilized,they are not evenly distributed globally,mainly on the land of the northern hemisphere,while observation data in sparsely populated areas such as deserts,plateaus,and oceans Limited;meteorological satellite cloud guide wind has the advantages of wide coverage and high temporal and spatial resolution,but the problem of height determination and designation has not been completely resolved,resulting in certain observation errors;satellite scatterometers can obtain real-time ocean surface wind field information,but It is mainly aimed at the lower atmospheric wind field in the ocean area,and the accuracy is poor under the condition of high wind speed.In order to realize the direct measurement of the global wind profile,the European Space Agency successfully launched the ADM-Aeolus satellite in 2018,which is equipped with an advanced laser Doppler wind radar.This satellite successfully realized the direct observation of the earth’s wind for the first time.The field is a landmark event in the field of international satellite remote sensing and atmospheric science.The remote sensing observations obtained by the Aeolus satellite are only the components of the earth’s atmosphere wind field along the satellite-target line of sight,and there are certain observation errors and deviations,so they cannot be used directly.The thesis studies the space-borne laser Doppler radar wind field retrieval technology based on machine learning,mainly considering the atmospheric wind vector obtained from the horizontal line-of-sight wind(HLOS)retrieval of the laser Doppler wind radar while ignoring the vertical component.Two wind field components are determined by a scalar observation HLOS,which is a typical ill-posed problem.In addition,due to the detection principle and various factors,the HLOS observation error model is complicated,and careful preprocessing and retrieval are required before retrieval.Quality control etc.In short,the wind field retrieval of spaceborne laser Doppler radar is a difficult scientific problem.In order to transform the Aeolus satellite wind data into a data form suitable for the habitual use of weather forecasters or directly used by the atmospheric data assimilation system,this paper proposes to use a variety of machine learning methods(K-nearest neighbor algorithm,support vector regression algorithm,random Forest algorithm and BP neural network),established a spaceborne laser Doppler radar wind field retrieval model,and successfully realized the function of obtaining the zonal wind and meridional wind components from the horizontal line-of-sight wind observation and retrieval.The retrieval process mainly includes the following four steps.First,extract the main features from the horizontal line-of-sight wind observation data of the Aeolus satellite,and perform screening and integration;select the data related to the retrieval of the horizontal line-of-sight wind of the Aeolus satellite as the input features of the retrieval process,which is mainly to re-analyze The ERA5 data is interpolated to the time and space position of the Aeolus data,and the interpolated zonal wind and meridional wind are used as the reference truth values for retrieval.Secondly,the horizontal line of sight wind data is screened using the experimental results of the observation system,and then the data is standardized using the Z-Score method to obtain the characteristic data and label data of the model.Then,use the preprocessed data of the Aeolus satellite data to establish a laser Doppler radar wind field retrieval model based on the machine learning method,and use the training set data constructed in the first step to train the machine learning model,and continuously adjust and get the various parameters of the adapted retrieval model.Finally,use the test set data to verify the accuracy of the output wind vector and other indicators on the retrieval model,and draw conclusions.It is concluded that the machine learning retrieval model established in this paper can successfully obtain the atmospheric zonal and meridional wind components from the horizontal line-of-sight wind observation.The retrieval accuracy is higher,but the inverted horizontal wind zonal wind component is more effective than the meridional wind component.Better;for the four different machine learning retrieval models,the BP neural network has the best retrieval effect;when the horizontal line of sight wind speed is greater than 10m/s,the retrieval effect is better than the wind speed less than10m/s,That is,the retrieval effect is better for the observation of the large value of the horizontal line of sight wind.