Wind Direction Retrieval From CYGNSS Sea Surface Data Based On Machine Learning

Wind speed and direction are important basic climate variables.As a new technology emerging in recent years,GNSS-R(Global Navigation Satellite Systems Reflection)includes different signal reception methods such as ground,airborne and spaceborne.The remote sensing technology originates from the reflected signals of global satellites and the signals come from a large number of GPS satellites.It has the advantage of dealing with all-weather remote sensing.In addition,GNSS-R provides a relatively cheap passive method to observe ocean characteristics,such as sea surface height and wind speed.Remote sensing of wind direction using GNSS-R method has also aroused great interest.However,the near-mirror geometry used in the GNSS-R system makes the retrieval of wind direction information challenging,so there are few studies on the retrieval of wind direction from spaceborne GNSS-R observation data,namely,the delayed Doppler map(DDM).The previous research mainly focused on the retrieval of the ocean surface roughness information directly related to the sea surface wind field.The selected method is to use the characteristic parameters obtained directly from the measured DDM,and connect them with the required geophysical parameters through empirical relations.These characteristic parameters are based on the measurement of DDM diffusion caused by surface roughness,but their relationship with wind direction is relatively small and greatly affected by other factors,resulting in poor performance of the model or great limitations of the model itself.Machine learning has taken the level of building complex models to a new level.So far,many people have applied it to the retrieval of sea surface height or ocean wind speed,and achieved good results.In this paper,a variety of machine learning algorithms are introduced into the wind direction inversion,with the focus on selecting appropriate characteristic parameters,data preprocessing,and model adjustment and optimization.Finally,a machine learning model that minimizes the ambiguity of wind direction inversion and has a high degree of generalization in the world is constructed.The main contents of this paper are as follows:1.It is difficult to establish a generalized inversion model of sea surface wind direction around the world,so this paper proposes a method of sea surface wind direction inversion based on support vector machine(SVM).Using the Cyclone Global Navigation Satellite System(CYGNSS)satellite data,according to the correlation between the characteristics of the sea surface reflection signal and the wind direction,a variety of characteristic parameters were selected,including the DDM corresponding to the CYGNSS satellite parameters and geometric characteristic parameters.Select the radial basis function(RBF),and optimize the parameters based on the grid search method through cross validation.Finally,the SVM model of sea surface wind direction inversion is established.2.Based on the large amount and wide coverage of CYGNSS L1 data,a sea surface wind direction inversion model based on three machine learning algorithms is established.The wind direction will cause the asymmetry of DDM.On this basis,this paper extracts two angle feature parameters from DDM.Compared with CYGNSS Full DDM,L1 Compact DDM size is reduced.Therefore,this paper introduces more characteristic parameters,including L1 parameters and geophysical parameters such as wind speed,mean sea surface pressure(MSL),sea surface temperature(SST).The wind speed,direction,MSL and SST are all from the European Center for Medium Range Weather Prediction(ECMWF).After the data preprocessing process,the experimental data set is generated.Based on this data set,this paper establishes the wind direction inversion models of SVM,back-propagation(BP)and convolution neural network(CNN),and verifies its model performance and generalization performance.In addition,a filter is constructed to optimize the accuracy of CNN model,and the inversion effect under different wind directions is further studied.