Retrieval And Fusion Of Multi-source Remote Sensing Sea Surface Wind Field Based On Machine Learning

Sea surface wind is the main driving force for the movement of sea water,and it is also an important channel for air-sea interaction.Accurately obtaining sea surface wind field information and generating high-quality sea surface wind field products have important research value and practical significance for marine economic development and disaster prevention.The development of satellite remote sensing technology has made up for the shortcomings of traditional sea surface wind field observation methods and provided a large amount of observation data for related research on sea surface wind fields.In recent years,the use of satellite remote sensing data to retrieve sea surface wind fields has become a research hotspot.Improving the retrieval accuracy is the focus of work,which is also the key to accurately obtaining sea surface wind field information.However,remote sensing sea surface wind field products obtained from a single data source will have defects such as low temporal and spatial coverage and low resolution.Therefore,it is necessary to carry out multi-source satellite remote sensing sea surface wind field fusion technology research.In this paper,we analyze different methods for retrieval of wind field observation data from spaceborne microwave sensors,and find that traditional methods generally have problems such as complex retrieval models,neglect of the spatial continuity of wind fields,and ambiguous solutions.In order to solve the above problems and improve the accuracy of wind field retrieval,we propose a field to field(F2F)wind field retrieval method(F2FCNN)based on Convolutional Neural Networks(CNN)for the observation data of the domestic Hai Yang-2B satellite(HY-2B)microwave scatterometer and the Fengyun3 D satellite(FY3D)microwave imager.This method inputs continuous wind field observation data within a specific range into the retrieval model at the same time,constructs multiple neural network convolutional layers to fully extract the spatial consistency and continuity characteristics of the wind field,and designs multiple neural network fully connected layers to integrate the entire wind field The wind vector of is obtained at the same time,without the need for ambiguous solutions removal process,and a smooth and continuous wind field retrieval result is obtained.The retrieval results are verified on three different data sets,and the root mean square error compared with the buoy observation data is used as the wind field accuracy measurement standard.The results show that:(1)The F2F-CNN method improves the wind speed retrieval accuracy of HY-2B scatterometer data to 0.75 m/s,and the wind direction retrieval accuracy to 0.18 rad(10.31°);(2)The F2F-CNN method has the ability to clearly and accurately retrieve the wind field structure of complex weather systems such as cyclones;(3)The F2F-CNN method is also effective for wind field retrieval of radiometer data,it not only improves the accuracy of wind speed retrieval of existing radiometers,but also provides an accurate and complete description of the structure of high wind speed wind fields.We also analyze the advantages and disadvantages of different wind field fusion methods,and proposes a space-time weighted wind field fusion method based on multilevel quality control,which solves the problem of ignoring single-point interpolation accuracy in traditional space-time weighted fusion methods.The wind field data participating in the fusion shall be strictly controlled in terms of observation quality and retrieval accuracy to ensure the integrity,consistency and accuracy of the wind field data.The root mean square error,correlation coefficient and other parameters of the wind farm products participating in the fusion are introduced into the interpolation weight setting process of the newly proposed fusion method,so that the quality evaluation parameters of the multisource wind farm have an impact on the setting of the interpolation weight of the newly proposed fusion method.Complete the wind field information fusion,and finally obtain a high-precision wind field fusion product with strong consistency characteristics and a global coverage in 2020,with a temporal and spatial resolution of 24 hours and 0.25°,respectively.